Adaptive Dual-path Spatial-Frequency Network for medical microstructure segmentation

• Published in: Expert systems with applications Vol. 275; p. 127032
• Main Authors: Xie, Qihang, Hao, Kuangrong, Wei, Bing, Ding, He, Zhang, Yumei, Ren, Lihong, Yang, Fei
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 25.05.2025
• Subjects: Dual-path encoding; Few-shot datasets; Irregular global structures; Local morphological blurring; Microstructure medical segmentation; Multi-scale diagonal attention; Multi-scale diagonal attention; Microstructure medical segmentation; Few-shot datasets; Dual-path encoding; Local morphological blurring; Irregular global structures
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2025.127032

Accurate segmentation of microstructures is critical in the medical field. It directly affects the accuracy of downstream tasks. However, this task faces challenges such as local morphological blurring, irregular global structures, and limited dataset resources. To address these issues, we propose an Adaptive Dual-path Spatial-Frequency network (ADF_UNet). It incorporates a frequency-domain enhancement module, combining dual-path spatial features encoding to extract complementary features from both the foreground and edges. Then, an adaptive fusion module integrates the dual-path features, enabling the comprehensive capture of local morphological details. To further address irregular global structures, the model introduces a Multi-scale Diagonal Attention (MDA) module in the skip Connections, effectively capturing features at various scales. In addition, due to the design of directional complementary convolution kernels in the encoder, ADF_UNet performs excellently on few-shot datasets, reducing dependency on large sample datasets and maintaining high-precision segmentation performance even with limited data. Experimental results show that ADF_UNet performs excellently on multiple densely microstructured datasets, including Polyetherimide (PEI), MoNuSeg, CPM17, and DSB18, and maintains outstanding segmentation accuracy even on the resource-limited CPM15 dataset. Reduction experiments with data amounts of 100%, 75%, 50%, and 25% show that the model outperforms others as the data decreases, validating its effectiveness and robustness in medical image segmentation tasks. •ADF_UNet integrates a frequency-domain module and dual-path encoding.•Directional and standard convolution kernels work together.•A multi-scale diagonal attention module is designed in the model’s residuals.•Our model outperforms others on few-shot datasets, even with reduced data.•mIoU score of our framework is higher than other state-of-the-art models.