1 Supplementary Information for

3 Extending resolution of structured illumination microscopy with sparse deconvolution

4 Weisong Zhao1,14, Shiqun Zhao2,14, Liuju Li2,14, Xiaoshuai Huang5, Shijia Xing2, Yulin Zhang2, Guohua

5 Qiu1, Zhenqian Han1, Yingxu Shang9, De-en Sun10, Chunyan Shan13, Runlong Wu2, Shuwen Zhang11,

6 Riwang Chen7, Jian Xiao12, Yanquan Mo2, Jianyong Wang6, Wei Ji11, Xing Chen10, Baoquan Ding9, Yanmei

7 Liu2,8, Heng Mao4, Baoliang Song12, Jiubin Tan1, Jian Liu1, Haoyu Li1✉, Liangyi Chen2,3,6✉

1School
9 of Instrumentation Science and Engineering, Harbin Institute of Technology, Harbin 150080, China.
2State
10 Key Laboratory of Membrane Biology, Beijing Key Laboratory of Cardiometabolic Molecular Medicine, Institute of

11 Molecular Medicine, Peking University, Beijing 100871, China.

3PKU-IDG/McGovern
12 Institute for Brain Research, Beijing 100871, China.
4School
13 of Mathematical Sciences, Peking University, Beijing 100871, China.
5Biomedical
14 Engineering Department, Peking University, Beijing 100191, China.
6Shenzhen
15 Bay Laboratory, Shenzhen 518055, China.
7School
16 of Software and Microelectronics, Peking University, Beijing 100871, China.
8Institute for
17 Brain Research and Rehabilitation (IBRR), Guangdong Key Laboratory of Mental Health and Cognitive Science,

18 South China Normal University, Guangzhou, China.

9CAS
19 Key Laboratory of Nanosystem and Hierarchical Fabrication, CAS Center for Excellence in Nanoscience, National

20 Center for Nanoscience and Technology, Beijing, China.

10College
21 of Chemistry and Molecular Engineering, Peking University, Beijing, China.
11National
22 Laboratory of Biomacromolecules, CAS Center for Excellence in Biomacromolecules, Institute of Biophysics,

23 Chinese Academy of Sciences, Beijing, China.

12Hubei
24 Key Laboratory of Cell Homeostasis, College of Life Sciences, Wuhan University, Wuhan, China.
13College
25 of Life Sciences, Peking University, 100871, Beijing, China.
14These
26 authors contributed equally: Weisong Zhao, Shiqun Zhao, Liuju Li.

27 ✉
e-mail: [email protected]; [email protected]

1
28 Content
29 Supplementary Notes .................................................................................................................................................3
30 Supplementary Note 1 | Image formation and the resolution limit to the Wiener deconvolution...................................... 3
31 Supplementary Note 2 | Computational superresolution. .................................................................................................. 4
32 Supplementary Note 3 | The development of the sparse deconvolution model. ................................................................ 8
33 Supplementary Note 4 | The details of algorithm execution............................................................................................ 14
34 Supplementary Note 4.1 | Iterative approximation based on the weighted constraints. .......................................... 14
35 Supplementary Note 4.2 | Iterative deconvolution. ................................................................................................. 16
36 Supplementary Note 4.3 | Background estimation. ................................................................................................. 17
37 Supplementary Note 4.4 | Upsampling operation. ................................................................................................... 18
38 Supplementary Note 5 | Using the sparse and continuity prior knowledge improves resolution and contrast. ............... 20
39 Supplementary Note 6 | Parameters in the sparse deconvolution software. .................................................................... 26
40 Supplementary Note 7 | Optimal values of the sparsity and the fidelity under different conditions. .............................. 29
41 Supplementary Note 8 | Parameters finetuning ............................................................................................................... 33
42 Supplementary Note 8.1 | Procedures for parameters finetuning. ........................................................................... 33
43 Supplementary Note 8.2 | Step-by-step examples. .................................................................................................. 33
44 Supplementary Note 8.3 | More explanations on the background estimation.......................................................... 36
45 Supplementary Note 9 | The calibration of nuclear pore diameters and the FWHM resolution of Sparse SD-SIM. ...... 44
46 Supplementary Note 9.1 | Calibrating fitted diameters of ring-shaped pores. ......................................................... 44
47 Supplementary Note 9.2 | The bead correction factor. ............................................................................................ 44
48 Supplementary Note 10 | Other sparsity-based reconstructions that improve spatial resolution in theory. ..................... 48
49 Supplementary Note 10.1 | FISTA method. ............................................................................................................ 48
50 Supplementary Note 10.2 | NLHT method. ............................................................................................................. 50
51 Additional Supplementary Figures .........................................................................................................................52
52 Supplementary Tables ..............................................................................................................................................55
53 Supplementary Table 1 | SSIM of different reconstructions from images corrupted with varying levels of noise. ........ 55
54 Supplementary Table 2 | PSNR of different reconstructions from images corrupted with varying levels of noise. ........ 55
55 Supplementary Table 3 | Imaging conditions................................................................................................................... 56
56 Supplementary Table 4 | Non-zero ratios and background selections of different images. ............................................. 57
57 Supplementary Table 5 | Computing time of Sparse deconvolution. ............................................................................... 59
58 Captions for Supplementary Videos. ......................................................................................................................60
59 References .................................................................................................................................................................78
60

2
61 Supplementary Notes

62 Supplementary Note 1 | Image formation and the resolution limit to the Wiener

63 deconvolution.
64 The image formation of an aberration-free optical system is governed by simple propagation models of
65 electromagnetic waves interacting with geometrically ideal surfaces of lenses and mirrors. Under such
66 conditions, the process can be described as a two-dimensional Fourier transformation1,
67 G (u , v) = H (u , v)  F (u , v) , (1)
68 in which G is the two-dimensional Fourier transform of the recorded image, F is the transform of the intensity
69 distribution of the object that gave rise to the image, and H is the spatial frequency transfer function of the
70 optical system, the OTF. Because the total amount of light energy entering an optical system is limited by a
71 physically limiting pupil or aperture, the OTF is controlled by the aperture function under aberration-free
72 conditions. Therefore, the OTF H(u, v) can be calculated from the aperture function a(x, y) by
 
73 H (u, v) =   a( x, y )a( x + u, y + v)dxdy . (2)
− −

74 Therefore, the OTF is the autocorrelation of the aperture function, and x and y denote coordinates in the spatial
75 domain, while u and v represent coordinates in the spatial frequency domain. The transmission of
76 electromagnetic waves within the aperture is perfect, with a(x, y) = 1, while outside of the aperture, no wave
77 can propagate and a(x, y) = 0. Therefore, the OTF changes to zero outside of a boundary defined by the
78 autocorrelation of the aperture function. Because no spatial frequency information from the object outside of
79 the OTF support can be transmitted to the recording device, the optical system itself is always bandwidth
80 limited1.
81 To recover the information of the original object from the recorded image, Eq. (1) can be rearranged into
82 the following form:
G (u, v)
83 F (u, v) = . (3)
H (u, v)
84 In the preceding equation, it is obvious that the retrieved object spatial frequency exists only where H(u, v)
85 ≠ 0, and division by null outside of the region of OTF support is impossible. This is why diffraction limits
86 the image resolution and causes the general belief that the information lost due to the cut-off of system OTF
87 cannot be retrieved by mathematical inversion in the frequency domain.

3
 88 Supplementary Note 2 | Computational superresolution.
89 Richardson and Lucy2, 3 (RL) deconvolution, followed later by Shepp and Vardi4 regarded the problem of
90 recovering the underlying image blurred by a known PSF as one of statistical estimation rather than an exercise
91 in direct solution. They proposed an iterative deconvolution method based on the Bayes' theorem on
92 conditional probabilities, and the following equation gives the basic algorithmic computation:
g ( x, y )
93 f n +1 ( x, y ) = f n (x, y)[( ) h( x, y )] . (4)
h ( x, y )  f n ( x, y )
94 For image data g(x, y) and the PSF h(x,y), the object at iteration n+1 is calculated from its estimation at n. The
95 starting approximation f 0(x, y) should be a smooth, nonnegative function having the same integrated intensity
96 as the observed image, thus adding a nonnegative a priori knowledge to the approximation. From this equation,
g ( x, y )
97 it is obvious that a large deviation of from unity on a large length scale compared to that
h ( x, y )  f n ( x, y )
n+1
98 of h(x,y) would result in a large correction to yield the new f (x, y), while small deviations lead to small
99 corrections3. Thus the algorithm nonlinearly approximates the maximum-likelihood estimation of the
100 observed sample.
101 Unlike direct inversion in the frequency domain as Wiener deconvolution in Eq. (3), this computational
102 process in Eq. (4) is conducted in the spatial domain, avoiding the problem of division by null outside of the
103 region of OTF support. In other words, instead of the direct inversion of the low-pass filter H(u, v) in the
104 frequency domain (G(u, v) = H(u, v) × F(u, v)), the RL deconvolution iteratively approximates with its
105 inverse Fourier transform, i.e., the convolution kernel h(x, y) in the spatial domain (g(x, y) = h(x, y)  f
106 (x, y)). In principle, the extension beyond the diffraction limit is possible. It may occur for two reasons: the
107 nonlinearity of the product of f n with the other quantities in the square brackets and the prior information of
108 nonnegative values of the object.
109 Because RL deconvolution will quickly converge to a solution dominated by the noise5, it is recommended
110 to stop the reconstruction after a small number of iterations in practice. However, if the spatial frequency is
111 sufficiently enormous, many iterations will be necessary to extract all the information6. To test this idea, we
112 have synthesized a ground-truth, noise-free image containing various-shaped structures. After the image was
113 convolved with a Gaussian PSF (FWHM of ~90 nm), we manually set spatial frequency beyond 90 nm to zero,
114 which trimmed off much of the high-frequency information. Thus the blurred image resembled that obtained
115 by optical microscopy of limited bandwidths. Unexpectedly, parallel lines with the spacing of 60 nm can be
116 resolved after RL deconvolution for 2 × 104 iterations, whereas their intensities also became more uniform
117 upon even more iterations (Supplementary Fig. 1). Therefore, our simulation confirmed that many iterations
118 of RL deconvolution extended bandwidth in synthetic images without noise, thus recovering the missing
119 information7. This simulation echoed the mathematical bandwidth extrapolation idea first proposed by Wolter8,
120 and later summarized in Goodman's book9. For an optical system, it follows that if an object has a finite size,
4
121 a unique analytic function exists that coincides inside G(u, v). By extrapolating the observed spectrum using
122 analytic continuation10 or the Gerchberg algorithm1, it is possible in principle to reconstruct the object with
123 arbitrary precision under noise-free conditions.
124 Despite the theoretical feasibility of mathematical SR, out-of-band extrapolation is usually unstable due
125 to noise in data. The noise in actual experiments may not be an analytic function, thus allowing no solution to
126 the bandwidth extrapolation based on analytic continuation10. From the perspective of information capacity,
127 bandwidth extrapolation based on analytic continuation extends the spatial bandwidth at the price of the signal-
128 to-noise ratio (SNR). For object fields containing large space-bandwidth products, even a small extension of
129 spatial bandwidth leads to greatly amplified noise in the final reconstruction11, 12.
130 To demonstrate possible effects of noise on RL deconvolution, we simulated parallel lines blurred with
131 PSF (FWHM of ~90 nm), spacing from 10 nm to 100 nm, and increasing at the interval of 10 nm
132 (Supplementary Fig. 2). Without noise, 2000 iterations of RL deconvolution correctly resolved lines 10 nm
133 apart (Supplementary Fig. 2c); in contrast, adding the Poisson noise rendered excessive RL iterations
134 improper, thus parallel lines with a distance smaller than 70 nm could not be resolved without many artifacts.
135 The RL deconvolution became less effective in increasing resolution as more Gaussian noise was added to the
136 image (Supplementary Fig. 2, e1-e3). Therefore, classic RL deconvolution may not enhance the resolution
137 of images corrupted with noise such as those captured by real-world microscopes, which agrees with the
138 majority of the imaging community's belief-the Rayleigh diffraction limit represents a practical frontier that
139 cannot be overcome with a conventional imaging system9.

5
140
141 Supplementary Fig. 1 | Low-pass filtered image recovered by RL deconvolution under noise-free
142 condition. (a) The ground-truth contains synthetic various-shaped structures with a pixel size of 20 nm. (b)
143 The low-pass filtered (equal to 90 nm resolution) image of (a). (c) RL deconvolution result of (b) with 2×104
144 iterations. (d) RL deconvolution with 2×105 iterations. (e-h) Fourier transforms of (a-d). (i-l) Magnified views
145 of white boxes from (a-d). (m) The corresponding profiles are taken from the regions between the yellow
146 arrowheads in (i-l).

