Frequency-domain Guided Attention-Differential Enhancement Network for medical image segmentation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Medical image segmentation remains a critical yet challenging task in pattern recognition, as extracting discriminative features is often hindered by complex anatomical structures with indistinct boundaries, low contrast, overlapping tissues, and inherent imaging noise. Consequently, these characteristics disrupt stable pattern representations, making segmentation approaches sensitive to intensity variations and anatomical ambiguities. To address these issues, we propose a Frequency-domain Guided Attention-Differential Enhancement Network (FDANet) for medical image segmentation. Our FDANet explicitly exploits frequency-domain representations, where low-frequency components preserve global anatomical layouts and high-frequency components emphasize structural details, while rich semantic information is simultaneously mined from the spatial domain. Additionally, to mitigate distribution discrepancies between the spatial and frequency domains while suppressing noise introduced during the fusion process, we design a Frequency-domain Differential Fusion Module (FDFM). Specifically, we first obtain two sets of feature embeddings in the frequency domain and spatial domain respectively, and utilize 2D Rotary Position Embedding (RoPE) to embed 2D relative spatial position information into these vectors to generate position-aware feature representations. Then we perform two independent self-attention operations within the spatial and frequency domains respectively, concatenate the resulting attention weights, and apply a learnable differential mechanism to adaptively attenuate redundant and noisy activations, thereby enabling more reliable exploitation of complementary cues. Finally, Semantic Consistency Loss (SCL) is employed to enforce semantic alignment between the fused frequency-domain and spatial-domain features, further enhancing the coherence of representations. We conduct extensive experiments on the CXRS, ISIC17, ISIC18, VinDr-RibCXR, and FracAtlas datasets. Experimental results and visualizations show that our method achieves competitive performance across all five benchmarks, consistently outperforming recent Transformer and Mamba-based baselines on most metrics under the same training protocol, while remaining on par with the strongest SAM-based models, illustrating its effectiveness. The code is publicly available at https://github.com/libaitian02/FDANet .