Enhancing Nuclei Segmentation in Histopathology Images via Quantitative to Qualitative Transitions in Supervised and Weakly Supervised Deep-Learning Approaches

Segmentation of nuclei in histopathology images stained with Hematoxylin and Eosin is essential for accurate disease diagnosis and tissue characterization. Traditional methods usually struggle with complicated tissue structures and overlapping nuclei. Deep learning-based approaches, particularly U-Net along with its variants, have demonstrated improved effectiveness but still face limitations, including restricted receptive fields, overfitting, boundary ambiguity, and uneven nuclei distribution. This paper presents the deep image segmentation network (DISegNet) customized for nuclei segmentation. DISegNet integrates enhanced lightweight convolutional modules with expanded receptive fields and introduces an Improved Global Context (IGC) block to capture long-range dependencies. A dedicated stem block further boosts feature extraction. These components improve boundary precision and model robustness in visually challenging conditions. We examine DISegNet on the MoNuSAC dataset, its subsets, and the MoNuSeg dataset. Experimental results show that DISegNet consistently outperforms preexisting fully supervised segmentation methods. Further, in a bid to mitigate the costly dense annotations, we introduce a weakly supervised approach using refined-point annotations. Point annotations are transformed to coarse labels via k-means clustering and Voronoi diagrams. This strategy significantly reduces annotation effort while maintaining high segmentation performance, yielding more efficient and accurate results than previous point annotation approaches. Future work will explore our models for broader medical imaging tasks involving intricate anatomical structures and overlapping regions.