AO-TransUNet: A multi-attention optimization network for COVID-19 and medical image segmentation

Background

The COVID-19 pandemic has created a significant demand for accurate and efficient diagnostic tools to support effective disease management. Medical images related to COVID-19 present unique challenges, as the lesions often appear in various forms (e.g., ground glass shadows and consolidation shadows) that vary significantly in size, shape, and distribution. Additionally, these lesions can share similar gray levels or texture features with normal lung tissue, making it difficult to delineate clear boundaries between affected and healthy areas.

Methods and procedures

To address these challenges, the paper introduces a novel network called Attention Optimization TransUNet (AO-TransUNet), which builds upon the foundation of TransUNet. The method incorporates multiple attention mechanisms aimed at minimizing the loss of key information during the dimensionality reduction phase of segmentation. AO-TransUNet enhances dense interactions across all pixels, ensuring that morphological details and feature information of the lesions are preserved. This comprehensive approach improves the model's ability to detect subtle structural differences and effectively segment complex COVID-19 lesions.

Results

The performance of AO-TransUNet was validated through experimental evaluations on four datasets. The results demonstrated that AO-TransUNet outperformed existing state-of-the-art networks, showcasing its effectiveness in medical image segmentation.

Conclusion:

The study underscores the potential of AO-TransUNet to contribute to the field of medical image segmentation by addressing the challenges of complex and variable lesions, such as those seen in COVID-19. The method's ability to maintain morphological details and improve pixel-level interactions suggests broader applicability for other medical image analysis challenges. All code is available at https://github.com/xiaqi7/AO-TransUNet.