[Paper Review] COVID-19 Chest CT Image Segmentation -- A Deep Convolutional Neural Network Solution

Qingsen Yan, Bo Wang+8 more|arXiv (Cornell University)|Apr 23, 2020

COVID-19 diagnosis using AI25 references115 citations

TL;DR

The paper introduces COVID-SegNet, a 3D CNN with a Feature Variation (FV) block and Progressive Atrous Spatial Pyramid Pooling (PASPP) to automatically segment COVID-19 infection regions and lungs from chest CT scans, trained on a large COVID-19 CT dataset.

ABSTRACT

A novel coronavirus disease 2019 (COVID-19) was detected and has spread rapidly across various countries around the world since the end of the year 2019, Computed Tomography (CT) images have been used as a crucial alternative to the time-consuming RT-PCR test. However, pure manual segmentation of CT images faces a serious challenge with the increase of suspected cases, resulting in urgent requirements for accurate and automatic segmentation of COVID-19 infections. Unfortunately, since the imaging characteristics of the COVID-19 infection are diverse and similar to the backgrounds, existing medical image segmentation methods cannot achieve satisfactory performance. In this work, we try to establish a new deep convolutional neural network tailored for segmenting the chest CT images with COVID-19 infections. We firstly maintain a large and new chest CT image dataset consisting of 165,667 annotated chest CT images from 861 patients with confirmed COVID-19. Inspired by the observation that the boundary of the infected lung can be enhanced by adjusting the global intensity, in the proposed deep CNN, we introduce a feature variation block which adaptively adjusts the global properties of the features for segmenting COVID-19 infection. The proposed FV block can enhance the capability of feature representation effectively and adaptively for diverse cases. We fuse features at different scales by proposing Progressive Atrous Spatial Pyramid Pooling to handle the sophisticated infection areas with diverse appearance and shapes. We conducted experiments on the data collected in China and Germany and show that the proposed deep CNN can produce impressive performance effectively.

Motivation & Objective

• Motivate fast, accurate automatic segmentation of COVID-19 infections in chest CT to assist diagnosis and treatment.
• Create a large annotated CT dataset of COVID-19 cases to enable robust deep learning-based segmentation.
• Develop a tailored network architecture that handles diverse infection appearances and boundaries.

Proposed method

• Propose COVID-SegNet, a 3D encoder-decoder network with residual blocks and two novel modules: Feature Variation (FV) block and Progressive Atrous Spatial Pyramid Pooling (PASPP).
• FV block combines contrast enhancement, position sensitivity, and identity branches to adaptively highlight infection boundaries via channel and spatial attention.
• PASPP progressively fuses multi-scale features using atrous convolutions with increasing dilation rates to capture varied infection sizes and shapes.
• Train end-to-end on a large annotated CT dataset (21,658 images from 861 patients) with a balanced loss combining Dice and cross-entropy.
• Evaluate with Dice, sensitivity, and precision metrics for both COVID-19 infection and lung segmentation; compare against FCN, UNet, UNet++, and VNet.

Experimental results

Research questions

• RQ1 Can a 3D CNN with specialized attention and multi-scale pooling improve COVID-19 infection segmentation on chest CT beyond existing architectures?
• RQ2 Does the FV block improve boundary delineation of infection regions, and does PASPP improve multi-scale contextual feature fusion?
• RQ3 How well does the proposed model generalize to data from different centers (China and Germany) on COVID-19 segmentation and lung segmentation?
• RQ4 What are the quantitative gains of COVID-SegNet over state-of-the-art methods across standard segmentation metrics?

Key findings

• COVID-SegNet achieves higher COVID-19 infection segmentation Dice of 0.726, with sensitivity 0.751 and precision 0.726 on the domestic test set.
• Lung segmentation Dice reaches 0.987 with sensitivity 0.986 and precision 0.990, outperforming several baselines.
• On the Germany dataset, COVID-SegNet produces near-manual quality COVID-19 segmentation, including small infection regions, outperforming FCN, UNet, UNet++, and VNet.
• Ablation studies show FV and PASPP blocks substantially improve COVID-19 and lung segmentation compared to baseline UNet4 and alternatives.
• Progressive fusion in PASPP yields better performance than standard ASPP or ResASPP due to more effective multi-scale feature integration.