[Paper Review] An automatic COVID-19 CT segmentation based on U-Net with attention mechanism

Tongxue Zhou, Stéphane Canu+1 more|arXiv (Cornell University)|Apr 14, 2020

COVID-19 diagnosis using AI54 citations

TL;DR

This paper proposes an attention-augmented U-Net with focal Tversky loss for automatic COVID-19 lung lesion segmentation in CT scans. By integrating spatial and channel attention mechanisms to refine feature representations and using a loss function tailored for small lesions, the method achieves high accuracy (Dice: 83.1%) and speed (0.29 s per slice), demonstrating strong performance on a 473-slice dataset.

ABSTRACT

The coronavirus disease (COVID-19) pandemic has led to a devastating effect on the global public health. Computed Tomography (CT) is an effective tool in the screening of COVID-19. It is of great importance to rapidly and accurately segment COVID-19 from CT to help diagnostic and patient monitoring. In this paper, we propose a U-Net based segmentation network using attention mechanism. As not all the features extracted from the encoders are useful for segmentation, we propose to incorporate an attention mechanism including a spatial and a channel attention, to a U-Net architecture to re-weight the feature representation spatially and channel-wise to capture rich contextual relationships for better feature representation. In addition, the focal tversky loss is introduced to deal with small lesion segmentation. The experiment results, evaluated on a COVID-19 CT segmentation dataset where 473 CT slices are available, demonstrate the proposed method can achieve an accurate and rapid segmentation on COVID-19 segmentation. The method takes only 0.29 second to segment a single CT slice. The obtained Dice Score, Sensitivity and Specificity are 83.1%, 86.7% and 99.3%, respectively.

Motivation & Objective

• Address the challenge of accurate and rapid segmentation of COVID-19 lung lesions in CT scans to support diagnosis and patient monitoring.
• Improve feature representation in U-Net by selectively emphasizing relevant spatial and channel-wise features using attention mechanisms.
• Overcome the class imbalance issue in small lesion segmentation by employing a focal Tversky loss function.
• Achieve high segmentation performance with minimal inference time, suitable for clinical deployment.

Proposed method

• Integrate a dual attention mechanism—spatial and channel attention—into the U-Net encoder-decoder architecture to re-weight feature maps based on contextual importance.
• Apply spatial attention to emphasize informative spatial regions and channel attention to highlight discriminative feature channels.
• Use the focal Tversky loss to focus training on hard-to-segment regions, especially small lesions, by down-weighting easy negatives.
• Train the network end-to-end on a dataset of 473 CT slices with ground-truth annotations for lung lesions.
• Leverage skip connections from encoder to decoder to preserve spatial details during up-sampling.
• Optimize the model using stochastic gradient descent with a learning rate schedule to improve convergence.

Experimental results

Research questions

• RQ1 Can attention mechanisms improve feature representation in U-Net for more accurate COVID-19 lesion segmentation in CT scans?
• RQ2 Does the focal Tversky loss enhance segmentation performance on small and sparse lung lesions compared to standard loss functions?
• RQ3 Can the proposed method achieve both high accuracy and low inference time for real-time clinical use?
• RQ4 How does the combination of attention and focal Tversky loss compare to standard U-Net in terms of Dice score, sensitivity, and specificity?

Key findings

• The proposed method achieved a Dice score of 83.1% on the test set, indicating strong overlap between predicted and ground-truth lesions.
• Sensitivity reached 86.7%, showing the model effectively detects most actual lesions despite their small size.
• Specificity was 99.3%, indicating very few false positives, which is crucial for clinical reliability.
• The model segmented a single CT slice in just 0.29 seconds, demonstrating high inference speed suitable for clinical deployment.
• The integration of spatial and channel attention improved feature representation by focusing on relevant regions and channels.
• The focal Tversky loss significantly improved performance on small lesions by reducing the impact of easy negatives during training.