COVID-19 CT Scan Lesion Segmentation (Short Doc)

Muhammad Shariq Shoaib MSCSF21M507

A semester project for the course of Deep Learning CS-568

Abstract

Image Processing has been always considered as top ranked research areas in Computer Sciences. Where Medical Image Processing/Analysis is among the key research topics as there are infinite problems relevant to medical domains that requires the fast computational power to help in order to save lives. In this project the goal is to design a lesion segmentation system by using a dataset of Lungs CT scans having lesions in Covid-19. Dataset consists of 2729 CT-scan image and ground truth mask pairs. All types of lesions are mapped to white color in the ground truth for consistency across dataset. This project will be done in incremental phases, where firstly I’ll be considering for such pre-processing techniques which result in highlighting lesions region the most . In the second phase, I’ll be looking forward for segmenting the lesion and comparing different techniques for segmenting in order to achieve better results. If these phases are done well before time and I’ll be comparing other researchers’ methodology and improve further.

Dataset Source: https://www.kaggle.com/maedemaftouni/covid19-ct-scan-lesion-segmentation-dataset Keywords: Image Processing; Medical Image Analysis; Covid-19; CT Scan; Lesion; Segmentation; Ground Truth;

APEER - Labeling tool

Code Help Source: https://github.com/bnsreenu/python_for_image_processing_APEER/blob/master/tutorial117_building_unet_using_encoder_decoder_blocks.ipynb

Introduction

A unexpected outbreak of the rare respiratory disease was reported in Wuhan, China, near the end of December 2019. The Chinese Center for Disease Control and Prevention examined the lower respiratory tracts of sick patients and discovered COVID19, a novel coronavirus. COVID-19 is a systemic infection that primarily affects the endothelium layer of the lungs and spreads from person to person, whether symptomatic or asymptomatic. Post-COVID-19 pulmonary fibrosis is characterised as the occurrence of fibrotic CT alterations on follow-up studies that are commonly clinically associated with decreased lung function. Even though histopathological patterns in post-COVID-19 pulmonary fibrosis have yet to be discovered, they are thought to be those of an organising and fibrotic phase of diffuse alveolar damage (DAD), fibrosing nonspecific interstitial pneumonitis, and organizing pneumonia. CT features being potentially identified in pulmonary fibrosis secondary to COVID-19 pneumonia have been suggested to include the presence of architectural distortion, reticular lesions, traction bronchiolectasis, ground-glass opacity, mosaic attenuation, and honeycombing in consideration of presumed histopathologic patterns. But according to the study by Pan et al (1), the CT features seen at one-year follow-up studies were those of fibrotic-like linear or multifocal reticular and/or cystic lesions. There were no typical CT DAD findings, such as fibrosing nonspecific interstitial pneumonitis or organised pneumonia patterns. Follow-up CT characteristics of severe acute respiratory syndrome-associated coronavirus (SARS-CoV) infection produced comparable results. Patients with underlying interstitial lung disease, particularly those with poor lung function and obesity, are at a higher risk of death from COVID-19. Finally, follow-up CT scans of COVID-19 pneumonia patients reveal persistent fibrotic-like linear or multifocal reticular/cystic lesions. So if we segment the lesion region, it’ll be helpful for post COVID treatment by looking at the affected areas of lungs.

Literature Review

1. Lung Lesion Segmentation Using Gaussian Filter and Discrete Wavelet Transform By Kamil Dimililer, Ali Hesri and Yoney Kirsal Ever ( 10.1051/itmconf/20171101018 ) 1 Electric and Electronic Engineering Department, Near East University, North Cyprus, Mersin 10 Turkey 2 Software Engineering Department, Near East University, North Cyprus, Mersin 10 Turkey Lesions inside the lungs are called a great early symptom for detection of Lung cancer. Lung cancer is a growing tumour that are cells covering the airways of the major respiratory system. Here they’ve used CT scans and applied image processing different applications and investigated the effectiveness of different types of discrete wavelets to obtain accurate results. Haar Wavelet, Db7 and Bi-orthogonal wavelet are being used along with Median and Gaussian Kernels.

