NON-INVASIVE LIVER CANCER DETECTION
USING DEEP LEARNING
Harshal Sonkusale Hitanshu Dadhich Nayan Dhurve
Yash Naktode Chiranjit Roy
[email protected] [email protected]
Department of Information Technology, Yeshwantrao Chavan College of Engineering, Nagpur, 441110
Abstract:
Non-invasive deep learning-based system for liver cancer detection is a promising area of
research which has the potential to transform liver cancer diagnosis. Such a system can detect
liver cancer early and provide significant cost savings. Additionally, the system has the
potential to improve the survival rate of liver cancer patients, personalize treatment plans,
monitor the response of patients to treatment, and reduce the need for invasive procedures.
This project aims to make a non-invasive deep learning-based system for detecting liver cancer
using CT scan images. We have collected data from CT scan images with and without liver
cancer and currently, we have developed our models of deep learning algorithms to detect liver
cancer lesions on CT scan images. The work is still in its early stages, but the results we got at
the initial stage are very promising. We used three deep learning models CNN, ResNet, and
MobileNetV1. The CNN model showed the highest accuracy in detecting liver cancer lesions,
with an accuracy of 91.67% followed by the MobileNetV1 model achieving an accuracy of
87.50% which are much better compared to the ResNet model which achieved an accuracy of
60.87%.
We believe that our work has the potential to make a significant impact in the fight against
liver cancer by developing an accurate and efficient non-invasive system based on deep
learning to detect liver cancer with it, we can help detect liver cancer earlier, improve patient
survival, and reduce healthcare costs.
Keywords: Non-Invasive, Liver cancer detection, cost saving, Deep learning, CT scan images, CNN
model, Resnet model, MobileNetV1 model.
1. Introduction most cases. To minimize the potential use of such
painful physical examinations, non-invasive
Liver cancer is a serious disease that is the third methods such as US, MRI, and CT are used to
leading cause of cancer-related death worldwide. It reliably screen patients at risk followed by the use of
has also been estimated that more than 900,000 new ventilation to reduce early detection of death can be
cases of liver cancer will be diagnosed and 830,000 done. We discuss these methods and protocols used,
more people will die from liver cancer by 2024. The on how to conduct detailed analysis.
prognosis for liver cancer is poor, even in patients
with severe disease has life expectancy of 5 years.
2.1.1. Ultrasonography (US)
However, early detection and diagnosis of liver The first commonly used imaging modality for the
cancer can significantly improve patient outcomes. abdomen is US parenchymal organs, these are non-
Conventional methods of detecting liver cancer, invasive, extensively available and inexpensive.
such as physical examination, testing and imaging, However, because of the variable diagnostic
are invasive and expensive. A biopsy involves sensitivity of hepatic fibrosis, and those with
taking a sample of liver tissue for testing purposes, positive experience and qualified staff. Use of
which can be painful and dangerous. Imaging tests, powerful sound waves to do so viewed from inside
such as CT (computed tomography) and MRI the body. A transducer is placed to concentrate a
(magnetic resonance imaging), are expensive and permeable substance containing gel in the skin for
time-consuming. Deep learning (DL) could sound waves. This is allowed because of the real-
revolutionize liver cancer diagnosis by providing time characteristics of active scanning [4].
non-invasive and accurate methods. Deep Learning
algorithms (DL) can be trained to detect liver cancer 2.1.2. Computed Tomography (CT)
from medical images such as CT scans and MRI
CT uses X-rays to acquire images using sequential
scans. Once these models of Deep Learning are
or spiral methods. CT is preferred for evaluating
trained, they can be used to detect and classify liver
hepatic features because of spiral and second
cancer in new patients. In this research paper, we
scanning. Axial reconstruction with images is the
propose a 5-step deep learning-based method for
best spacing resulting in a decrease in the average
liver cancer detection. Deep learning algorithms are
lesion volume. Similarly, it can monitor liver
well suited for image recognition tasks.
enlargement in arterial portal delayed modes.