6
147
148 Supplementary Fig. 2 | The evolution of noise effects on parallel paired-line structures resolved by the
149 RL-deconvolution. (a-b) We created a ground-truth containing twelve parallel paired lines departed with
150 different distances (from 10 nm to 120 nm), and convoluted by a PSF with a 90 nm FWHM. (c) Without any
151 synthetic noise (and sufficient spatial sampling, 10 nm/pixel), the RL-deconvolution could distinguish paired
152 lines 10 nm apart. (d1-d4) We simulated images captured by real-world microscopes with large pixel sizes
153 and different noise levels. After downsampling the image in (b) by two folds (20 nm/pixel), we added Poisson
154 (d1), Poisson + 10% Gaussian (d2), Poisson + 50% Gaussian (d3), and Poisson + 100% Gaussian (d4) noise.
155 (e1-e4) Upon introduction of even weak noise, RL-deconvolution failed to resolve paired lines 70 nm apart.
156 In images corrupted with more than 50% Gaussian noise, the RL-deconvolution failed to improve the
157 resolution. We used 2000, 15, 15, 5, and 5 times of iterative deconvolution for images of different noise levels
158 (from 0% to 100%).

7
159 Supplementary Note 3 | The development of the sparse deconvolution model.
160 As explained in Supplementary Note 2, noise ultimately limited any possible extension in resolution by the
161 RL deconvolution. Under any fluorescence microscopes, images captured by the photoelectric sensor13 can be
162 regarded as the convolution of the object with the point spreading function (PSF) of the microscope. Thus a
163 fluorescent image can be expressed as:

164 image = N S  ( I (ω)  x(ω))  A(ω, θ)  , (5)

165 where ω is the spatial coordinate, and I, x, and A represent the illumination, the objective, and the PSF of the

166 microscope, respectively. The S and N denote the sampling and noise. The image is the final collected image

167 of the microscope, and θ is the aberration in the microscope.

168 In real optical instruments, deviations may arise from the Poisson noise intrinsic to the observation, the
169 inevitable read-out noise of the hardware14-18, the existence of aberration in the optical path19, and the effect
170 of finite sampling20, 21. To accurately deconvolve the object from images corrupted with noise always confers
171 an ill-posed inverse problem. Therefore, by building up a constraint model using the prior knowledge widely
172 shared by fluorescent microscopes, we wanted to decode the fluorescent signals as close to their ground-truth
173 positions as possible. Taken an example of single-molecule localization microscopy (SMLM), it achieves
174 super-resolution down to 10~20 nm by assuming each isolated fluorescence puncta stemmed from a single
175 molecule22, 23. However, such strong prior knowledge needs specifically designed instruments and fluorescent
176 probes, which may not apply to conventional microscopes. Therefore, we introduce two pieces of prior
177 knowledge, i.e., sparsity and continuity, which are common features of fluorescence generation that can be
178 used in our reconstruction model to achieve super-resolution.
179 Continuity in xy-t(z) axes:
180 For any microscope systems, to ensure sufficient Nyquist sampling criteria for maximal spatial resolution
181 dictated by the optics, their PSF must occupy more than 3 × 3 pixels in space (Supplementary Fig. 3), which
182 constitutes the basis for the continuity along x and y axes of any fluorescence microscope24. Moreover, if the
183 distance that the biological sample moved between two consecutive time points was smaller than the
184 microscope's lateral resolution, structures were considered to be continuous along the t axis. Given that cells
185 may demonstrate high-speed dynamics sometimes (such as Ca2+ diffusion within live cells, Extended Data
186 Fig. 17), the parameter for the continuity along t axis can be turned off under those conditions. As we have
187 demonstrated earlier, the application of continuity constraint substantially suppresses reconstruction artifacts
188 caused by Gaussian-type random noise24.

8
189 Sparsity:
190 Although the application of continuity a priori can increase the SNR, it also obscures the image and reduces
191 the resolution. For any fluorescence microscopes, an increase in spatial resolution always leads to smaller PSF.
192 Thus, as compared with a conventional microscope, the convolution of the object with the smaller PSF in
193 super-resolution images always confers relative sparsity. Therefore, we propose introducing the sparsity as
194 another prior knowledge to extract the fluorescent object's high-frequency information.

195 Absolute sparsity ( 0 norm, the total number of non-zero elements) has been linked with compressive

196 sensing (CS), which has been successfully applied to different fields of signal processing25-27. If the signal is
197 sparse (mostly zeros) or sparse after some transformations, CS can precisely recover the signal from the highly

198 noisy or corrupted measurements25. However, instead of 0 norm (only absolute sparsity) used as the start point

199 in CS theory, we directly used the 1 norm (sparsity score in Supplementary Fig. 3a), i.e., the sum of absolute

200 values of all elements.

201 We simulated an array of sparsely distributed fluorescence molecules and the diffusive square
202 fluorescence structures without fluorescence at the array complementary to the previous structure, and
203 convolve these ground truth structures with PSFs of variable sizes to mimic images detected by microscopes
204 of different spatial resolutions. No matter whether these structures were absolutely sparse or not, we showed
205 that the sparsity scores of images of the high resolution were always smaller than those of low resolution
206 (Supplementary Fig. 3a, 3b). Therefore, our sparsity term can handle both absolutely and relatively sparse
207 structures and constrain the extraction of high spatial frequency information, as shown in numerous examples
208 listed in Supplementary Fig.4.
209 In summary, the continuity means positively correlated fluorescence among neighboring 2×2 or 3×3
210 pixels, which equals the size of PSF of the microscope under the efficient sampling condition required by the
211 Nyquist sampling theorem. On the other hand, any captured microscope images are always denser than the
212 real object because of the convolution of the object with the system PSF, and high resolution always means
213 an increase in relative sparsity (Supplementary Fig. 3c). Therefore, we believe that continuity and sparsity
214 are general features of the fluorescence microscope, which could be used as the prior knowledge to suppress
215 noise and facilitate high-frequency information extraction collaboratively (Supplementary Fig. 3c-3d).
216 The sparse deconvolution model:
217 Therefore, we have proposed the following loss function containing these two priors, in which the Hessian
218 matrix continuity to reduce artifacts and increase robustness at the price of reduced resolution, and the sparsity

9
219 to balance the extraction of high-frequency information, which gives:
 
arg min  f − b − Ax +R Hessian (x)+L1 x 1  .
2
220 (6)
x ,b 2 2

221 The first term on the left side of the equation is the fidelity term, representing the distance between the
222 recovered image x and the result obtained with the Wiener filtered SIM image f. A is the PSF of the optical
223 system. Here, b is the background, which is assumed to be smooth, varying in the whole field-of-view. We
224 involved such background b in Eq. (6), considering that the fluorescent images may contain strong
225 background components from the out-of-focus emission light, fluorescent background, and cellular
226 autofluorescence. We estimated and removed such background fluorescence, which is critical for the high-
227 fidelity reconstruction of objects inside the cell. The second and third terms are the continuity and sparsity

228 priors, respectively.  1 and  2 are the 1 and 2 norms, respectively, and λ and λL1 denote weight factors to

229 balance image fidelity with sparsity. The continuity prior regarding the structural continuity along the xy-t(z)
230 axes is defined as follows:

g xx g xy t g xt
RHessian (g ) = g yx g yy t g yt
231 . (7)
t g tx t g ty t g tt
1

= g xx 1 + g yy + t g tt 1 + 2 g xy + 2 t g xt 1 + 2 t g yt
1 1 1

232 In Eq. (7), λt denotes the regularization parameters that present the continuity along the t axis, which can be
233 turned off upon imaging objects moving at high speed. Besides, we split the prior knowledge constrain and
234 deconvolution subproblems into a two-step operation to gradually and effectively search the final solution x.
235 The full three-step is given as follow:
236 Step. 0 (optional): Because not all of the samples may have the background (listed in Supplementary
237 Table 4), we estimate the background b in Eq. (6) by an optional isolate step before iterative optimization.
238 We used a modified iterative wavelet transform method (lowest frequency wavelet bands of f) to estimate b28.
239 Step. 1: We solve the Eq. (6) gradually with the following Eqs. (8) and (9). The detailed procedure of Step.
240 1 can be found in Supplementary Note 4.1.

 
arg min  f − b − g 2 +R Hessian (g)+L1 g 1  .
2
241 (8)
g 2 

242 Step. 2: Iterative deconvolution to achieve final solution x. The detailed procedure can be found in
243 Supplementary Note 4.2.

244
x

arg min g − Ax
2
2 . (9)

10
245 Because the pixel size reconstructed may not match with the improved resolution, we introduce a pre
246 upsampling step (optional, marked as ×2) to reconstruct the structure on smaller grids before the execution of
247 Step. 1. The detailed notation is provided in Supplementary Note 4.4.

11
248
249 Supplementary Fig. 3 | The concept of sparse deconvolution. (a, b) Specific examples for absolute (a) and
250 relative (b) sparsity. (c) Visualization of the working process of our sparse deconvolution. (d) Reconstructions
251 with only sparsity, only continuity, or both the sparsity and continuity priors. Actin filaments labeled with
252 LifeAct-EGFP in a live COS-7 cell in Fig. 4a under 2D-SIM (left, top panel) and CCPs labeled by Clathrin-
253 EGFP in Fig. 5a under SD-SIM (left, bottom panel), followed by sparsity-constraint deconvolution (middle
254 left), continuity-constraint deconvolution (middle right), or both sparsity and continuity-constraint
255 deconvolution (right). Scale bars: 500 nm.
12
256
257 Supplementary Fig. 4 | Different decreases in non-zero ratios of images after the sparse deconvolution.
258 We collected the number of non-zero values of images of this paper before and after our sparse deconvolution
259 (details also in Supplementary Table 4). To demonstrate these numbers more clearly, they are normalized by
260 the total pixel number of images. The resulting ratios could be clustered by the K-means method29 into three
261 categories in different colors (green, red, and blue). Each scatter denotes a non-zero ratio of one frame before
262 (x-axis) and after (y-axis) the sparse deconvolution.

13
263 Supplementary Note 4 | The details of algorithm execution.

264 Supplementary Note 4.1 | Iterative approximation based on the weighted constraints.

265 Searching for the best solution of the loss function in Eq. (8) can be translated into a convex optimization
266 problem (COP). To do that, we adopted the split Bregman algorithm widely used in total variance (TV)
267 problems30 due to its fast convergence speed. Using Bregman splitting, we first replaced variable g with the
268 intermediate variable u. Then, the loss function could be transformed into a constrained minimization problem:

 
 2 f − b − g 2 + u xx 1 + u yy 1 + utt
2
  
min  f − b − g 2 +  (u)  = min  , (10)
2 1
269
g ,u
2  g ,u  + u xy + u xt 1 + u yt + u 1 
 1 1 

270 in which u xx = g xx , u yy = g yy , u tt =  t  g zz , u xy = 2  g xy , u xt = 2  t  g xt , u yt = 2  t  g yt ,u = L1  g .

