Efficient nuclei semantic segmentation in histopathology images: A weakly supervised approach with color and sure-foreground extraction

Nuclei segmentation, a critical task in cancer diagnosis, presents substantial challenges in histopathological image analysis due to the labor-intensive and time-consuming nature of manual pixel-wise annotation required for fully supervised training. In this study, we address these challenges by proposing a novel weakly supervised segmentation framework. Notably, we introduce two novel methods: color region extraction and sure-foreground extraction. These methods address the limitations of existing approaches by generating high- quality weak masks, effectively mitigating the impact of unwanted image imperfections and irrelevant tissue elements. Furthermore, we enhance the LinkNet architecture by incorporating Dual-branch Spatial Blocks (DBS Blocks) within the decoder section. This novel component empowers the network to capture fine-grained details crucial for accurate segmentation. Finally, we introduce a boundary-aware loss function to refine the model during training, specifically focusing on enhancing nuclei boundary accuracy. Our framework has achieved competitive performance compared to state-of-the-art methods on two benchmark datasets. Notably, our approach effectively addresses unwanted image imperfections and eliminates irrelevant tissue elements, leading to precise segmentation results compared to existing weakly supervised methods.