This pipeline is an open-source pipeline for MRI image segmentation, registration, and quantitative analysis.

The current code uses the command line interface for use. Pull requests for a GUI to command-line translation are welcome.

This repo is to serve as an open-source location for developers to add MRI processing techniques. This includes, but is not limited to:

• image processing tasks (denoising, super-resolution, segmentation, etc)
• relaxation parameter analysis (T1, T1-rho, T2, T2*, etc)
• anatomical features (patellar tilt, femoral cartilage thickness, etc)

We hope that this open-source pipeline will be useful for quick anatomy/pathology analysis from MRI and will serve as a hub for adding support for analyzing different anatomies and scan sequences.

Currently, this pipeline supports analysis of the femoral cartilage in the knee using cubequant, cones, and DESS scanning protocols. Details are provided below.

The following scan sequences are supported. All sequences with multiple echos, spin_lock_times, etc. should have metadata in the dicom header specifying this information.

All data should be provided in the dicom format. Currently only sagittal orientation dicoms are supported.

Dicom files should be named in the format 001.dcm: echo1, 002.dcm: echo2, 003.dcm: echo1, etc.

T₂: Calculate T₂ map using dual echos

Segmentation

Analysis for the following anatomical regions are supported

Tissues: Femoral Cartilage

Download this repo to your disk. Note that the path to this repo should not have any spaces. In general, this pipeline does not handle folder paths that have spaces in between folder names.

We recommend using the (Anaconda)[https://www.anaconda.com/download] virtual environment to run python.

An environment.yml file is provided in this repo containing all libraries used.

For pretrained weights for MSK knee segmentation, request access to this Google Drive. Note that these weights are optimized to run on single-echo RMS DESS sequence as used in the OA initiative.

Save these weights in an accessible location. Do not rename these files.

To run the program from a shell, run python -m opt/path/pipeline with the flags detailed below. opt/path is the path to the file python

usage: pipeline [-h] [-d [D]] [-l [L]] [-s [S]] [-e [E]] [--gpu [G]]
            {dess,cubequant,cq,knee} ...

Pipeline for segmenting MRI knee volumes

positional arguments:
  {dess,cubequant,cq,knee}
                        sub-command help
    dess                analyze DESS sequence
    cubequant (cq)      analyze cubequant sequence
    knee                analyze tissues

optional arguments:
  -h, --help            show this help message and exit
  -d [D], --dicom [D]   path to directory storing dicom files
  -l [L], --load [L]    path to data directory to load from
  -s [S], --save [S]    path to directory to save mask. Default: D/L
  -e [E], --ext [E]     extension of dicom files. Default 'dcm'
  --gpu [G]             gpu id

The DESS protocol used here is detailed in this paper referenced below:

Chaudhari, Akshay S., et al. "Five‐minute knee MRI for simultaneous morphometry and T2 relaxometry of cartilage and meniscus and for semiquantitative radiological assessment using double‐echo in steady‐state at 3T." JMRI 47.5 (2018): 1328-1341.

The figure below details the DESS protocol used in this paper:

Figure 1: Supported DESS protocol as referenced here

usage: pipeline dess [-h] [-rms] [-t2] {segment} ...

positional arguments:
  {segment}   sub-command help

optional arguments:
  -h, --help  show this help message and exit
  -rms        use root mean square (rms) of two echos for segmentation
  -t2         compute T2 map

usage: pipeline dess segment [-h] [--model [{unet2d}]] [--weights_dir WEIGHTS_DIR]
                         [--batch_size [B]] [-fc]

optional arguments:
  -h, --help            show this help message and exit
  --model [{unet2d}]
  --weights_dir WEIGHTS_DIR
                        path to directory with weights
  --batch_size [B]      batch size for inference. Default: 32
  -fc                   handle femoral cartilage

The cubequant protocol used here is detailed below:

usage: pipeline cubequant [-h] [-t1rho] [-fm [FM]] {interregister} ...

positional arguments:
  {interregister}  sub-command help

optional arguments:
  -h, --help       show this help message and exit
  -t1rho           do t1-rho analysis
  -fm [FM]         focused mask to speed up t1rho calculation

Register cubequant scan to a target scan

usage: pipeline cubequant interregister [-h] [-ts TS] [-tm [TM]]

optional arguments:
  -h, --help  show this help message and exit
  -ts TS      path to target image (nifti)
  -tm [TM]    path to target mask (nifti)

The cones protocol used here is detailed below:

usage: pipeline cones [-h] [-t2star] [-fm [FM]] {interregister} ...

positional arguments:
  {interregister}  sub-command help

optional arguments:
  -h, --help       show this help message and exit
  -t2star          do t2* analysis
  -fm [FM]         focused mask to speed up t1rho calculation

usage: pipeline cones interregister [-h] [-ts TS] [-tm [TM]]

optional arguments:
  -h, --help  show this help message and exit
  -ts TS      path to target image (nifti)
  -tm [TM]    path to target mask (nifti)

usage: pipeline knee [-h] [--orientation [{RIGHT,LEFT}]] [-fc] [-t2] [-t1_rho]
                     [-t2_star]

optional arguments:
  -h, --help            show this help message and exit
  --orientation [{RIGHT,LEFT}]
                        knee orientation (left or right)
  -fc                   analyze femoral cartilage
  -t2                   quantify t2
  -t1_rho               quantify t1_rho
  -t2_star              quantify t2_star

All weights/parameters trained for any task are likely to be most closely correlated to data used for training. If scans from a particular sequence were used for training, the performance of those weights are likely optimized for that specific scan type. As a result, they may not perform as well on segmenting images acquired using different scan types.

If you do train weights for any deep learning task that you would want to include as part of this repo, please provide a link to those weights and detail the scanning parameters/sequence used to acquire those images. All data contributed to this pipeline should be made freely available to all users.

We detail use cases that could be useful for analyzing data. We assume that all scans are stored per patient, meaning that the folder structure looks like below:

research_data
    | patient01
        | dess
            | I001.dcm
            | I002.dcm
            | I003.dcm
            ....
        | cubequant
        | cones
        | <OTHER SCAN SEQUENCE DATA>
    | patient02
    | patient03
    | unet_weights
    ...

All use cases assume that the current working directory is this repo. If the working directory is different, make sure to specify the path to pipeline.py when running the script. For example, python -m ~/MyRepo/pipeline.py if the repo is located in the user directory.

Analyze patient01 knee T₂ properties using DESS sequence

1. Calculate 3D T₂ map

python -m pipeline -d research_data/patient01/dess -s research_data/patient01/data dess -t2

2. Segment femoral cartilage using RMS of two echo dess echos

python -m pipeline -d research_data/patient01/dess -s research_data/patient01/data dess -rms segment --weights_dir unet_weights

Note steps 1 and 2 can be combined as the following:

python -m pipeline -d research_data/patient01/dess -s research_data/patient01/data dess -rms -t2 segment --weights_dir unet_weights

3. Calculate T₂ time for femoral cartilage

python -m pipeline -l research_data/patient01/data -s research_data/patient01/data knee -fc -t2