271 By using the Lagrange multiplier method to enforce the constraints weakly, we obtained the following
272 unconstrained problem:

  2 2

  u xx − g xx 2 + u yy − g yy 2 + utt − t  g tt 2 + u xy − 2  g xy
2 2
  ,
min  f − b − g 2 +  (u) +  
2 2
273  (11)
2 2 2
2 2 + u xt − 2 t  g xt + u yt − 2 t  g yt + u − L1  g 
g ,u

 
 2 2

2

274 where μ is the Lagrange multiplier. Finally, we strictly enforced the constraints by applying the simplified
275 Bregman iteration, which was proved by Goldstein and Osher 31:
  u − g − v 2 + u − g − v 2 + u −  g − v 2 
  xx xx xx 2 yy yy yy 2 tt t tt tt 2 
    2 2  
min  f − b − g 2 +  (u) +   + u xy − 2  g xy − v xy + u xt − 2 t  g xt − v xt +
2
276  , (12)
2 2  2

g ,u 2

 
2
 u yt − 2 t  g yt − v yt + u − L1  g − v 2
  2 2  
277 where v is used to reduce the computational complexity in the iteration.
278 The iterative minimization procedure is given as follows:

  u − g − v 2 + u − g − v 2 + u −  g − v 2 
  xx xx xx 2 yy yy yy 2 tt t tt tt 2 
    2 2  
g k +1 = min  f − b − g 2 +   + u xy − 2  g xy − v xy + u xt − 2 t  g xt − v xt +
2
279  . (13)
2 2  2

g 2

 
2
 u yt − 2 t  g yt − v yt + u − L1  g − v 2
  2 2  

14
   2
u kxx+1 = min  u xx 1 +  u xx − g kxx+1 − v kxx 
u xx
 2 2

  2
u kyy+1 = min  u yy +  u yy − g kyy+1 − v kyy 
u yy
 1 2 2

  2
uttk +1 = min  utt 1 +  utt − t  g ttk +1 − v ttk 
utt
 2 2

  2
280 u kxy+1 = min  u xy +  u xy − z  g kxy+1 − v kxy  . (14)
u xy
 1 2 2

  2
u kxt+1 = min  u xt 1 +  u xt − 2 t  g kxt+1 − v kxt 
u xt
 2 2

  2
u kyt+1 = min  u yt +  u yt − 2 t  g kyt+1 − v kyt 
u yt
 1 2 2

  2
u k +1 = min  u 1 +  u − L1  g k +1 − v k 
u
 2 2

v kxx+1 = vxxk + g kxx − u xxk
v kyy+1 = v kyy + g kyy − u kyy
v ttk +1 = v ttk + t  g ttk − uttk
281 v kxy+1 = v kxy + 2  g kxy − u kxy . (15)
v kxt+1 = v kxt + 2 t  g kxt − u kxt
v kyt+1 = v kyt + 2 t  g kyt − u kyt
v k +1 = v k + L1  g k − u k

282 Using Eqs. (13–15), the COP is split into three steps (smooth convex). Then, the final iterative steps can be
283 written as follows:

  
 FFT ( Lk sHe ) +  FFT ( f − b ) 

g k +1 = iFFT  
 FFT (  2 +  2 +  2  2 + 4  2 + 4  2 + 4  2 ) +  2 +  
 xx yy t tt xy t xt t yt L1
 , (16)
284  
Txx ( u kxx − v kxx ) + Tyy ( u kyy − v kyy ) + z Ttt ( u ttk − v ttk ) + 2 Txy ( u kxy − v kxy ) 
 
k
L sHe = 
+2 t  xt ( u xt − v xt ) + 2 t  yt ( u yt − v yt ) + L1 ( u − v ) 
T k k T k k k k

285 where the FFT (iFFT) denotes the (inverse) fast Fourier transforms,  xx is the two-order derivation operator

286 in the x-direction:  xx = 1, −2,1 , and  yy , tt ,  xy ,  xt ,  yt are similarly defined:

15
  k +1 1 k +1 1 
 g xx + v xx − , g xx + v xx   ,  
k k

   
  1 1
u kxx+1 =  0, g kxx+1 + v kxx   − , 
   
 1  1
g kxx+1 + v kxx + , g kxx+1 + v kxx   −, − 
   
 1
=shrink  g kxx+1 + v kxx , 
 
 1
u kyy+1 = shrink  g kyy+1 + v kyy , 
 
 1
uttk +1 = shrink  t  g ttk +1 + v ttk , 
 
 1
u kxy+1 = shrink  2  g kxy+1 + v kxy , 
 
 1
u kxt+1 = shrink  2 t  g kxt+1 + v kxt , 
 
 1 . (17)
287 u kyt+1 = shrink  2 t  g kyt+1 + v kyt , 
 
 1
u k +1 = shrink  L1  g k +1 + v k , 
 

288 Using Eqs. (15–17), the COP can easily be resolved.

289 Supplementary Note 4.2 | Iterative deconvolution.

290 To achieve the final solution x, we iterative minimize the Eq. (10). For the data with a high SNR, we applied
291 the Landweber (LW) deconvolution32, in which the gradient descent method with Nesterov momentum
292 acceleration was used33:
y ( j +1) = x ( j +1) +  k (x ( j ) − x ( j −1) )
x ( j +1) = y ( j +1) + AT (g − Ay ( j +1) )
293 1 , (18)
j=
1
2
( j −1 j −1 )
4 2
+  4 2 − 2
j −1

 j = 1 −  −1 j −1
294 where g is the image after sparsity reconstruction and xj+1 is the image after j+1 iterations. Iteration is
295 terminated at an early stage to constrain the resulting data to avoid artifacts34. The numbers of deconvolution
296 iterations are listed in the parameter table in the user manual of Supplementary Software.
297 For the low SNR data corrupted with excessive noise, we used a Bayesian-based Richardson Lucy (RL)2,
3
298 deconvolution algorithm to process these images. The accelerated RL-based deconvolution35 algorithm

16
299 follows:
 g 
y j +1 = x j   hT  
 hxj 
v j = x j +1 − y j
, (19)
300
 j +1
=
v j
 v j -1
v j −1
 v j -1
x j +1 = y j +1 +  j +1  ( y j +1 − y j )

301 where g is the image after sparsity reconstruction and xj+1 is the image after j+1 iterations. Specifically, the
302 adaptive acceleration factor (α) was introduced by Andrews et al.35 and represented the computational iteration
303 step length, which can be estimated directly from experimental results. The accelerated RL algorithm
304 convergences rapidly, and the required number of iterations 5~10 times less than that of the classical RL
305 method.

306 Supplementary Note 4.3 | Background estimation.

307 To estimate the background b appearing in Eq. (8), we modified the iterative wavelet transform method28. For
308 the data that contains weak background or even no background (e.g., TIRF imaging mode), we remove such
309 background estimate operation, i.e., set b as zero (corresponded to the No Background in the Supplementary
310 Table 4) to avoid removal of information. Images from low-dose illumination exhibited a low and stable
311 noise-like distribution of background fluorescence only. Thus, values over the mean value of the image were
312 set to zero directly, and the resulting residual image was used for the following background estimation. Under
313 other conditions, we estimated and removed the background as follows:
314 i. Original images from high-dose illumination containing a strong background fluorescence signal
315 originating from out-of-focus emissions and cellular autofluorescence were regarded as the input images in
316 the flowchart. The background can be iteratively estimated from the lowest frequency wavelet bands
317 associated with the input images (Wiener filter). To extract the lowest band of the frequency domain, we used
318 2D Daubechies-6 wavelet filters to decompose the signal up to the 7th level.
319 ii. To prevent accidental removal of minor useful signals, we performed an inverse wavelet transform on
320 the lowest band of the frequency information to the spatial domain and compared the result with half of the
321 square root of the input image. We merged these two images by keeping the minimum values at each pixel.
322 This operation removed high-intensity pixels resulting from inaccurate background estimation.
323 iii. The resulting low band off-peak background estimation data in ii were used again as the input image
324 in i, and the abovementioned loop was repeated several times. In general, we set the number of iterations

17
325 rounds to three to estimate the background with minimal contribution from the real fluorescent signals.

326 Supplementary Note 4.4 | Upsampling operation.

327 The pixel size could be the factor that limits SR image reconstruction. For example, although the Sparse SIM's
328 theoretical resolution could reach ~60 nm, the 32.5 nm pixel size may limit its resolution due to the insufficient
329 spatial sampling. Likewise, when we used the EMCCD camera for the SD-SIM, 94 nm pixel size severely
330 limited the resolution that could be achieved by the sparse deconvolution (down to ~90 nm). Therefore, we
331 introduced an upsampling operation in the reconstruction pipeline.
332 As a result, the loss function in Eq. (8) for the sparse reconstruction can be rewritten as follows:

 
arg min  f − b − Dg +R Hessian (g)+L1 g 1  ,
2
333 (20)
g 2 2


334 where D is the downsampling matrix. In practical, we used the Fourier interpolation to upsample the 2D-SIM
335 data (pad the zeros out of SIM OTF)36, and the spatial upsampling method (pad zeros around each pixel) to
336 upsample the SD-SIM data, which was less sensitive to the SNR and less prone to snowflake artifacts.
337

338
339 Supplementary Fig. 5 | Flowchart of background estimation based on multilevel wavelets.

18
340 Key messages:
Algorithm | Sparse deconvolution.
Input: time-lapse/volumetric image stack f
STEP 0.
if background estimation:
Background estimation using the procedure in Supplementary Note 4.3.
else:
b = 0.
Output: b
Input: f, b, A
STEP 1.
if upsampling:
Minimize Eq. (20):
 
arg min  f − b − Dg +R Hessian (g )+L1 g 1  .
2

g 2 2

else:
Minimize Eq. (8):

 
arg min  f − b − g +R Hessian (g)+L1 g 1  .
2

g 2 2

STEP 2.
Minimize Eq. (9):

x

arg min g − Ax
2
2 .
Output: x

19
341 Supplementary Note 5 | Using the sparse and continuity prior knowledge improves

342 resolution and contrast.

343 By adding the sparsity and the continuity prior knowledge to constrain the iterative deconvolution, we were
344 able to enhance the resolution of synthetic images of different structures (Extended Data Fig. 2, 3, 4), images
345 captured by the wide-field microscope (Extended Data Fig. 5), and by SIM in resolving the DNA origami
346 structures (Fig. 1), paired lines in the Argo slides (Fig. 2a-2b, Extended Data Fig. 6), fluorescence-labeled
347 nuclear pores (Fig. 2c-2g, Extended Data Fig. 11), actin mesh pores (Fig. 4a-4i), caveolae (Fig. 4j-4l), and
348 intricate intermediate fusion pores (Fig. 4o-4p). However, most of the structures resolved were of regular
349 shapes. To test whether we could use the sparse deconvolution to reconstruct erratic structures, we have
350 synthesized the bisected ring and irregular, periodic lines. As shown in Supplementary Fig. 6, sparse
351 deconvolution was able to resolve these structures correctly.
352 As the RL deconvolution failed to improve resolution in images corrupted with noise (Supplementary
353 Fig. 2), we wondered how SNR conditions affect our sparse deconvolution performance. Thus, we synthesized
354 parallel paired-lines from 10 nm to 120 nm apart, convolved them with the PSF with 90 nm FWHM, and
355 systematically added Poisson and Gaussian noise of different amplitudes (Supplementary Fig. 7). While the
356 RL deconvolution failed to improve the resolution, the sparse deconvolution consistently outperformed the
357 former in separating paired lines 50~60 nm apart. However, we also noticed that parallel lines under excessive
358 noise sometimes manifested as twisted and fluctuated fluorescence profiles (Supplementary Fig. 7),
359 indicative of artifacts. To quantitatively evaluate the quality of reconstruction, we have proposed a line
360 restoration quality (LRQ) criterion:

2  Avg (Region 0 )
361 LRQ = . (21)
Avg (Region1 ) + Avg (Region 2 )

362 The Region1, Region2, and Region0 represent the parallel lines and the region in between, and Avg calculates
363 the mean intensity of pixels within the corresponding area. Because the LRQ value always decayed rapidly
364 from a stable high plateau to 0.4 as the distance between paired lines decreased, we selected it as the threshold
365 of successful reconstruction (Supplementary Fig. 8). According to this criterion, RL deconvolution stopped
366 at resolving lines 100 nm apart at 50% Gaussian noise, while the sparse deconvolution could separate lines 50
367 nm apart. Increasing Gaussian noise limited the ability of the sparse deconvolution in improving
368 resolution. With 100% Gaussian noise, the sparse deconvolution could reach 60~70 nm resolution. With 150%
369 Gaussian noise, the sparse deconvolution could not faithfully extract information beyond the OTF

20
370 (Supplementary Fig. 8). However, compared to the RL deconvolution, it significantly increased the contrast
371 of high-frequency information within the OTF and enabled lines 100 nm apart to be separated
372 (Supplementary Fig. 7).
373 These data were also recapitulated in images of real samples taken by microscopes. Under the regular 200
374 mW illumination and 10 ms exposure (1 SNR), 2D-SIM, RL-SIM, and Sparse-SIM resolved fluorescent lines
375 carved in the commercial Argo-SIM slide 150, 120, and 60 nm apart, respectively (Extended Data Fig. 9).
376 Sparse-SIM resolution deteriorated to 90 nm under 25 mW illumination and 10 ms exposure (1/8 SNR) and
377 returned to the optical limit of ~120 nm under 25 mW illumination and 5 ms exposure (1/16 SNR). Despite
378 the failure to extend resolution at this extremely low SNR, sparse deconvolution was superior to the 2D-SIM
379 and RL-SIM in generating reconstructions of good contrast and minimal artifacts. Finally, in the live COS-7
380 cell weakly expressed with tubulin-EGFP, images captured by SD-SIM were of low SNR (Supplementary
381 Fig. 9). Under such a circumstance, the FRC resolutions of images under the SD-SIM and RL SD-SIM were
382 similar (~250 nm). In contrast, although Sparse SD-SIM did not increase resolution beyond the system OTF,
383 it increased the image contrast to ensure the maximal resolution allowed by the optics (~130 nm).
384 Key messages:
385 1) Reconstructing with the sparse and the continuity prior constrains the following RL deconvolution to
386 increase resolution in images with noise.
387 2) In images corrupted with excessive noise, the sparse deconvolution pipeline improves the contrast but not
388 the resolution.

