[论文解读] COVID CT-Net: Predicting Covid-19 From Chest CT Images Using Attentional Convolutional Network

Shakib Yazdani, Shervin Minaee+3 more|arXiv (Cornell University)|Sep 10, 2020

COVID-19 diagnosis using AI参考文献 21被引用 42

一句话总结

本文提出一种带注意残差卷积网络，用于从胸部CT图像分类COVID-19，注意力图突出感染区域并公开发布标注的感染区域。

ABSTRACT

The novel corona-virus disease (COVID-19) pandemic has caused a major outbreak in more than 200 countries around the world, leading to a severe impact on the health and life of many people globally. As of Aug 25th of 2020, more than 20 million people are infected, and more than 800,000 death are reported. Computed Tomography (CT) images can be used as a as an alternative to the time-consuming "reverse transcription polymerase chain reaction (RT-PCR)" test, to detect COVID-19. In this work we developed a deep learning framework to predict COVID-19 from CT images. We propose to use an attentional convolution network, which can focus on the infected areas of chest, enabling it to perform a more accurate prediction. We trained our model on a dataset of more than 2000 CT images, and report its performance in terms of various popular metrics, such as sensitivity, specificity, area under the curve, and also precision-recall curve, and achieve very promising results. We also provide a visualization of the attention maps of the model for several test images, and show that our model is attending to the infected regions as intended. In addition to developing a machine learning modeling framework, we also provide the manual annotation of the potentionally infected regions of chest, with the help of a board-certified radiologist, and make that publicly available for other researchers.

研究动机与目标

• Motivate rapid COVID-19 diagnosis using chest CT when RT-PCR is time-consuming or inconclusive.
• Develop an attentional convolutional framework that focuses on infected lung regions for improved prediction.
• Provide qualitative and quantitative analysis including attention visualization and comparison to baselines.
• Offer manually annotated infected regions to support further research.

提出的方法

• Adopt a residual attention network where attention modules combine trunk and soft mask branches to emphasize infected areas.
• Use a single ResNet-56 based attention module unit within a residual attention framework (parameters p=1, t=2, r=1).
• Train on a public SARS-CoV-2 CT scan dataset and augment data to increase training samples by a factor of 4.
• Evaluate with sensitivity, specificity, ROC (AUC), and precision-recall analyses; visualize attention via Grad-CAM and Grad-CAM++.
• Annotate infected regions with radiologist input and release these annotations publicly.

实验结果

研究问题

• RQ1 Can an attentional convolutional network improve COVID-19 prediction accuracy from chest CT images compared with baselines?
• RQ2 Do the model's attention maps correspond to radiologist-identified infected regions in the lungs?
• RQ3 What is the impact of data augmentation and hyper-parameter choices on predictive performance?
• RQ4 Can the method provide reliable visualization of infected regions to aid clinical interpretation?

主要发现

• The model achieves competitive accuracy on the COVID vs non-COVID task with multiple optimizer choices (e.g., Adam with lr 0.001 yields 0.907±0.001 accuracy).
• Using BCE loss with lr 0.01 leads to 0.920±0.001 accuracy in validation.
• Attention maps (Grad-CAM and Grad-CAM++) predominantly highlight infected lung regions, aligning with radiologist annotations.
• The approach provides ROC and precision-recall analyses and histograms of predicted scores showing separation between classes.
• Data augmentation increased training samples by a factor of 4, aiding model performance.
• The study provides visually interpretable attention maps and generates a heatmap of potentially infected regions via a sliding-window occlusion method.