2. Method Preview Moreover, registration events due to inhalation are
less frequent than during a single breath. Finally,
2.1. Image acquisition spiral CT scans provide three-dimensional imaging
information. Usually, a CT scan of the abdomen is
Patients with chronic liver disease often have performed after the assessment of variable tissue
complications due to liver failure. The diagnosis and contrast agents with volume, size, and
medical examination of such chronic diseases reconstruction time. Sometimes a series of scans are
should account for the high fraction of liver fibrosis also done for further investigation.
as a sufficient predictor of HCC LI-RADS [2,3] and
the Barcelona Staging System [3,6] is well known as 2.1.3. Magnetic Resonance Imaging (MRI)
an indicator of the liver cancer stage. However, for
MRI is used for detailed body image analysis. It has
multiple diagnoses such as cirrhosis and benign,
some advantages over CT such as non-ionizing, high
hepatocellular and cholangiocarcinoma, we need
contrast, radiation, and wide-field imaging
intelligent computer-aided diagnostic systems.
capabilities. MRI uses magnetic waves to align the
Liver biopsy is a test for the degree of fibrosis and is photons of interest with the field. Thus the radio-
still considered the gold standard for stag frequency current pulses were accurate with a
production. However, it is an embodied process that magnetic field and were used to analyze tissue
can have consequences; such as 30 to 40% pain in particles. MRI of the abdomen can often be
performed with contrast-enhanced axial spin-echo object recognition. CNNs work by extracting
T1-weighted and fast spin-echo T2-weighted features from images by applying a series of
imaging. T2-weighted imaging is commonly convolutional and pooling layers.
preferred for abdominal examination [2,3,4].
There are many models of Convolutional Neural
In this project, we used CT scans to detect and Networks (CNNs). We are using CNN, ResNet and
classify liver cancer. We trained a deep-learning MobileNet which are forms of convolutional neural
model on a dataset of 33 cancer images, 34 non- networks, but they have some key important
cancerous images and a training set of 117 images. differences.
2.2. Image Processing
Medical images from CT scans are mostly structural
images that may contain artifacts, blurs, and
environmental device noise due to unintentional
human interference da because These factors can
affect image quality and diagnostic accuracy when
processed and analysed with state-of-the-art systems
are also available [3,4,5]. Nonlinear bilateral mean
filters are widely used for noise removal [4,5]. But
multivariate multidimensional wavelet-based
methods have also been proposed to reduce motion
artefacts and noise [5,6] Deep learning has also been
widely used to discover and remove various
distorted images. Several techniques such as artifact
removal, registration, noise removal, blurring and
segmentation are used for this purpose. We use
CNNs are a type of deep learning model well suited
Segmentation, based on a deep learning model
for image classification and object recognition tasks.
which has revolutionized image classification by
CNNs work by extracting features from the images
providing more accurate and efficient solutions.
by applying a series of layers. Here is the detailed
This example systematically implements the method
explanation of CNNs step-by-step
of visualization, where multiple filter layers are used
to the input image to extract high-quality features ➢ Image input: The CNN is provided with an
deep learning-based models can achieve accuracy as input image, such as a CT scan image of the
available today in a variety of image classification liver.
tasks on even more challenging datasets.
➢ Convolutional layers: Convolutional layers
2.3. Deep Learning Models apply a series of filters to the input image
to identify specific features. Each filter
Deep learning models are a type of machine learning
consists of a small matrix of weights
model that uses artificial neural networks with many
applied to a subset of the input image. The
hidden layers to learn from data. Deep learning
output of each convolutional layer is a
models help in identifying complex patterns and
feature map, which is a map of the
relationships in data that are difficult for traditional
activation values of each filter at each
machine learning models to learn. Deep learning
location in the input image.
models are used in various areas such as image
classification, speech recognition, object
➢ Pooling layers: Feature maps of
recognition, and natural language processing.
convolutional layers produced by
Convolutional Neural Networks (CNNs) are well-
downsampling (reducing the number of
suited for tasks such as image classification and
pixels in an image which indirectly reduces 2.4. Attribute Analysis
file size) pooling layers. This is done by
taking the average or maximum value of a 2.4.1. Features
subset of the feature map and assigning it
The development of Computer-Aided Diagnosis
to the corresponding output pixel. Pooling
(CAD) systems requires deep learning algorithms to
layers help reduce noise and improve the
be trained with discriminating features so that the
computational efficiency of the model.