21
389
390 Supplementary Fig. 6 | Partial ring/line simulation. (a) We created synthetic partial (left) and complete
391 (right) ring structures with an 80 nm diameter as ground-truth (top left). (b) We created four synthetic partial
392 lines as ground-truth (top left). Each partial line consists of multiple narrowed placed dots with a distance of
393 80 nm. The ground-truth in a and b were convolved with PSF with FWHM of either 110 nm (top right) or 45
394 nm (bottom left), and such image was subsequently subsampled 16 times (pixel sizes of 16 nm), and corrupted
395 with Poisson noise and 5% Gaussian noise. Images at the top right in a and b were sparse deconvolved (bottom
396 right).

22
397
398 Supplementary Fig. 7 | Effects of the combined Poisson-Gaussian noise on the paired lines resolved by
399 the RL and Sparse deconvolution. We synthesized a ground-truth image containing ten parallel paired lines
400 departed by different distances (from 10 nm to 100 nm), and convoluted it with a 90 nm FWHM PSF. We set
401 the maximum fluorescence intensities of lines to 9216 (a), 576 (b), and 36 (c) a.u., respectively, and added a
402 Poisson-distributed noise plus 10% (first column), 50% (second column), 100% (third column), and 150%
403 (fourth column) Gaussian distributed noise (compared to the maximum intensities of lines). After that, we
404 deconvolved these raw images (left, labeled as 'Raw') with RL (middle, labeled as 'RL') or Sparse
405 deconvolution (right, labeled as 'Sparse').

23
406
407 Supplementary Fig. 8 | The line restoration quality (LRQ) values of images in Supplementary Fig. 7.
408 Avg represents the averaged intensity of the pixels in the corresponding region. We use the value 0.4 (green
409 lines) as the threshold (discernment criterion) to determine whether the lines are separated successfully.

24
410
411 Supplementary Fig. 9 | Sparse deconvolution increases the contrast of microtubules of low SNR
412 captured by the SD-SIM. (a-c) A live COS-7 cell labeled with Tubulin-EGFP captured by SD-SIM. The raw
413 image was shown in (a), which was deconvolved by the RL (b) or Sparse deconvolution (c). (d) Magnified
414 views of white boxes in (a-c). The FRC resolution of raw SD-SIM, RL SD-SIM, and Sparse SD-SIM were
415 estimated to be 249 nm, 248 nm, and 136 nm, respectively. Scale bars: (a) 5 μm; (d) 3 μm.

25
416 Supplementary Note 6 | Parameters in the sparse deconvolution software.
417 In the Supplementary Software user interface (UI), we included thirteen parameters to adapt to different
418 hardware environments, experimental conditions, and fluorescence microscopes (Supplementary Fig. 10).
419 To simplify the usage of this software, we have classified them into three categories: fixed parameters, image
420 property parameters, and content-aware parameters.
421 The system hardware determines fixed parameters, and the image property parameters are associated
422 with image quality, such as high-or-low SNR and strong-or-weak background. Ten parameters in these two
423 categories are selected based on the system and the image property and need little tuning. The three parameters
424 left belong to Content-aware parameters. They need to be adjusted carefully to achieve the optimal
425 reconstruction results. We introduced a four-step workflow to finetune these content-aware parameters
426 (shown in Supplementary Fig. 10 and 14), which serves as a general guide. We provided the detailed step-
427 by-step illustration in Supplementary Note 8.
428 The parameters description:
429 Fixed parameters:
430 Pixel size: The physically equivalent pixel size of the final images.
431 Wavelength: The emission wavelength of the fluorescence probes.
432 Effective numerical aperture: Effective NA of the optical system based on its spatial resolution.
433 3D imaging: This is the option to choose whether the input images are volumetric or not.
434 GPU acceleration: This is the option to choose whether the calculation uses CUDA-GPU or not.
435 Image property parameters:
436 Background (Step 0): Based on the actual background fluorescence of the images processed, users can choose
437 no background or the background under high-dose (HI) or low-dose of illumination (LI). Both the types HI
438 and LI have weak and strong magnitude options. Thus five options of background are: No, Weak-HI, Strong-
439 HI, Weak-LI, and Strong-LI. We want to emphasize that the background estimation step is an optional
440 operation that may not be necessary. For example, when we processed Ca2+ imaging data in Extended Data
441 Fig. 17, we selected no options to avoid removing the baseline signal. We have also provided parameters used
442 for processing images in this manuscript for reference in Supplementary Table 4.
443 Upsampling (Step 0): In conditions of inadequate Nyquist sampling, we manually up-sample images to
444 achieve the theoretical resolution increase posed by the sparse deconvolution. We usually choose the spatial
445 upsampling method for low-SNR images and the Fourier upsampling method for high-SNR ones.
446 t (z)-axial continuity (Step 1): This parameter is for adjusting the continuity along the input dataset's t or z-
26
447 axis. We set the t (z)-axial continuity less than or equal to 1 to avoid temporal blurring. For the fast time-lapse
448 imaging, while the fidelity is less than 100, the t (z)-axial continuity is usually assigned as one-hundredth of
449 the fidelity. If the object being imaged has undergone fast movements, we need to set this parameter to a small
450 number (0.1) or even zero to avoid causing motion artifacts.
451 Sparse iteration times (Step 1): Usually, we set it to 100 iterations. If spatial over-sampling is used, we need
452 to increase the number to 200 or 300 to ensure the sparsity reconstruction convergence.
453 Iterative deconvolution (Step 2): We usually choose the Richardson-Lucy algorithm (RL) to deconvolve
454 low-SNR images and the LandWeber deconvolution (LW) to deconvolve high-SNR ones.
455 Content-aware parameters:
456 Image fidelity (Step 1): This parameter denotes the distance between the image before and after the sparse
457 reconstruction and is the inverse of the xy continuity. Usually, we use a large value (1000~300) for high SNR
458 images.
459 Sparsity (Step 1): This parameter represents the relative sparsity constraint enforced on the reconstruction
460 (Step 1). Usually, we pre-set this value to one-tenth of the image fidelity term (Supplementary Fig. 5).
461 However, an unnecessary high sparsity value may remove weak signals. Thus we need to finetune this
462 parameter back-and-forth based on the final deconvolution result.
463 Iterative deconvolution times (Step 2): This parameter sets the times for the post iteration deconvolution.
464 Because the LW method is slower than the vector extrapolation version of the RL method in reaching
465 convergence, we choose 5~15 iteration times for the RL algorithm and 30~50 for the LW algorithm.
466

467 Key messages:

468 The ten parameters in the first two categories are primarily determined by the optical system and image
469 property and need little tuning. Only the three content-aware parameters need to be adjusted back-and-forth
470 carefully by visual examination of the reconstruction results.

27
471
472 Supplementary Fig. 10 | The sparse deconvolution software. The top panel shows the software UI, and the
473 bottom panel shows the simplified workflow of parameters adjustment. The detailed workflow is shown in
474 Supplementary Fig. 14.

28
475 Supplementary Note 7 | Optimal values of the sparsity and the fidelity under different

476 conditions.
477 Among content-aware parameters, we usually set the iterative deconvolution time as a constant of 10, 50 for
478 the RL, LW deconvolution method, respectively. As we used the fidelity term (1/continuity) to adjust the
479 weight of continuity in the software, we only need to finetune the sparsity and the fidelity parameters in
480 practice.
481 By collecting all fidelity and sparsity values carefully chosen for images used in this work, we find that
482 the sparsity value roughly follows a linear relationship with the fidelity one (1:5~1:20, Supplementary Fig.
483 11-13). Although we do not understand the underlying mechanism, it represents the suitable range for
484 processing different image datasets.
485 Because we want to improve the resolution in deconvolving high SNR images and the contrast in
486 deconvolving low SNR images, these two parameters needed to be adjusted carefully. To explore the effects
487 of different sparsity and fidelity values on reconstruction results, we synthesized ring-shaped and punctated
488 structures of 80 nm in diameters, convoluted them with a PSF of 110 nm in full width at half maximum
489 (FWHM) (Supplementary Fig. 12a), and corrupted them with the noise of various amplitudes
490 (Supplementary Fig. 12b). With 2% noise, sparse deconvolution resolved both the ring-shaped and the
491 punctuated structures with high fidelity and sparsity values, which approximately followed a 5:1 relationship
492 (1000:200, 950:200, 400:80). However, at low fidelity numbers (smaller or equal to 100), the fluorescent ring
493 was not resolved, indicating the failure to improve resolution (Supplementary Fig. 12d). With 10% noise,
494 the ring became irregular using the fidelity: sparsity ratio of 1000:200 or 950:200, while a balance of 400:80
495 was optimal for the resolution enhancement. As the SNR of the image deteriorated, it was evident that large
496 fidelity and sparsity values led to artifacts, while small numbers were better for the extraction of real structures
497 from noise. Under 80% noise, a fidelity: sparsity ratio of 20:5 allowed the ring to be resolved with sharp edges.
498 With 100% noise, we could only use small fidelity and sparsity values such as 10 and 2 to avoid artifacts.
499 Despite the failure to improve resolution, we observed reduced background and much-improved contrast,
500 which may help subsequent visualization, segmentation, tracking, and analysis.
501 Because the ground-truth for this synthetic image was known, we systematically examined different peak
502 signal-to-noise ratios (PSNRs) obtained under different sets of fidelity and sparsity values (Supplementary
503 Fig. 13a). If we selectively highlighted regions of PSNR larger than 25 (or 22.6 at 100% noise), the weight of
504 fidelity linearly correlated with that of the sparsity within a narrow range at most noise levels (Supplementary

29
505 Fig. 13b). These data reinforce the roughly linear relationship between optimal choices of the fidelity and
506 sparsity values.
507

508 Key messages:

509 1) We only need to adjust the values of fidelity and sparsity carefully.
510 2) The optimal values of fidelity and sparsity follow an approximately linear relationship.
511 3) Images of high SNR afford large fidelity values, while low SNR images require small fidelity values.
512

513
514 Supplementary Fig. 11 | Linear relationship between optimal values of sparsity and fidelity. We have
515 identified an approximately linear relationship of sparsity and fidelity (~1:5~1:20) from the reconstruction
516 results of SIM (blue) and SD-SIM (orange) in this work.

30
517
518 Supplementary Fig. 12 | Exploring optimal fidelity and sparsity choices for reconstructions of various
519 SNR images. (a) We synthesized ring-shaped and punctated structures of 80 nm in diameter as the ground-
520 truth (left), convoluted them with a PSF of either 110 nm (middle) or 45 nm (right) in FWHM. (b) The image
521 in the middle of (a) was subsequently subsampled 16 times (pixel sizes of 16 nm), and corrupted with Poisson
522 noise plus 2%, 5%, 10%, 20%, 50%, 80%, and 100% Gaussian noise. (c) The 7 ×7 table of the reconstruction
523 results. The column and row represent the noise amplitude and values of priors for the reconstruction,
524 respectively.

31
525
526 Supplementary Fig. 13 | PSNR heat maps at different values of sparsity and fidelity. (a) The maps of the
527 PSNR values (c.f., images under different noise amplitudes in Supplementary Fig. 12) at different values of
528 sparsity (row) and fidelity (column). (b) PSNR maps in (a) with values greater than 25 (2%, 5%, 10%, 20%,
529 50%, and 80% noise) or 22.6 (100% noise) at different conditions.

32
530 Supplementary Note 8 | Parameters finetuning
531 In this note, we intend to provide a comprehensive guide for the readers/users to process their data, including
532 the detailed procedures to adjust the Fidelity and Sparsity values (Supplementary Note 8.1) and some step-
533 by-step examples (Supplementary Note 8.2).

534 Supplementary Note 8.1 | Procedures for parameters finetuning.

535 STEP 0. Set the Fixed parameters and Image property parameters according to the microscope hardware,
536 computing system, and properties of images collected as described in Supplementary Note 6.
537 STEP 1. Choose the initial Fidelity value according to the image SNR. We recommend starting from 1000 for
538 high SNR images or 500 for low SNR ones. For starters, we also recommend choosing the Sparsity as one-
539 twentieth of the Fidelity, which is to avoid over-filtering artifacts due to the inappropriately high Sparsity
540 value.
541 STEP2. Adjust the Fidelity and Sparsity downwards simultaneously with the ratio maintained. Because the
542 Fidelity is the opposite of the xy continuity term, its reduction constrains the reconstruction with more weight
543 on the continuity. However, unnecessary small Fidelity values may over-smooth the structures and reduce the
544 resolution. Therefore, we need to examine the reconstruction results and terminate the downwards adjustment
545 once image over-smoothing is apparent.
546 STEP 3. Fine-tune Sparsity upwards. Large Sparsity values help us recover the high-frequency information
547 on the one hand but also may over-filter the weak signals on the other hand. Therefore, we increase the weight
548 of Sparsity until over-filtering artifacts emerge. Please note that adjusting the Sparsity value changes the ratio
549 of Fidelity and Sparsity.
550 STEP 4. Stop at the value of Sparsity with the right balance of resolution and minimally removing weak
551 fluorescence signals.

