MFNet: Multi-fusion network for medical image segmentation

Abstract

Medical image segmentation can distinguish and determine various structures, tissues, or lesions, providing crucial information for clinical diagnosis. However, it faces numerous challenges. On the one hand, medical images possess complex structures, diverse morphologies, uneven contrast, and blurred borders between target tissues and the background, all of which complicate the segmentation process. On the other hand, there exists semantic gaps between low-level and high-level features as well as between the encoder and decoder, which greatly impacts the segmentation effectiveness. In order to overcome these drawbacks, a Multi-Fusion Network (MFNet) is presented to integrate semantic and feature fusion. In this method, two novel modules including Multi-Level Semantic Fusion (MLSF) module and Multi-Scale Progressive Fusion (MSPF) module are designed to heighten the representation capability of capturing diverse semantic and scale information. Moreover, a Multi-Stage Progressive Fusion Decoder (MSPFD) model is developed to substitute the traditional bottom-up aggregation decoder with a hierarchical fusion decoder to integrate features from different levels step by step. Meanwhile, an Interaction and Fusion of Adjacent Levels (IFAL) module is introduced to merge higher-level and lower-level features, effectively learning the semantic consistency and reducing this semantic gap. To evaluate the performance of our designed network, we evaluate it against several SOTA methods on four benchmark datasets including ISIC2018, GlaS, ACDC, and Synapse. Comparative results indicate that MFNet achieves remarkable ability on medical image segmentation.