DBFA-Net: Diffusion-Based Bidirectional Fusion and Axial Attention Network for medical image segmentation

Abstract

In recent years, Denoising Diffusion Probabilistic Models (DDPMs) have shown strong modeling capability and are increasingly applied to medical image segmentation tasks. However, existing diffusion models still struggle with fine-grained structures and multi-scale context, particularly with semantic–noise feature fusion and global dependency modeling. To address these challenges, this paper proposes a Diffusion-Based Bidirectional Fusion and Axial Attention Network (DBFA-Net). Specifically, we design a Bidirectional Feature Fusion Module (BiFFM) that facilitates deep fusion of semantic and noise features through a bidirectional guidance mechanism. Additionally, we introduce a Frequency-domain Axial Cross-Attention Module (FACAM), which captures cross-scale global dependencies via axial attention in the frequency domain, enhancing long-range dependency modeling. Moreover, we introduce a Multi-Scale Axial Aggregation Attention Module (MSA3M) to aggregate multi-scale spatial information during the diffusion decoding process, enhancing the network’s ability to capture fine-grained structures. We conducted experiments on three public datasets to evaluate the effectiveness of DBFA-Net and compared it with existing methods. The results show that DBFA-Net achieves competitive or superior segmentation performance compared with most of the evaluated methods under the adopted experimental settings.