TransCA: Transformer-Based Context Aggregator for Medical Image Segmentation

The diversity and complexity of medical images, as well as the problems such as boundary blurring, bring challenges to the segmentation task. However, the traditional mainstream methods are still faced with problems such as low accuracy and insufficient robustness to different types of images. Transformer and its variants have been widely used for medical image segmentation. However, its overall processing of input sequences makes it difficult for the model to focus on the relationship between different channels. To alleviate this problem, we propose a Transformer-based Context Aggregator Network. First, we use an encoder-decoder structure to extract feature representations of different levels. Secondly, we design a block Context Aggregator to improve the information extraction capability of the model in channel dimension. Finally, we introduce Context Aggregator into decoder and propose TransCA model. Extensive experiments on three datasets demonstrate that our proposed model outperforms current advanced methods in the segmentation of skin lesion images and cell nucleus images. Furthermore, our approach exhibits greater robustness in various scenarios with blurred object boundaries and occlusions.