Medical Image Segmentation Based on TransUnet and A2B

Abstract

The objective of medical picture segmentation technology is to accurately delineate and partition regions affected by pathology, which is of great significance to medical diagnosis and treatment. Thanks to deep learning's quick development, prior efforts mainly utilize the convolutional neural networks equipped with powerful feature representation capability. In this paper, to make the segmentation performance a petter appearance, we introduce a new model in medical image segmentation technology, TransUNet, which combines Transformer and U-Net and uses labeled and unlabeled images for training. To enhance the feature quality and boost the accuracy and robustness of the segmentation results, we further introduce the A2B (Attention in Attention Block) attention mechanism in the TransUNet backbone network. At the same time, we use the CNN-Transformer hybrid model and conduct ablation experiments to verify the improvement effect of the performance indicators. The empirical findings show that TransUNet has accomplished success in tasks requiring medical picture segmentation and can help doctors and researchers more accurately identify and locate key structures.