Multi-Organ Segmentation Network for Female Pelvic MR Images Based on Hierarchical Decoupled Fusion and Multi-Scale Feature Processing

Abstract

Accurate segmentation of the female pelvic structures is crucial for diagnosing and treating female pelvic disorders. However, the complex and variable shapes of the pelvic organs, along with their blurred boundary features, present significant challenges. To address the aforementioned issues, HDFMFP-Net, a multi-organ segmentation network for female pelvic MR images, is constructed based on hierarchical decoupled fusion and multi-scale feature processing. Firstly, to address the problem of the complex and variable shapes of the pelvic organs, a hierarchical decoupled fusion module is constructed as a fundamental unit. This module is used to build a lightweight encoder-decoder structure, which reduces the accumulation of estimated bias while effectively extracting the complex boundary features of pelvic organs through multi-scale feature learning and cross-layer information interaction. Secondly, to address the challenge posed by the blurred boundary features of pelvic organs, a multi-scale feature processing module is integrated into the skip connections. This module captures features at multiple scales from both local and global perspectives, enhancing the network's ability to represent the fuzzy boundary features of organs. Finally, the algorithm was evaluated on the female pelvic dataset to segment the bladder, uterus, and rectum in female pelvic images, and the mIoU, mDice, and mPA of HDFMFP-Net reached 92.98%, 96.3%, and 96.24%, respectively, with a model size of only 2.80 M. The results demonstrate that the proposed method offers a promising approach for the automatic segmentation of female pelvic organs.