system can detect any anomalies in its optimal
performance because the extraction of these features
➢ Fully connected layers: Fully connected
is classified in method appropriate over
layers produce the final result by
morphological and structural properties [4].
combining features extracted from
previous layers, such as class labels or 2.4.1.1. Morphological Features
probability distributions among different
class labels Table 1 shows the various common examples in
practice for morphological feature extraction. The
CNN (Convolutional Neural Network) is a type of
following examples are described as follows.
deep learning model which is well-suited for image
recognition and classification. This Model has been Statistical-based models, the statistical methods are
used in our project Non-Invasive Liver Cancer mainly divided into primary and secondary
Detection which is achieving an accuracy of components. The first-order feature provides us with
91.67%. statistical properties associated with the grey-level
distribution and an image, ignoring pixel proximity
ResNets is a new CNN introduced by ResNets in
or spatial properties. These parameters include
2015 that addresses the problem of vanishing
mean, slope, kurtosis peak, variance, skewness,
gradients, which complicates the training of deep
standard deviation, etc. [4]. Alternatively, secondary
CNN models. ResNets uses residual connectivity,
features depend on co-occurrence properties in
which skips a few layers in the network and directly
neighbouring pixels. These features specify some
connects the input layer to the output layer, which
spectral properties related to the image properties,
allows it to learn from its mistakes and improve its
such as grey-level co-occurrence matrix (GLCM)
performance. This Model has been used in our
and geometric property-based texture features, etc.
project Non-Invasive Liver Cancer Detection which
is achieving an accuracy of 60.87%. Similarly, few studies have examined texture
features based on geometric properties that provide
MobileNet is a small deep-learning system designed
information about the material profile and body
to run on mobile devices. This model uses depth-
structure. Typically, benign tumors have a smooth
separating diffractions to reduce the number of
profile, while malignant lesions tend to have
parameters and calculations required. This model
irregular shapes. Thus, spatial features such as
has shown comparable accuracy to other CNN
solidity, compactness, ROI area, radial length,
models but with a much smaller footprint. This
roundness, rectangularity, etc. Which provides
Model has been used in our project Non-Invasive
valuable specific information for lesion
Liver Cancer Detection which is achieving an
classification.
accuracy of 87.50% followed by CNN.
Processing-based models are rotated with a mask
Overall, we got that the CNN model has the highest
image of various sizes to quantify textural
accuracy which is followed by the MobileNetV1
information [4]. The first obtain such a feature by
model which is much better in comparison to the
graphically rotating a mask of size 5 x 5. Since then,
ResNet model.
the Laws’ features have been used in various fields
to find value-added features. The Frequency-Based
Models, several transformation models based on
frequency-domain features have been proposed [5-
7] these are shown to be more efficient than spatial- method to optimize the best combination of spatial
based features. features for a desired number of results. Based on
the loss function defined in the back-propagation,
Table 1
these weights are optimized to minimize the errors
Method Extracted Features through multiple runs. Since the results we obtained
Statistical Based Models are of the best quality, no further selection is
Gray-level Co- Correlation information, required. However, such algorithms require large
occurrence Matrix Inverse difference moment, data inputs to optimize the performance and avoid
(GLCM), Gray- Angular second moment overfitting.
level Difference Sum of the square, Gradient,
Matrix (GLDM) Sum of variance, Sum of 2.4.1.3. Feature Selection Based on
average, Sum of entropy, Algorithms
Contrast, Difference entropy,
A selection procedure is performed to select inputs
etc.
that have a strong statistical relationship with the
Gray-level Run Gray-level (GL) run length, output. Studies [4-6] have used some methods such
Length Matrix GL run number, Run as iterative elimination, statistical testing, genetic
(GLRLM) emphasis, GL nonuniformity, algorithms, etc.
etc.
