Glandular morphological features serve as a crucial basis for the accurate diagnosis of colorectal cancer. Existing gland segmentation methods primarily focus on separating adherent glands, overlooking the inter-class visual consistency that stems from the similarity between large-gland lumens and the surrounding background. In addition,because they operate solely in the spatial domain,these methods struggle to delineate gland morphologies with blurred edges in complex tissue contexts,leading to boundary loss.To address these challenges, we propose a method that integrates wavelet features with spatial-domain representations for gland segmentation. Specifically, we observed that the presence of large white lumens weakens the already sparse glandular boundaries,making it difficult to establish long-range dependencies.The absence of accurate high-frequency features hinders gland localization and results in boundary loss.In contrast,wavelets are sensitive to fine texture details and can provide continuous high-frequency information.To this end,we propose a wavelet-based,edge-and-texture-aware network for colorectal cancer gland segmentation (WET-Net),which comprises a wavelet texture enhancement framework (WTEF),a wavelet edge perception module (WEPM),and a spatial-domain edge network. The spatial edge network first extracts sparse glandular edge features to serve as a foundation for constructing continuous boundaries. Then,the WTEF progressively enhances high-frequency details across multiple scales using a multi-branch structure,enabling precise localization of gland borders. Finally,the WEPM integrates sparse spatial edges with the enhanced high-frequency texture to establish global relationships between edge pixels,yielding refined,smooth high-frequency contours. This approach enriches edge structural diversity while reducing interference from complex background information. The experimental results demonstrate that WET-Net achieves highly competitive performance on two widely used datasets and is particularly effective in segmenting large challenging glands.
Luo et al. (Mon,) studied this question.