2. COVID-19 Lesion Segmentation Using Lung CT Scan Images: Comparative Study Based on Active Contour Models By Younes Akbari, Hanadi Hassen, Somaya Al-Maadeed and Susu M.Zughaier ( https://doi.org/10.3390/app11178039 ) 1 Department of Computer Science and Engineering, Qatar University, Doha 122104, Qatar 2 College of Medicine, QU Health, Qatar University, Doha 122104, Qatar Pneumonia being a lung infection is highly threatening. COVID-19 causing pneumonia is being segmented here using CT scans to find how effective Active Contour Models (ACM) are. This research provides the basis on which further ACM related Lung Segmentation can be done in the COVID-19 domain. They’ve shown how initially contour results and improvements in result if (Gac, C-V, MAC, Lsacm, Frbacm) Models of Active Contour are used.

3. Dual-branch combination network (DCN): Towards accurate diagnosis and lesion segmentation of COVID-19 using CT images By Kai Gao, Jianpo Su, Zhongbiao Jiang et al. ( https://doi.org/10.1016/j.media.2020.101836 ) 1 College of Intelligence Science and Technology, National University of Defense Technology, Changsha, Hunan, China 2 Department of Radiology, Third Xiangya Hospital, Central South University, Changsha, Hunan, China 3 Department of Radiology, Second Xiangya Hospital, Central South University, Changsha, Hunan, China College of Computer Science and Technology, National University of Defense Technology, Changsha, Hunan, China As the COVID-19 outbroke from Wuhan China and as a result, many different diagnostic tools were developed, CT(computer tomography)-based diagnostic tool faced further complications due to lack of adequate manuallydelineated samples. So dual-branch combination network (DCN) was developed to achieve classification and lesion segmentation which resulted in high accuracy. CT scan images are preprocessed first using U-net and then DCN is applied which uses FC Net to get lesion segmentation.

4. Lung Lesion Detection in CT Scan Images Using the Fuzzy Local Information Cluster Means (FLICM) Automatic Segmentation Algorithm and Back Propagation Network Classification By M Lavanya and P Muthu Kannan ( https://dx.doi.org/10.22034%2FAPJCP.2017.18.12.3395 ) 1 Department of Electrical and Electronics Engineering, Saveetha School of Engineering, Saveetha University, Thandalam, Chennai-602 105, India. Lungs’ cancer has always been a deadly and diagnostic method for lung tumours that weren’t up to the mark then due to variation in lesions. These researchers proposed a segmentation technique using the FLI Cluster Mean algorithm for segmentation and further clustering is done to achieve accurate results after which Edge detection was done to extract features and Classification is achieved using Back Propagation Network. It resulted in far better accuracy than existing methods of that time.

5. Dense GAN and multi-layer attention based lesion segmentation method for COVID-19 CT images By Ju Zhang, Lunduan Yu at el. ( https://doi.org/10.1016/j.bspc.2021.102901 ) 1 Zhijiang College of Zhejiang University of Technology, Shaoxing 312030, China 2 College of Computer Science and Technology Zhejiang University of Technology, Hangzhou 310023, China 3 College of Information Engineering, Zhejiang University of Technology, Hangzhou 310023, China 4 Department of Medical Imaging, Zhejiang Hospital, Hangzhou 310013, China Testing and screening during COVID-19 pandemic was a headache for governments. AS CT-scan images were clear diagnostic medium, medical imaging researchers proposed a technique in which novel Dense GAN was developed to get clear images and U-NET combined with multi-layer attention mechanism was used to get accurate lesion segmentation. Further related work resulted in Dense Generative Adversarial Network(DGAN) Precision by this proposed method was as high up to 0.94.

Framework

Tensorflow Keras, Numpy, Matplotlib and other dependencies.

How to Run?

Use Anaconda 3 with default environment and make sure you have latest numpy. Connect Kernal with Jupyter notebook and Run :)

Methodology

1 Pre-processing

Due to the limitation of hardware resources, and time, No preprocessing was done and all images are fed as they were collected

2 Model Selection

Currently, I have used U-Net with the attention model having following details. Again due to hardware limitation 851 unique images were selected to train our DL model (U-Net) as it allows to have a versatile boundary.

11 hidden layers Total params: 31,117,985 Trainable params: 31,105,953 Non-trainable params: 12,032

3 Evaluation

Intersection over union IoU has been used as the evaluation metric. And for training accuracy was being calculated at each epoch.

Results

Data was divided into 80 percent training and 20 percent testing and by UNet with attention model we achieved IoU of 0.60990273 on the pixel based resulting in form of mask. Predicted masks are attached in the PDF report attached(DL_Project_MSCSF21M507.pdf) in current directory.