[论文解读] Residual Attention U-Net for Automated Multi-Class Segmentation of COVID-19 Chest CT Images

Xiaocong Chen, Lina Yao+1 more|arXiv (Cornell University)|Apr 12, 2020

COVID-19 diagnosis using AI参考文献 35被引用 187

一句话总结

论文提出了一种 Residual Attention U-Net，将 Aggregated Residual Transformations (ResNeXt blocks) 与 locality-sensitive hashing attention 相结合，在 chest CT 图像上实现多类别 COVID-19 感染区域分割，并在公开数据集上相比 U-Net 显著提升。

ABSTRACT

The novel coronavirus disease 2019 (COVID-19) has been spreading rapidly around the world and caused significant impact on the public health and economy. However, there is still lack of studies on effectively quantifying the lung infection caused by COVID-19. As a basic but challenging task of the diagnostic framework, segmentation plays a crucial role in accurate quantification of COVID-19 infection measured by computed tomography (CT) images. To this end, we proposed a novel deep learning algorithm for automated segmentation of multiple COVID-19 infection regions. Specifically, we use the Aggregated Residual Transformations to learn a robust and expressive feature representation and apply the soft attention mechanism to improve the capability of the model to distinguish a variety of symptoms of the COVID-19. With a public CT image dataset, we validate the efficacy of the proposed algorithm in comparison with other competing methods. Experimental results demonstrate the outstanding performance of our algorithm for automated segmentation of COVID-19 Chest CT images. Our study provides a promising deep leaning-based segmentation tool to lay a foundation to quantitative diagnosis of COVID-19 lung infection in CT images.

研究动机与目标

• Motivate automated multi-class segmentation of COVID-19 lung infection regions in chest CT images.
• Develop a deep learning model that improves feature representation and attention to discriminate infection patterns.
• Demonstrate superiority over baseline U-Net and analyze component contributions through ablation.

提出的方法

• Use Aggregated Residual Transformations (ResNeXt blocks) in the encoder to learn robust features.
• Incorporate a locality-sensitive hashing (LSH) attention mechanism in the decoder to enhance multi-class segmentation.
• Employ skip connections in a U-Net-like architecture to fuse encoder and decoder features.
• Train with multi-class cross-entropy loss for pixel-wise labeling across multiple infection classes and background.
• Perform data augmentation (rotations and scaling) and compare with baseline U-Net under augmentation and non-augmentation settings.
• Evaluate with Dice Score, accuracy, and precision using 10-fold cross-validation on a public COVID-19 CT dataset.

实验结果

研究问题

• RQ1 Does a Residual Attention U-Net improve multi-class segmentation of COVID-19 infection regions over standard U-Net?
• RQ2 What is the contribution of ResNeXt blocks and LSH-based attention to segmentation performance?
• RQ3 How does the model perform under data augmentation vs. no augmentation on a public dataset?

主要发现

• The proposed model achieves higher segmentation performance than U-Net across metrics with data augmentation: DSC 0.94, Accuracy 0.89, Precision 0.95 for the proposed method; U-Net yields DSC 0.82, Accuracy 0.79, Precision 0.83.
• With no augmentation, the proposed method still outperforms U-Net (DSC 0.83 vs 0.75; Accuracy 0.79 vs 0.70; Precision 0.82 vs 0.72).
• The reported improvements over U-Net are 14.6% (DSC), 12.7% (Accuracy), and 14.5% (Precision) with augmentation, and 10.7% (DSC), 12.9% (Accuracy), and 13.9% (Precision) without augmentation.
• Ablation studies show that removing either Attention or ResNeXt blocks reduces performance, indicating both components contribute to gains over U-Net.
• Model without Attention (M-A) achieves DSC 0.85, Acc 0.82, Precision 0.84 (with augmentation) and DSC 0.79, Acc 0.74, Precision 0.77 (no augmentation); Model without ResNeXt (M-R) achieves DSC 0.84, Acc 0.81, Precision 0.83 (with augmentation) and DSC 0.77, Acc 0.76, Precision 0.77 (no augmentation).
• Overall, the approach yields a promising tool for automated, multi-class COVID-19 infection quantification in CT images.