552 Supplementary Note 8.2 | Step-by-step examples.

553 Here, we provided two simulation examples and four experimental examples from imaging systems of 2D-
554 SIM (c.f., Fig. 4a), TIRF-SIM (c.f., Fig. 4j), and SD-SIM (c.f., Extended Data Fig. 15) microscopes,
555 respectively, to illustrate the four-step parameters adjustments procedure in practice.
556 Examples for the finetuning of the Fidelity and Sparsity values

557 Example 1. Sparse deconvolution of synthetic ring-shaped and punctuated structures corrupted with noise

558 of different amplitudes (Supplementary Fig. 15)

33
559 5% noise condition.

560 1) For synthetic ring-shaped and punctuated structures corrupted with 5% noise, we selected the Fidelity and
561 Sparsity as 1000 and 50 as recommended.
562 2) Reducing the Fidelity from 1000 to 800 and maintaining the Fidelity:Sparsity ratio did not significantly
563 change image quality. Therefore, we stopped the Fidelity at 1000.
564 3) Next, we increased the Sparsity while keeping the Fidelity stable. Once we raised the Sparsity to 150, the
565 ring-shaped structure emerged. However, when the Sparsity was adjusted to 250, the ring-shaped structure
566 became interrupted. Therefore, we selected 1000 and 200 as the optimal values for the Fidelity and
567 Sparsity for this image with low noise.
568 80% noise condition.
569 1) For synthetic ring-shaped and punctuated structures corrupted with 80% noise, we selected the Fidelity
570 and Sparsity as 500 and 50 as recommended.
571 2) At this ratio, irregularly sharpened structures were apparent. Therefore, we reduced the Fidelity while
572 maintaining the Fidelity:Sparsity ratio to suppress artifacts. The contrast of two fluorescence puncta
573 increased at the 20:1 ratio, while reducing the Fidelity to an even smaller value did not improve further.
574 3) Thus we fixed the Fidelity to 20, and increased the Sparsity term. Once we raised the Sparsity to 6, the
575 ring-shaped structure emerged, and even larger Sparsity led to over-sharpening artifacts. Therefore, we
576 stopped at 20 and 6 as the optimal values for the Fidelity and Sparsity for this image with high noise.

577 Example 2. Sparse deconvolution of dense actin filaments obtained by the 2D-SIM (Supplementary Fig.

578 16)

579 1) For dense actin structures under the 2D-SIM, we selected the Fidelity and the Sparsity as 1000 and 50 as
580 recommended. With these parameters, actin filaments were intermittent, and the snowflake-like artifacts
581 emerged. Therefore, we needed to reduce the Fidelity (increase the continuity) to reduce artifacts caused
582 by noise amplification.
583 2) We started with a large searching step for the general evaluation to enable rapid convergence and then
584 gradually reduced the step size for fine adjustment. By decreasing the Fidelity (from 1000 to 150) and
585 maintaining the Fidelity:Sparsity ratio, the snowflake-like artifacts were suppressed at the price of reduced
586 resolution. When we selected 150 and 7 for the Fidelity and Sparsity, respectively, actin filaments were
587 relatively continuous without much blurring.
588 3) Next, we kept the Fidelity to 150 and increased the Sparsity to further improve the resolution and contrast.

34
589 Once we selected a Sparsity value larger than 10, some weak actin filaments were trimmed off. Therefore,
590 we set the Fidelity value of 150 and the Sparsity value of 10.

591 Example 3. Sparse deconvolution of the caveolae obtained by the TIRF-SIM (Supplementary Fig. 17)

592 1) For caveolae under the TIRF-SIM, we selected the Fidelity and Sparsity as 1000 and 50 as recommended.
593 Interestingly, when we concomitantly reduced the Fidelity and the Sparsity values, the ring-shaped
594 caveolae were blurred. Thus we stopped the Fidelity value at 1000.
595 2) Next, we kept the Fidelity to 1000 and increased the Sparsity to further improve the resolution and contrast.
596 As we increased the Sparsity value, ring-shaped structures became more apparent. On the other hand, once
597 we selected a Sparsity value of 130, the caveolae ring became interrupted. Therefore, we stopped the
598 Sparsity value at 90 for optimal extraction of high-frequency caveolae structures.

599 Example 4. Sparse deconvolution of the actin filaments obtained by the SD-SIM (Supplementary Fig. 18)

600 1) Live-cell images obtained by our SD-SIM configurations (Supplementary Table 3) were usually of low
601 SNR. Therefore, we set the values of the Fidelity and Sparsity as 500 and 25, respectively.
602 2) At the initial recommended values of the Fidelity and Sparsity, actin filaments were intermittent and
603 corrupted with snowflake-like artifacts. By decreasing the Fidelity and maintaining the Fidelity: Sparsity
604 ratio, the snowflake-like artifacts were suppressed. When we selected 40 and 2 for the Fidelity and Sparsity,
605 respectively, actin filaments were relatively continuous without much blurring.
606 3) Next, we kept the Fidelity to 40 and increased the Sparsity to further improve the resolution and contrast.
607 However, once we increased the Sparsity value, some weak actin filaments were trimmed off. Therefore,
608 we set the optimal Fidelity and Sparsity values to 40 and 2.

609 Example 5. Sparse deconvolution of the CCPs obtained by the SD-SIM

610 1) For CCPs labeled by clathrin-DsRed and observed under the SD-SIM (c.f., Extended Data Fig. 15), we
611 needed to select the Strong background (HI) to remove the cytosol signal before sparse deconvolution
612 (Supplementary Fig. 19). It was apparent that our background estimation method was able to remove
613 different amplitudes of background fluorescence from different regions.
614 2) Upon selecting the Fidelity and Sparsity as 500 and 25, ring-shaped CCPs were oversharpened and
615 brightening single pixels (Supplementary Fig. 20). As we reduced the Fidelity and Sparsity
616 concomitantly, artifacts disappeared along with the blurring of rings. We selected 60 and 3 for the Fidelity
617 and Sparsity without unnecessary blurring.
618 3) Next, we kept the Fidelity to 60 and increased the Sparsity to improve the resolution and contrast further.

35
619 We stop at the optimal Fidelity and Sparsity values of 60 and 5, which reached a good balance between
620 resolving intricate signals and maintaining weak signals.

621 Supplementary Note 8.3 | More explanations on the background estimation

622 To demonstrate the effects of different background estimation choices, we chose cytosolic lysosomes from
623 Fig. 4m as the example. Upon selecting either No background, Weak-HI, or Strong-HI, it was apparent that
624 the vesicular structure remained largely unchanged while the signal-to-background ratio increased
625 substantially (Supplementary Fig. 21a).
626 Weakly fluorescent lysosomes at the upper left corner of the FOV, which was most likely due to the
627 microscope's uneven illumination, might be removed by selecting the Strong-HI parameter (Supplementary
628 Fig. 21b). Therefore, we could use the BaSiC correction37 method to correct the skewed intensity distribution
629 in the FOV so that obscure lysosomes with low intensities emerged in the image before the deconvolution
630 (Supplementary Fig. 21c). Subsequent sparse deconvolution revealed multiple lysosomes with the resolution
631 and contrast similar to lysosomes in the middle of the FOV (Supplementary Fig. 21d).
632 Finally, we also showed the effects of different background estimation choices on dense actin filaments
633 revealed by the sparse deconvolution (Supplementary Fig. 22). Choosing either Weak-HI or Strong-HI
634 removed the background without much affecting fluorescence profiles of filaments.

36
635
636 Supplementary Fig. 14 | The complete workflow of adjusting parameters in the sparse deconvolution
637 software in handling various fluorescence images.

37
638

639 Supplementary Fig. 15 | The process of selecting optimal parameters for the sparse deconvolution of
640 synthetic ring-shaped and punctuated structures corrupted with small (a) or large (b) noise (c.f.,
641 Supplementary Fig. 12).

38
642
643 Supplementary Fig. 16 | The process of selecting optimal parameters for the sparse deconvolution of
644 actin filaments obtained by the 2D-SIM.

39
645
646 Supplementary Fig. 17 | The process of selecting optimal parameters for the sparse deconvolution of
647 caveolae obtained by the TIRF-SIM.

40
648
649 Supplementary Fig. 18 | The process of selecting optimal parameters for the sparse deconvolution of
650 actin filaments obtained by the SD-SIM.

41
651
652 Supplementary Fig. 19 | Estimating background of different amplitudes at different regions. The raw
653 SD-SIM image (left) and the estimated background by using the Strong-HI parameter (right).

654
655 Supplementary Fig. 20 | The subsequent process of selecting optimal parameters for the sparse
656 deconvolution of CCPs obtained by the SD-SIM.

42
657
658 Supplementary Fig. 21 | Images reconstructed with different background estimations and the effect of
659 the BaSiC correction. (a) The 2D-SIM image (c.f., Fig. 4m) after reconstructed with No background, Weak-
660 HI, or Strong-HI. Scale bars: left, 3 μm; right, 500 nm. (b, c) The 2D-SIM image (c.f., Fig. 4m) before (b)
661 and after (c) BaSiC illumination correction. (d, e) The sparse deconvolution results of (b) and (c). Scale bars:
662 (a) left, 3 μm; right, 500 nm; (e) left, 5 μm; right, 1 μm.

663

664
665 Supplementary Fig. 22 | Actin filaments reconstructed with different background estimations (c.f.,
666 Extended Data Fig. 15). Scale bars: left, 3 μm; right, 500 nm.

43
667 Supplementary Note 9 | The calibration of nuclear pore diameters and the FWHM

668 resolution of Sparse SD-SIM.

669 Unexpectedly, we found that when the sizes of system PSF and the size of camera pixel were comparable to
670 the size of the structure been imaged, fitted diameters of punctated and ring-shaped structures differently
671 deviated from their real values. For large punctated fluorescence beads, the fitted FWHMs were larger than
672 the PSF of the microscope. For ring-shaped pores under the Sparse-SIM, their fitted diameters were smaller
673 than their real values. Therefore, we used simulations to derive correction factors under different situations.

674 Supplementary Note 9.1 | Calibrating fitted diameters of ring-shaped pores.

675 To examine the relationship between fitting diameter and the real diameter, we created synthetic ring structures
676 (1 nm sampling) with different diameters (60, 70, 80, 90, 100, 110, 120 nm). These ground truth structures are
677 then convolved with different sized PSF (10, 20, 30, 40, 50 nm FWHMPSF). The results are subsampled by 16
678 times to match the experimental condition (Supplementary Fig. 23a). Resultant images are fitted with a
679 double-peak Gaussian function to extract the fitted diameters (Supplementary Fig. 23a). As can be seen in
680 Supplementary Fig. 23b, with 1 nm pixel size and no subsampling, the fitted diameter (FD) was smaller than
681 the real diameter (RD) due to the convolution of system PSF (FWHMPSF) with the ring, which could be
682 approximated by the equation:

683 RD = FD 2 + FWHM PSF

2
. (22)

684 Interestingly, if the pixel was subsampled to be 16 nm per pixel, the apparent FD was further reduced by
685 the aliasing effect of a large pixel. Under such circumstance, the relationship between FD and RD could be
686 approximated by a linear function:
687 RD = a  FD + b . (23)
688 Therefore, we used a linear regression with parameters (a = 0.9335, b = 22.8085) to correct for the nuclear
689 pore measurement in Sparse-SIM 2× configuration (50 nm resolution, 16 nm pixel size). To correct the pore
690 size measured by TIRF-SIM (90 nm resolution, 32 nm pixel size), we used another set of parameters (a =
691 0.7740, b = 62.9420).

