MSWAFFNet: improved segmentation of nucleus using feature fusion of multi scale wavelet attention

Introduction Nucleus segmentation plays an essential role in digital pathology,particularly in cancer diagnosis and the evaluation of treatment efficacy. Accurate nucleus segmentation provides critical guidance for pathologists. However, due to the wide variability instructure, color, and morphology of nuclei in histopathological images, automated segmentation remains highly challenging. Previous neural networks employing wavelet-guided, boundary-aware attention mechanisms have demonstrated certain advantages in delineating nuclear boundaries. However, their feature fusion strategies have been suboptimal, limiting overall segmentation accuracy. Methods In this study, we propose a novel architecture—the Multi-Scale Wavelet Fusion Attention Network (MSWAFFNet)—which incorporates an Attention Feature Fusion (AFF) mechanism to effectively integrate high-frequency features extracted via 2D Discrete Wavelet Transform (DWT) from different Unet scales. This approach enhances boundary perception and improves segmentation performance. To address the variation across datasets, we apply a series of preprocessing steps to normalize the color distribution and statistical characteristics, thereby ensuring training consistency. Results and Discussion The proposed method is evaluated on three public histopathology datasets (DSB, TNBC, CoNIC), achieving Dice coefficients of 91.33%, 80.56%, and 91.03%, respectively—demonstrating superior segmentation performance across diverse scenarios.