Enhancing Deep Learning-Based Subabdominal MR Image Segmentation During Rectal Cancer Treatment: Exploiting Multiscale Feature Pyramid Network and Bidirectional Cross-Attention Mechanism.

Abstract

Background: This study is aimed at solving the misalignment and semantic gap caused by multiple convolutional and pooling operations in U-Net while segmenting subabdominal MR images during rectal cancer treatment. Methods: We propose a new approach for MR Image Segmentation based on a multiscale feature pyramid network and a bidirectional cross-attention mechanism. Our approach comprises two innovative modules: (1) We use dilated convolution and a multiscale feature pyramid network in the encoding phase to mitigate the semantic gap, and (2) we implement a bidirectional cross-attention mechanism to preserve spatial information in U-Net and reduce misalignment. Results: Experimental results on a subabdominal MR image dataset demonstrate that our proposed method outperforms existing methods. Conclusion: A multiscale feature pyramid network effectively reduces the semantic gap, and the bidirectional cross-attention mechanism facilitates feature alignment between the encoding and decoding stages.