692 Supplementary Note 9.2 | The bead correction factor.

693 In our experiments, the diameters of fluorescent beads (100 nm) were comparable in scale to the spatial
694 resolution of the Sparse SD-SIM (~90 nm). Under such circumstances, the measured FWHMs were also
695 affected by the size of the beads, which did not accurately report the actual microscope resolution. Therefore,

44
696 we intend to use the 'beads correction factor' equation derived in the PSFj38 to correct for such effects (details
697 in PSFj), and retrieving the real resolution of the optic system (FWHMPSF). In addition, per personal
698 communication with the authors of the PSFj, we also tried to simulate a fluorescent bead to be a 3D sphere
699 with constant fluorophore concentration over the bead volume, in which gives:
C , 2 x 2 + y 2 + z 2  d
700 f ( x, y , z ) =  0 bead
. (24)
 0, otherwise
701 Since there is no analytical solution for the convolution of 3D sphere bead and PSF, we solved this
702 convolution numerically with the specific experiment parameters as shown in Supplementary Fig. 24. In our
703 simulations, the 100 nm 3D sphere was convolved with a 3D PSF (FWHMPSF of 88 nm × 88 nm × 260 nm),
704 and subsequently subsampled to match the experimental condition (the pixel size of 38 nm). The resulting
705 image had a measured bead FWHM as 110 nm, which was the same as our experimental fitted result 110 nm.
706 For a system with higher or lower spatial resolution (FWHMPSF: 80 nm or 100 nm) would lead to measured
707 bead FWHM varied between 102 nm and 120 nm. Therefore, we concluded that the actual FWHM PSF of our
708 Sparse SD-SIM should be approximated 90 nm.

45
709
710 Supplementary Fig. 23 | Relationships between fitted and real diameters of ring structures of various
711 sizes as observed under microscopes of different resolutions and spatial sampling frequencies. (a) The
712 simulation workflow. A representative example of ring structures with different diameters was convolved with
713 the PSF with FWHM of either 10 nm (upper panels) or 50 nm (bottom panels). These images were shown
714 without (pixel size of 1 nm, left) and with spatial subsampling (pixel size of 16 nm, right). The histogram in
715 the left-bottom corner showed examples of fitted diameters (FD) of the 80 nm ring obtained under different
716 conditions. (b) Linear relationships between fitted and real diameters (RD) under 1 nm (left) or 16 nm (right)
717 spatial sampling. Different colors indicate relationships obtained with microscopes of different spatial
718 resolution (FWHMPSF). With sufficient spatial sampling, RD could be derived from the fitted result using the

equation: RD = FD + FWHM PSF . When the pixel size of the final image was comparable to the FWHMPSF
2 2
719

720 and the diameter of the ring, the relationships were approximated by a linear function offset by a constant,

721 RD = a  FD + b (a = 0.9335, b = 22.8085).

46
722
723 Supplementary Fig. 24 | A similar size of the synthetic bead after convolution with the PSF of the system
724 as compared with that detected in the real experiment. (a) The 3D rendering of a synthetic 100 nm bead
725 in three dimensions. (b) The 2D image of the bead convolved with a 3D PSF with the FWHMxyz of 88 nm ×
726 88 nm × 260 nm (1 nm pixel size). (c) The image in (b) was downsampled by 38 times. (d) The experimental
727 image of 100 nm fluorescent bead (c.f., Extended Data Fig. 14a), which was nearly identical to the simulated
728 image in (c).

47
729 Supplementary Note 10 | Other sparsity-based reconstructions that improve spatial

730 resolution in theory.

731 In this part, we implemented two previous reconstruction procedures based on the sparsity concept, FISTA
732 (Fast Iterative Shrinkage-Thresholding Algorithm)39, and NLHT (Non-Local Hard Thresholding)26, 40

733 algorithms, which have been proposed to extend spatial resolution under coherent imaging of very sparse
734 samples.

735 Supplementary Note 10.1 | FISTA method.

736 Similar to bandwidth interpolation methods that were based on compressive sensing to reduce the acquisition
737 time26, 40, FISTA reconstructs with l1 constraint that is mathematically expressed as:

738
x x

arg min  f ( x) +  x 1 = arg min Ax − b +  x 1 ,
2
2
 (25)

739 where the first term on the left side is the fidelity term, representing the distance between recovered image x
740 and the image obtained after the Wiener filtered result b. A is the point spread function (PSF) of the imaging

741 system. The second term represents the sparsity prior,  1 and  2 are the l1 and l2 norms, respectively. λ

742 denotes the weight factor, balancing the images fidelity and the sparsity prior.
743 We reconstructed the SR images with the sparsity prior based on Eq. (25), a convex optimization problem
744 resolved with the FISTA (with backtracking) algorithm39.

For the simplicity of derivation, we termed f (x) = Αx − b

2
745 . First, consider the unconstrained
2

746 minimization problem of a continuously differentiable function f (x) :

747 min{ f (x)}. (26)

748 The simplest methods to solve Eq. (26) is the gradient descent method:

749 xk = xk −1 − tk f (xk −1 ), (27)

750 where tk  0 is a suitable step distance. The gradient iteration Eq. (27) can be regarded as an extremal solution

751 of a quadratic equation (proximal regularization of the linearized function f (x) at x k −1 ), which could be

752 rewritten in the equivalent form as

 1 
753 x k = arg min  f (x k −1 ) +  x − x k −1 , f (x k −1 ) + x − x k −1 2  , (28)
x  2tk 

754 where ,  denotes the inner product of two vectors.

48
755 Adopting this same basic gradient solution to the non-smooth l1 regularized problem equal to Eq. (25)

 1 
756 x k = arg min  f (x k −1 ) +  x − x k −1 , f (x k −1 ) + x − x k −1 2 +  x 1  . (29)
x  2tk 
757 After removing constant terms, the equation can be rewritten as

 1 
x − ( (x k −1 − tk f (x k −1 ) ) +  x 1  .
2
758 x k = arg min  (30)
x  2tk 

759 For any L > 0, considering the quadratic approximation of F ( x)  f ( x) +  x 1 at a given point x k −1 , we define

760 QL (x, xk −1 ) as:

L
761 QL (x, x k −1 )  f (x k −1 ) +  x − x k −1 , f (x k −1 ) + x − x k −1 2 + x 1 , (31)
2

762 where L is Lipschitz constant of smooth convex function f. Through formula derivation, we rewrite the formula
763 above as follows:

L  1 
2


764 x k = arg min  x −  (x k −1 − f (x k −1 )  +  x 1  . (32)
x

 2  L  


765 where Eq. (32) can be simply interpreted that the constant terms tk of Eq. (30) are replaced by Lipschitz

766 constant terms L to ensure the convergence. Based on Eq. (32), we defined a unique minimization operator

767 pL ( y) as:


 L  1 
2


768 pL (x k −1 )  arg min QL (x k , x k −1 ) = arg min  g (x) + x −  (x k −1 − f (x k −1 )  . (33)
x x

 2  L  


1
769 Considering g (x) = x 1 , we can use the soft-threshold algorithm (1) to solve pL ( y ) , and let u = y − f ( y ) ,
L


770 a= , the expression is given as,
L

771 soft (u, a)  sign(u ) max{ u − a,0}. (34)

772 As a result, the complete steps of the FISTA algorithm can be concluded as follows:

Step 0. Let L0 > 0,  > 1, and x0  . Take y1 = x0 , m1 = 1 .

n
773

Step k. ( k  1 ) search the smallest nonnegative integers ik subject to L =  k Lk -1

i
774

775 F ( pL (y k ))  QL ( pL (y k ), y k ). (35)

49
 Let L =  k Lk -1 and proceed
i
776

1 
777 xk = soft ( yk − f ( yk ), ). (36)
L L

778 Combine the Nesterov33 method to accelerate gradient update:

1 + 1 + 4mk2
779 mk +1 = . (37)
2

 m −1 
780 yk +1 = xk +  k  ( xk − xk −1 ). (38)
 mk +1 

781 Supplementary Note 10.2 | NLHT method.

782 Adding some constraints to the FISTA algorithm yields the Non-Local hard Thresholding algorithm (NLHT)40.
783 Initially, the results of l1 constraint are solved by the Basis Pursuit De-Noising algorithm (BPDN), in which
784 the pixels with values close to zero are located by performing a nonlocal thresholding step. Each element of x
785 will be zeroed out along with its neighbors when its value is below a fixed threshold. The BPDN step is
786 repeated with the additional constraint that the locations corresponding to the off-support are set to zero40.
787 Because of the slow convergence speed, we replaced the BPDN with the FISTA to solve the convex
788 optimization problem and conducted non-local hard thresholding. A more detailed version of the algorithm is
789 provided as follows:
790 Step 1: Use FISTA to solve

791
x
 2

arg min Ax − b +  x 1 , x  j  = 0, j  S .
2
(39)

792 where S is the index set of the zero value.

793 Step 2: Search all j  such that x  j    max( x) which are distanced from j  to the right or left by 1 pixel.

794 Then add j  to S  and set S = S S  . If the index set S was not updated, increase  by  . It is worth denoting

795 that,  and  are threshold and increment in the threshold respectively.

796 Comparison of the reconstruction results using Sparse-SIM based algorithm with FISTA and NLHT are
797 shown in Supplementary Fig. 25. It is worth noticing that, to avoid over-filtering structural information by
798 the NLHT (Supplementary Fig. 25c, pointed by yellow arrows), the iteration of these two steps needed to be
799 terminated after several rounds.

50
800
801 Supplementary Fig. 25 | Comparisons of our sparse deconvolution pipeline along with pre-existing
802 reconstruction methods based on the sparsity a priori. (a-c) Actin filaments labeled with LifeAct-EGFP in
803 a live COS-7 cell in Fig. 4a under the 2D-SIM, followed by the FISTA, or NLHT reconstruction, or
804 reconstruction with our sparse algorithm. (d and e) CCPs labeled by Clathrin-EGFP under the SD-SIM in
805 Extended Data Fig. 15, followed by the FISTA, or NLHT reconstruction, or reconstruction with our sparse
806 algorithm. Under this condition of compromised SNR, neither FISTA nor NLHT method could resolve any
807 ring-shaped structures. Scale bars: (a) 1 μm; (b and c) 500 nm; (d) 3 μm; (e) 1 μm.

51
808 Additional Supplementary Figures

809

810 Supplementary Fig. 26 | Comparing results of SD-SIM, RL deconvolved SD-SIM, and Sparse SD-SIM
811 (c.f., Fig. 5k). (a) OMM in a live COS-7 cell (labeled with TOM20-mCherry) under SD-SIM, SD-SIM
812 deconvolved by RL with 3 and 15 iterations, and Sparse SD-SIM, respectively. (e) Magnified views from
813 white boxes in (a-d) respectively. Scale bars: (a-d) 5 μm; (e) 1 μm.

52
814

815 Supplementary Fig. 27 | Comparing results of SD-SIM, PURE denoised SD-SIM, Sparse SD-SIM, and
816 RL deconvolved SD-SIM. (c.f., Fig. 6e). (a-c) HeLa cell labeled with tubulin-EGFP (magenta), Pex11a-BFP
817 (cyan), and Lamp1-mCherry (yellow) captured by SD-SIM, PURE denoised41 SD-SIM, Sparse SD-SIM. (d)
818 Raw SD-SIM deconvolved by RL with 3 iterations. (e) Raw SD-SIM deconvolved by RL with 15 iterations.
819 (f) Raw SD-SIM deconvolved by RL with 50 iterations. PURE: Poisson Unbiased Risk Estimate41. Scale bars:
820 (a-f) 5 μm; (a-f, inset) 3 μm.

53
821

822 Supplementary Fig. 28 | Comparing results of SD-SIM, TV SD-SIM, and Sparse SD-SIM (c.f., Fig. 6j).
823 (a-c) Hoechst in a live COS-7 cell under the SD-SIM (axial position at +3 μm) (a), TV30 SD-SIM (b), and
824 Sparse SD-SIM (c), respectively. (d-f) Corresponding z-axial views from the white dashed lines in (a-c). TV:
825 Total Variance30; Scale bars: (a-c) 5 μm; (d-f) 2 μm.

54
826 Supplementary Tables

827 Supplementary Table 1 | SSIM of different reconstructions from images corrupted with varying levels
828 of noise.

Without noise G (11%) G (25%) G (50%) G (80%)

Wiener-SIM 0.5612 0.2352 0.1773 0.1417 0.1445
Continuity 0.5500 0.2303 0.1724 0.1311 0.1545
Continuity+sparsity 0.8276 0.6741 0.4766 0.5652 0.5418
Continuity+deconv 0.8311 0.2886 0.1724 0.1790 0.2107
Sparsity+deconv 0.9430 0.5203 0.4178 0.2884 0.3191
Sparse-SIM 0.9570 0.6832 0.6431 0.6167 0.6376
829 G represented the Gaussian noise. From Extended Data Fig. 3.
830
831 Supplementary Table 2 | PSNR of different reconstructions from images corrupted with varying levels
832 of noise.

