M3-Net: A Multi-Scale Nuclei Segmentation Model for Breast Cancer Histopathology Using Contextual Patches and Attention Mechanism

Histopathological analysis, particularly of nuclear morphology, is critical for identifying malignancies. Accurate nuclei segmentation plays a pivotal role in this process, as it enables detailed assessment of nuclear size, shape, and distribution patterns. Traditional segmentation methods, however, often fail to capture fine details, lack broader context, and struggle with overlapping nuclei of varying sizes and shapes, especially when relying on single-scale approaches. To address these challenges, we propose a multi-scale context-based encoder-decoder model named M3-Net (Multi-Scale, Multi-Level, Multi-Stream Network) that integrates both global and local tissue features. Evaluations demonstrate that M3-Net effectively segments overlapping nuclei and diverse structures, providing a robust solution for automated nuclei segmentation in breast cancer pathology.