A Wavelet-Enhanced Attention Network with Elliptical Directed Depthwise Convolutions for Nuclei Segmentation

Abstract

In digital pathology, accurate nuclei segmentation from histopathology images is crucial for detecting cancer in various organs. Overlapping nuclei, uneven cellular architec-ture, varying tissue arrangements, and staining variations pose challenges for pathologists conducting morphological analysis. By enabling highly accurate nucleus detection and enhancing diagnostic results, computer-aided approaches overcome these challenges. A wavelet-enhanced attention-aided (WEAM) deep learning model for segmenting nuclei from histopathology im-ages is presented in this paper. Anisotropic nuclear features are extracted by the Elliptical Directed Depthwise Convolution (EDDC) module, and model performance is enhanced by patch-based learning and thorough data augmentation. Improved performance over baseline models is demonstrated by evaluation on the MoNuSeg and PanNuke benchmark datasets, which yield Dice scores of 0.8100 and 0.8036 and Intersection over Union (IoU) scores of 0.6816 and 0.6717, respectively.