Without noise G (11%) G (25%) G (50%) G (80%)

Wiener-SIM 18.08 10.90 9.39 8.27 11.32
Continuity 17.94 10.85 9.40 8.23 9.73
Continuity+sparsity 21.63 18.10 16.09 16.13 16.40
Continuity+deconv 23.01 13.93 11.77 10.68 13.22
Sparsity+deconv 27.04 17.74 15.88 15.29 15.24
Sparse-SIM 27.92 18.72 18.08 17.97 18.11
833 G represented the Gaussian noise. From Extended Data Fig. 3.

55
834 Supplementary Table 3 | Imaging conditions.

Illumination
Wavelength Exposure
Microscope Figure Label intensity
(nm) (ms)
(W/cm2)
Fig. 2c Nup-98GFP 488 7×9 21
Fig. 4a LifeAct-EGFP 488 20×9 2
Fig. 4j Caveolin-EGFP 488 7×9 14
LAMP1-EGFP
Fig. 4m LysoView 488 488 20×9 2
LipidSpot 488
SIM Fig. 4o VAMP2-pHluorin 488 0.2×9 187
ExFig. 5f Clathrin-EGFP 488 7×9 14

MitoTracker Green 488 7×9 14

ExFig. 14a
Tom20-mScarlet 561 7×9 50
MitoTracker Green 488 7×9 14
ExFig. 14f
Sec61β-mCherry 561 7×9 100
Expansion- α-tubulin
Fig. 3 488 20×9 20
SIM Sec61β-GFP
Fig. 5a Clathrin-EGFP 488 200 0.91
Lifeact –EGFP 488 200 0.89
Fig. 5g, ExFig. 15
Clathrin-DsRed 561 300 12.11
Fig. 5k Tom20-mCherry 561 200 8.84
Hoechst H1399 405 200 0.87
Tubulin-EGFP 488 200 0.91
Fig. 5h
Lamp1-mCherry 561 200 8.84
MitoTracker® Deep Red FM 640 50 1.07
Fig. 6a Sec61β-EGFP 488 200 0.91
SD-SIM
Pex11a-BFP 405 100 0.87
Fig. 6e Tubulin-EGFP 488 300 0.91
Lamp1-mCherry 561 200 4.96
Hoechst H1399 405 50 0.87
Fig. 6h Tubulin-EGFP 488 200 0.89
MitoTracker® Deep Red FM 640 50 0.27
Sec61β-EGFP 488 200 0.77
ExFig. 16
Lysotracker Deep Red 640 200 0.92
ExFig. 17 GCaMP6s 488 90 0.27
Et.: 112
ExFig. 18a Mab414 590
Dt.: 229×106
Et.: 74
Tom20 594
Dt.: 48×106
ExFig. 18f
Et.: 97
Tubulin 635
Dt.: 14×106
STED
Et.: 109
ExFig. 19a Lifeact-GFP 488
Dt.: 16×106
Et.: 156
ExFig. 19b Sec61β-GFP 488
Dt.: 13×106
Et.: 307
ExFig. 19c SiR-Tubulin 651
Dt.: 158×106

MTPM ExFig. 20 Thy1-GFP 920 3.98×106

835 ExFig.: Extended Data Figure; Et.: Excitation laser; Dt.: Depletion laser.

56
836 Supplementary Table 4 | Non-zero ratios and background selections of different images.

Non-zero number Non-zero ratio

Figure Label Modality Pixel number Background
(Raw / Sparse) (Raw / Sparse)

1024 × 1024
Fig. 2c Nup98-GFP TIRF-SIM 1037042 / 159384 99 % / 4 % Weak-HI
(× 2)

Fig. 4a LifeAct-EGFP 2D-SIM 1675100 / 589824 71 % / 25 % 1536 × 1536 No

Caveolin- 793 × 871

Fig. 4j TIRF-SIM 566376 / 1215637 82 % / 44 % No
EGFP (×2)

Fig. 4m LAMP1-EGFP 2D-SIM 2312110 / 117965 98 % / 5 % 1536 × 1536 Strong-HI

Fig. 4m LysoView 488 2D-SIM 2345140 / 212336 99 % / 9 % 1536 × 1536 Strong-HI

LipidSpot
Fig. 4m 2D-SIM 2264924 / 94371 96 % / 4 % 1536 × 1536 Strong-HI
488

VAMP2- 512 × 144

Fig. 4o TIRF-SIM 64659 / 1180 88 % / 0.04 % No
pHluorin (× 2)

Fig. 5a Clathrin-EGFP SD-SIM 751564 / 37578 100 % / 5 % 899 × 836 Strong-HI

Lamp1-
Fig. 5h SD-SIM 1327104 / 185795 100 % / 14 % 1152 × 1152 Strong-HI
mCherry
MitoTracker®
Fig. 5h SD-SIM 1327104 / 225608 100 % / 17 % 1152 × 1152 Strong-HI
Deep Red FM

Hoechst
Fig. 5h SD-SIM 1327104 / 1008599 100 % / 76 % 1152 × 1152 Strong-LI
H1399

Fig. 5h Tubulin-EGFP SD-SIM 1327104 / 1167852 100 % / 88 % 1152 × 1152 Weak-LI

Tom20- 1042 × 975

Fig. 5k SD-SIM 1015950 / 547597 100 % / 54 % Strong-HI
mCherry 3D-MIP

57
837 Continued Table of Supplementary Table 4...

Non-zero number Non-zero ratio

Figure Label Modality Pixel number Background
(Raw / Sparse) (Raw / Sparse)

Sec61β- 512 × 512

Fig. 6a SD-SIM 262144 / 555745 100 % / 53 % Weak-LI
EGFP (× 2)
512 × 512
Fig. 6e Tubulin-EGFP SD-SIM 262144 / 786432 100 % / 75 % No
(× 2)
512 × 512
Fig. 6e Lamp1-mCherry SD-SIM 262144 / 94371 100 % / 9 % Strong-LI
(× 2)
Pex11a- 512 × 512
Fig. 6e SD-SIM 262144 / 41943 100 % / 4 % No
BFP (× 2)
512 × 512
Fig. 6h Hoechst H1399 SD-SIM 262144 / 576717 100 % / 55 % Strong-HI
(× 2) 3D-MIP
MitoTracker® Deep 512 × 512
Fig. 6h SD-SIM 262144 / 377487 100 % / 36 % Strong-HI
Red FM (× 2) 3D-MIP
512 × 512
Fig. 6h Tubulin-EGFP SD-SIM 262144 / 450888 100 % / 43 % Strong-LI
(× 2) 3D-MIP

ExFig.13a MitoTracker Green 2D-SIM 249037 / 99615 95 % / 38 % 512 × 512 No

ExFig.13a Tom20-mScarlet 2D-SIM 209715 / 39322 80 % / 15 % 512 × 512 No

ExFig.13e MitoTracker Green 2D-SIM 254280 / 18350 97 % / 7 % 512 × 512 No

ExFig.13e Sec61β-mCherry 2D-SIM 214958 / 28836 82 % / 11 % 512 × 512 No

ExFig.17 GCaMP6s SD-SIM 2303136 / 852160 100 % / 37 % 1454 × 1584 No

58
838 Supplementary Table 5 | Computing time of Sparse deconvolution.

Figure Label Size (x × y × t/z) GPU time CPU time

Fig. 2c Nup-98GFP 2048 × 2048 × 200 (t) ~1800 s ~1812 min

Fig. 3a α-tubulin 2048 × 2048 × 10 (t) ~90 s ~92 min
Fig. 3d Sec61β-GFP 2048 × 2048 × 10 (t) ~90 s ~92 min
Fig. 4a LifeAct-EGFP 1536 × 1536 × 20 (t) ~90 s ~46 min
Fig. 4j Caveolin-EGFP 2048 × 2048 × 200 (t) ~1680 s ~1708 min
Fig. 4m LAMP1-EGFP 1536 × 1536 × 50 (t) ~188 s ~120 min
Fig. 5a Clathrin-EGFP 873 × 873 × 100 (t) ~140 s ~80 min
Fig. 5g Clathrin-DsRed 1152 × 1152 × 200 (t) ~425 s ~329 min
Fig. 5k Tom20-mCherry 1152 × 1152 × 110 (z) ~231 s ~181 min
Fig. 5h Hoechst H1399 1152 × 1152 × 100 (t) ~210 s ~165 min

59
839 Captions for Supplementary Videos.

840

841 Supplementary Video 1 | Ring-shaped Nup98 pores resolved with the Sparse-SIM 2×. Nuclear pores
842 labeled with Nup98-GFP in a live COS-7 cell with 4 s interval imaged by 2D-SIM, 2D-SIM followed by RL
843 deconvolution and Sparse-SIM 2× (c.f., Fig. 2c) demonstrate that only Sparse-SIM 2× can resolve the ring-
844 shaped Nup98.

60
845

846 Supplementary Video 2 | Dense actin mesh network revealed due to improved spatial resolution and
847 enhanced contrast. Part I compares the actin results by 2D-SIM, Hessian-SIM and Sparse-SIM,
848 demonstrating that finer structures of the dense actin mesh network can be revealed only with Sparse-SIM
849 (c.f., Fig. 4a). Part II exhibits the corresponding profiles of the red and blue lines from 2D-SIM (left) and
850 Sparse-SIM (right) over 20 frames at 5 s intervals, showing the enhanced contrast available with Sparse-SIM.

61
851

852 Supplementary Video 3 | Ring-shaped caveolae revealed by the Sparse-SIM 2×. The caveolae in a COS-
853 7 cell at 37°C transfected with Caveolin-EGFP by TIRF, TIRF-SIM, Sparse-SIM and Sparse-SIM 2× (c.f., Fig.
854 4j).

62
855

856 Supplementary Video 4 | Improved contrast of fluorescent vesicles deep in the cytosol under the Sparse-
857 SIM. Part I the structures labeled with LAMP1-EGFP, LysoView, and LipidSpot in COS-7 cells are shown
858 from left to right at 5 s intervals (c.f., Fig. 4m). Part II compares the 2D-SIM, Hessian-SIM and Sparse-SIM
859 performances on three individual organelles and the corresponding magnified regions of interest.

63
860

861 Supplementary Video 5 | Visualization of ultrafast vesicle fusion (played 4x slower than real-world
862 speed). Fusion pores by TIRF-SIM (top) and Sparse-SIM (bottom) played at a speed 4x slower than real world
863 rates. Vesicle in an INS-1 cell labeled with VAMP2-pHluorin by TIRF-SIM and Sparse-SIM, respectively (c.f.,
864 Fig. 4o).

64
865

866 Supplementary Video 6 | Vesicle with a 60 nm pore diameter captured by the Sparse-SIM at 564 Hz.
867 The initial opening of the fusion pore occurred much earlier, and its duration was much longer under Sparse-
868 SIM than under TIRF-SIM (same data as in Supplementary Video 5).

65
869

870 Supplementary Video 7 | Asynchronized movements between the outer and inner mitochondrial
871 membranes revealed by dual-color Sparse-SIM. Outer and inner mitochondrial membranes (OMM and
872 IMM, labeled by Tom20- mScarlet and Mito-Tracker Green, respectively) by 2D-SIM and Sparse-SIM for 30
873 frames with a 0.2 s acquisition time (c.f., Extended Data Fig. 13a). The extension of the inner mitochondrial
874 membrane was not enclosed by the Tom20-labeled structures captured by Sparse-SIM in a few frames,
875 demonstrating the nonhomogenous distribution of Tom20 on the OMM.

66
876

877 Supplementary Video 8 | Relative dynamics between ER tubules and inner mitochondrial membrane
878 recorded by dual-color Sparse-SIM. The inner mitochondrial membrane (cyan) and ER (magenta) in a COS-
879 7 cell labeled by MitoTracker Green and Sec61β-mCherry (c.f., Extended Data Fig. 13f). Part I compares the
880 IMM, and ER by dual-color 2D-SIM and Sparse-SIM. In Part II, the ER contacts the IMM, rearranging the
881 orientations of the inner cristae structures. Part III shows the equal probability of ER tubules contacting the
882 mitochondria at cristae regions or in the matrix between cristae.

67
883

884 Supplementary Video 9 | Ring-shaped structure of clathrin-coated pits resolved by the Sparse SD-SIM.
885 Comparison of clathrin-coated pits (CCPs) in COS-7 cells expressing clathrin-EGFP light chains for 100 time
886 points with a 0.2 s exposure time imaged by SD-SIM and Sparse SD-SIM, respectively (c.f., Fig. 5a).
887 Disintegration (part II of the video) and disappearance (part III) events of CCPs are resolved only by Sparse-
888 SD-SIM.