Statistical testing is used to select features based on
First Order Mean, Peak, Slope, internal relationships with input and between each
Statistics (FOS) Variance, Standard output variable. These tests are considered valid
deviation, Kurtosis, keeping in mind some prior assumptions about the
Skewness etc. data distribution. Furthermore, features are also
Geometric features Coarseness, Eccentricity, selected based on the mean of freedom and
Periodicity, Roughness, minimum variance.
Regional area, Compactness,
Recursive feature elimination (RFE) [4] selects a
Margin, Circularity,
subset of features based on model performance.
Rectangularity, Periphery,
Some simple items are removed by repeated runs of
Border, Width, Depth etc.
the model and analysis of cross-validation scores.
Processing based Models Because of its iterative process, RFE requires high
Laws’ Mask Laws’ mask of length range, computational costs for large numbers of data sets.
Analysis edges, corners, texture, Therefore, highly correlated features should have to
shapes, etc be eliminated before using the RFE method for
feature selection.
Frequency-based Models
Transform domain Wavelet transform, Fourier Genetic algorithms [4] are another method of
power spectrum, Gabor selecting crossovers and mutations to find the best
wavelet transform, chromosome. A particular segment is counted as a
Contourlet transform, chromosome, whereas each gene has specific
Shearlet transform, Ranklet information such as a combination of selection
transform etc. segments according to the required trait locus to
select individual genes based on an extreme criterion
of fitness to the test survival. Multiple cross-overs
are performed to recombine the best chromosomes
2.4.1.2. Hierarchical Feature
and mutations for new arrangements from different
Recent studies [4] have used DNN (Deep neural genomes. In addition, some other methods, such as
network) models for logical feature extraction. NNS grasshopper or swarm particle optimization have
(Neural Network search) uses a back-propagation also been used for optimal feature selection [4,5,6].
2.4.2. Feature Extraction and Selection 4. Models: Accuracies Graph
Feature extraction and selection are the most
important part of the training of the models. Here we
are using Laws’ Mask Analysis which is a
processing base model that extracts the edges,
corners, shapes, texture, etc from the images as the
features during training. These extracted features are
the selected features on which checking of these
features will happen when the actual input Ct scan
image has been provided to the model. These
selected features by the models will help to
determine whether the input image has cancer or not.
3. System architecture
This figure shows an overview of the accuracies of
three different models we used: CNN, ResNet and
MobileNetV1. The graph shows that CNN has the
highest accuracy followed by MobileNetV1.
5. Future work
Future work on non-invasive deep learning-based
• Image input: The system takes a CT scan systems for liver cancer detection could focus on the
image of the liver as input. following areas.
• Pre-processing: The system pre-processes
• Improved accuracy and strength.
the image to improve its quality and
facilitate the deep learning model. This • Expansion of power.
includes steps such as contrast • Development of user-friendly interface.
enhancement, noise reduction, and • Dealing with social and moral concepts.
normalization. • Fusion of CT scan and MRI scan images.
• Segmentation using deep learning model: Overall, the future of non-invasive deep learning-
The system uses a deep learning model to based systems for detecting liver cancer is
segment the liver and gall bladder. Deep
promising. This project has the potential to
learning models help in identifying transform the diagnosis and treatment
complex patterns in data, making them of liver cancer.
ideally suited for tasks such as image
segmentation. 6. Result
• Tumor removal: The system removes the
tumor from the liver to remove all voxels Non-invasive deep learning-based systems for liver
that are not part of the tumor. cancer detection may revolutionize liver cancer
• Classifying: The system classifies tumor as diagnosis by reducing the need for invasive testing.
Cancerous or Non-cancerous. Such programs can detect liver cancer at an early
stage reducing the death rate, improving patient 6] Abhishek Midya, Jayasree Chakraborty, Rami
situation and potential benefits for its patients. Srouji, Raja R.Narayan, Thomas Boerner, Jian
Diagnostic features of liver cancer that are not Zheng, Linda M. Pak, John M. Creasy, Luz A.
implants can also provide significant cost savings. Escobar, Michael I. D'Angelica “Computerized
Work to develop a non-invasive device-based Diagnosis of Liver Tumors From CT Scans Using a
system for the detection of liver cancer has yielded Deep Neural Network Approach”, IEEE Journal of
very promising results. The models have Biomedical and Health Informatics Volume: 27,
demonstrated high accuracy in the detection of liver Issue: 5, May 2023.