68
889

890 Supplementary Video 10 | Four-color sub-90-nm resolution live-cell imaging of lysosomes (yellow),
891 mitochondria (green), nuclei (blue), and microtubules (magenta). A COS-7 cell labeled with LAMP1-
892 mCherry (yellow), Mito-Tracker (green), Hoechst (blue), and Tubulin-EGFP (magenta) imaged by SD-SIM
893 and Sparse SD-SIM for 100 frames at a 2.3 s exposure time (c.f., Fig. 5h). Part I shows the SD-SIM
894 performance of the four organelles, which gradually transfer to Sparse SD-SIM in sequence. Then, the four
895 organelles under SD-SIM and Sparse SD-SIM are compared simultaneously showing the striking contrast.
896 Part II displays the dynamics of the magnified selection region in part I.

69
897

898 Supplementary Video 11 | Live-cell 3D image of the mitochondrial membrane in a dividing cell with the
899 depth of field of up to ~7 µm. Mitochondrial expressions of the outer membrane marker TOM20-mCherry
900 captured by SD-SIM and Sparse SD-SIM. (c.f., Fig. 5k). Part I shows the color-coded projection of the
901 mitochondrial membrane in the z direction as the depth increases gradually and the y-z orthoslices along the
902 x-axis. Part II displays multi-angle views of the 3D rendering volume of the mitochondrial membrane showing
903 the details recorded by SD-SIM (red) and Sparse SD-SIM (green).

70
904

905 Supplementary Video 12 | ER tubular structures under Nyquist insufficient sampling paradigm
906 resolved by the Sparse SD-SIM 2× configuration. Tubular ER in the periphery of a COS-7 cell labeled with
907 Sec61β-EGFP. Part I shows a comparison of the tubular ER by SD-SIM, Hessian SD-SIM, Sparse SD-SIM,
908 and Sparse SD-SIM 2× (c.f., Fig. 6a). Part II compares the tubular ER by SD-SIM (left) with those by Sparse
909 SD-SIM (right) 2× for 500 time points at a 0.2 s acquisition time. A magnified view of the selected region in
910 Part II is shown in Part III.

71
911

912 Supplementary Video 13 | Microtubules regulate lysosome-peroxisome membrane contacts in live cells
913 captured by three-color Sparse SD-SIM 2× imaging. Part I: A time-lapse full-field view comparison of SD-
914 SIM and Sparse SD-SIM 2×. Part II: The two zoomed-in examples show a comparison of the images by SD-
915 SIM and Sparse SD-SIM 2× (c.f., Fig. 6e). The first example shows a lysosome moving along the microtubule
916 to make membrane contact the peroxisome for close to 2 mins; the deformation of the lysosome membrane
917 during the contact period can be detected only by Sparse SD-SIM 2×. The second example shows a lysosome-
918 peroxisome membrane in contact moving along microtubules together in a relatively dense microtubule region;
919 the movement can be detected only by Sparse SD-SIM 2×.

72
920

921 Supplementary Video 14 | Three-color 3D volume rendered views under the Sparse SD-SIM 2×
922 configuration. Live-cell three-color 3D imaging of a COS-7 cell labeled with Tubulin-EGFP (green), Hoechst
923 (cyan), and MitoTracker® Deep Red FM (magenta) under Sparse SD-SIM 2× (c.f., Fig. 6h). Part I shows x-y
924 orthoslices gradually forming color-coded volumes of nuclei (left), mitochondria (middle), and microtubules
925 (right). Part II displays three-color 3D volume rendered views of nuclei, mitochondria, and microtubules from
926 various angles.

73
927

928 Supplementary Video 15 | Benchmarks of time-varying actin filaments under different approaches
929 according to the ground truth 2D-SIM images. Two branched actin filaments were recorded by Wide-Field,
930 Wide-Field + deconvolution, Sparse 2×, and 2D-SIM (c.f., Extended Data Fig. 5e). Using 2D-SIM as the
931 ground truth, Wide-Field with Sparse 2× reveals fine structure such as two-branched actin filaments separated
932 correctly.

74
933

934 Supplementary Video 16 | Relative dynamics of both actin filaments and CCPs captured by the dual-
935 color Sparse SD-SIM. Actin filaments (labeled by LifeAct-EGFP) and CCPs (labeled by Clathrin-DsRed) in
936 COS-7 cells for 200 time points at a 5 s interval (c.f., Extended Data Fig. 15). A CCP stably docked at the
937 intersection of two actin filaments that begins to disappear from the confocal plane as the neighboring
938 filaments close up and meet (marked with a yellow circle).

75
939

940 Supplementary Video 17 | ER-lysosome contacts in live cells captured by the dual-color Sparse SD-SIM.
941 Part I: The time-lapse full-field ER-lysosome (labeled with Sec61β-EGFP and Lysotracker Deep Red) view
942 comparison of SD-SIM and Sparse SD-SIM (c.f., Extended Data Fig. 16). Part II: The zoom-in example
943 shows typical lysosome-ER contact dynamics by SD-SIM and Sparse SD-SIM. The lysosome-ER contact
944 dynamics can be clearly viewed under Sparse SD-SIM.

76
945

946 Supplementary Video 18 | Three-dimensional images of neuronal dendrites and spines observed by the
947 Sparse-MTPM. Three-dimensional distributions of neuronal dendrites and spines within a volume of 190 ×
948 190 × 110 μm3 from the brain of a Thy1-GFP transgenic mouse observed with MTPM, RL deconvoluted
949 MTPM (RL-MTPM), and Sparse-MTPM (c.f., Extended Data Fig. 20). Part I demonstrates x-y comparison
950 of these three methods. Part II shows the x-y orthoslices gradually forming color-coded volumes of MTPM,
951 RL deconvoluted MTPM (RL-MTPM), and Sparse-MTPM from left to right respectively. Part III displays 3D
952 volume rendered views of neuronal dendrites and spines from various angles.

77
953 References

954 1. Hunt, B.R. Super-resolution of images: Algorithms, principles, performance. International Journal of Imaging
955 Systems and Technology 6, 297-304 (1995).
956 2. Richardson, W.H. Bayesian-based iterative method of image restoration. Journal of The Optical Society of America
957 A 62, 55-59 (1972).
958 3. Lucy, L.B. An iterative technique for the rectification of observed distributions. The Astronomical Journal 79, 745
959 (1974).
960 4. Shepp, L.A. & Vardi, Y. Maximum likelihood reconstruction for emission tomography. IEEE Transactions on
961 Medical Imaging 1, 113-122 (1982).
962 5. Dey, N. et al. Richardson–Lucy algorithm with total variation regularization for 3D confocal microscope
963 deconvolution. Microscopy research and technique 69, 260-266 (2006).
964 6. Lucy, L.B. Resolution limits for deconvolved images. The Astronomical Journal 104, 1260-1265 (1992).
965 7. Lucy, L.B. Statistical limits to super resolution. Astronomy & Astrophysics 261, 706-710 (1992).
966 8. Wolter, H. On basic analogies and principal differences between optical and electronic information, Vol. 1. (Elsevier,
967 1961).
968 9. Goodman & W, J. Introduction to Fourier optics. (Roberts and Company Publishers, 2005).
969 10. Bertero, M. & De Mol, C. Super-Resolution by Data Inversion, Vol. 36. (Elsevier, 1996).
970 11. Lindberg, J. Mathematical concepts of optical superresolution. Journal of Optics 14, 083001 (2012).
971 12. Cox, c.I. & Sheppard, C. Information capacity and resolution in an optical system. Journal of The Optical Society of
972 America A 3, 1152-1158 (1986).
973 13. Aspelmeier, T., Egner, A. & Munk, A. Modern Statistical Challenges in High-Resolution Fluorescence Microscopy.
974 Annual Review of Statistics and Its Application 2, 163–202 (2015).
975 14. Robbins, M. & Hadwen, B.J. The noise performance of electron multiplying charge-coupled devices. IEEE
976 Transactions on Electron Devices 50, 1227-1232 (2003).
977 15. Foi, A., Trimeche, M., Katkovnik, V. & Egiazarian, K. Practical Poissonian-Gaussian Noise Modeling and Fitting for
978 Single-Image Raw-Data. IEEE Transactions on Image Processing 17, 1737-1754 (2008).
979 16. Mortensen, K.I., Churchman, L.S., Spudich, J. & Flyvbjerg, H. Optimized localization-analysis for single-molecule
980 tracking and super-resolution microscopy. Nature Methods 7, 377 - 381 (2010).
981 17. Hirsch, M., Wareham, R.J., Martin-Fernandez, M., Hobson, M.P. & Rolfe, D. A Stochastic Model for Electron
982 Multiplication Charge-Coupled Devices: From Theory to Practice. PLoS ONE 8 (2013).
983 18. Liu, S. et al. sCMOS noise-correction algorithm for microscopy images. Nature Methods 14, 760-761 (2017).
984 19. Born, M. & Wolf, E. Principles of optics, Edn. 7th. (Cambridge University Press, Cambridge; 1999).
985 20. Heintzmann, R. & Sheppard, C. The sampling limit in fluorescence microscopy. Micron 38 2, 145-149 (2007).
986 21. Lohmann, A. et al. Space bandwidth product of optical signals and systems. Journal of The Optical Society of America
987 A 13, 470-473 (1996).
988 22. Betzig, E. et al. Imaging Intracellular Fluorescent Proteins at Nanometer Resolution. Science 313, 1642 - 1645 (2006).
989 23. Rust, M., Bates, M. & Zhuang, X. Sub-diffraction-limit imaging by stochastic optical reconstruction microscopy
990 (STORM). Nature Methods 3, 793-796 (2006).
991 24. Huang, X. et al. Fast, long-term, super-resolution imaging with Hessian structured illumination microscopy. Nature
992 biotechnology 36, 451-459 (2018).
993 25. Candes, E.J. & Tao, T. Near-optimal signal recovery from random projections: Universal encoding strategies? IEEE
994 Transactions on Information Theory 52, 5406-5425 (2006).
995 26. Szameit, A. et al. Sparsity-based single-shot subwavelength coherent diffractive imaging. Nature materials 11, 455-
996 459 (2012).
997 27. Zhu, L., Zhang, W., Elnatan, D. & Huang, B. Faster STORM using compressed sensing. Nature Methods 9, 721-723

78
 998 (2012).
999 28. Galloway, C., Le Ru, E. & Etchegoin, P. An iterative algorithm for background removal in spectroscopy by wavelet
1000 transforms. Applied Spectroscopy 63, 1370-1376 (2009).
1001 29. Lloyd, S.P. Least squares quantization in PCM. IEEE Trans. Inf. Theory 28, 129-136 (1982).
1002 30. Wang, Y., Yang, J., Yin, W. & Zhang, Y. A new alternating minimization algorithm for total variation image
1003 reconstruction. SIAM Journal on Imaging Sciences 1, 248-272 (2008).
1004 31. Goldstein, T. & Osher, S. The split Bregman method for L1-regularized problems. SIAM journal on imaging sciences
1005 2, 323-343 (2009).
1006 32. Landweber, L. An iteration formula for Fredholm integral equations of the first kind. American journal of
1007 mathematics 73, 615-624 (1951).
1008 33. Sutskever, I., Martens, J., Dahl, G. & Hinton, G. in International conference on machine learning 1139-1147 (2013).
1009 34. Veklerov, E. & Llacer, J. Stopping rule for the MLE algorithm based on statistical hypothesis testing. IEEE
1010 Transactions on Medical Imaging 6, 313-319 (1987).
1011 35. Biggs, D.S. & Andrews, M. Acceleration of iterative image restoration algorithms. Applied optics 36, 1766-1775
1012 (1997).
1013 36. Stein, S.C., Huss, A., Hähnel, D., Gregor, I. & Enderlein, J. Fourier interpolation stochastic optical fluctuation
1014 imaging. Optics Express 23, 16154-16163 (2015).
1015 37. Peng, T. et al. A BaSiC tool for background and shading correction of optical microscopy images. Nature
1016 Communications 8, 14836 (2017).
1017 38. Theer, P., Mongis, C. & Knop, M. PSFj: know your fluorescence microscope. Nature Methods 11, 981-982 (2014).
1018 39. Beck, A. & Teboulle, M. A Fast Iterative Shrinkage-Thresholding Algorithm for Linear Inverse Problems. SIAM J.
1019 Imaging Sciences 2, 183-202 (2009).
1020 40. Gazit, S., Szameit, A., Eldar, Y.C. & Segev, M. Super-resolution and reconstruction of sparse sub-wavelength images.
1021 Optics Express 17, 23920-23946 (2009).
1022 41. Luisier, F., Vonesch, C., Blu, T. & Unser, M. Fast interscale wavelet denoising of Poisson-corrupted images. Signal
1023 Processing 90, 415-427 (2010).
1024

79