cancer lesions, with an accuracy of 91.67% on the
7] Venkateswarlu Gavini, G.R. Jothi Lakshmi
CNN model, followed by 87.50% on MobileNetV1
,"Liver Tumor Grade Detection using CNN Based
which are much better compared to the ResNet
LSTM Model with Corelated Feature Set from CT
model achieving an accuracy of 60.87% on the
Images", Published in 2022 International
dataset, as well as feasibility in distinguishing
Conference on Automation 07 February 2023.
between non-cancerous and cancerous lesions. By
providing a detailed analysis of tumor burden and 8] Amritha M, Manimegalai P, "Liver Tumor
morphology, this system can help physicians Segmentation and Classification Using Deep
accurately diagnose a patient’s cancer. Learning" Published in 2023 Fifth International
Conference on Electrical, Computer and
7. Reference Communication Technologies (ICECCT), pp 638-
650, 21 July 2023.
1] Piyush Kumar Shukla, Mohammad Zakaria,
Vessam Atef Hatamleh, Hussam Taraji, and Basant 9]Chao-Lun Kuo, Shyi-Chyi Cheng, Chih-Lang Lin,
Tiwari, “AI-Driven Novel Approach to Liver Kuei-Fang Hsiao, Shang-Hung Lee, "Texture-based
Cancer Diagnosis & Prognosis Using Cascaded treatment prediction by automatic liver tumor
Fully Convolutional Neural Networks”, Journal of segmentation on computed tomography", Published
Healthcare Engineering Vol.23, pp 7-14, Sep 2022. in 2017 International Conference on Computer,
Information and Telecommunication Systems
2] Anita Bakrania, Narotam Joshi, Xun Zhao, Gang
(CITS), pp 406-423,14 September 2017.
Zheng, Mamata Bhat,” Artificial intelligence in liver
cancer: decoding the impact of machine learning 10] Aparna P R, Libish T M, "Automatic
models on clinical diagnosis of primary liver cancer segmentation and classification of the liver tumor
and liver cancer of metastasis”, Journal of using deep learning algorithms", Published in 2023
Pharmacological Research, pp 12-20, Feb 2023. 3rd International Conference on Advances in
Computing, Communication, Embedded and Secure
3] Esam Othmen, Muhammad Mahmood, Habib
Systems (ACCESS), pp 146-151, 07 August 2023.
Dhari, Meena Ibrahim, Awais Mahmood, and
Hatem Abdulkader,” Automated detection of liver 11]Hameedur Rahman, Tanvir Fatima Naik Bukht,
cancer using pre-trained hybrid models”, National Azhar Imran, Junaid Tariq, Shanshan Tu,
Institutes of Health July 2022. Abdulkareem Alzahrani, “A deep learning approach
for liver and tumor segmentation in CT images using
4] Ryan Azam Khan, Yigang Luo, and Fang Xiang
ResUNet”, Publish in Bioengineering 2022 and
Wu,” Machine Learning Based Liver Disease
article of Artificial intelligence in Advanced
Evaluation: A Systematic Review”, Journal of
Medical Imaging 5th Aug 2022.
Neurocomputing Volume 468 11 January 2022.
12]Xin Dong, Yizhao Zhou, Lantian Wang,
5]Heng Zhang, Cawen Luo, Ren Deng, Shenglin Li,
Jingfeng Peng, Yanbo Lao, Yiqun Fan, “Liver
Shukai Duan,” Deep learning-based CT imaging for
Cancer Detection Using Hybridized Fully
detection of liver tumors”, National Library of
Convolutional Neural Network Based on Deep
Medicine article May 2022.
Learning Framework”, IEEE (Access), pp 99, July
2022
13]Josephine Yu Yan Yap, Laura Shih Hui Goh,
Ashley Jun Wei Lim, Samuel S. Chong, Lee Jin
Lim, Caroline G. Lee, “Machine learning identifies
a signature of Nine exosomal RNAs that predicts
hepatocellular carcinoma”, Journal of Cancers 2023
Vol. 15, 24th July 2023.
