AMN network beats 8 baselines on nuclei segmentation with 0.82 Dice

A new research paper from Spoorthi M and Suja Palaniswamy presents AMN (Adaptive Multi-Scale Nuclei Network), a segmentation framework designed to improve nuclei subtype classification in histopathology images — a critical step for tumor grading, immune infiltrate quantification, and prognosis. Unlike existing approaches that rely solely on CNNs or transformers, AMN uses a dual-encoder architecture: a Swin Transformer captures long-range spatial context, while a ResNet-50 feature pyramid handles fine-grained local textures. The two encoders are fused via a learned per-channel gating mechanism that dynamically weights each encoder's contribution at every scale, allowing the network to adaptively prioritize the most relevant features per pixel. AMN is trained with a multi-objective loss that combines class-weighted focal loss (to handle class imbalance), boundary-aware loss with positive-pixel emphasis (to sharpen nuclear edges), and a novel uncertainty-modulated classification term that suppresses overconfident erroneous predictions. On the CoNIC benchmark (seven nuclei classes), AMN achieves a mean Dice of 0.82 and mean F1 of 0.68, with an F1 of 0.67 on the diagnostically challenging lymphocyte class. It outperforms eight baselines including U-Net, ResU-Net, DeepLabV3+, SegNet, ViT-Small, HmsU-Net, ConvFormer-UNet, and BEFUnet. Cross-dataset evaluation on MoNuSeg confirmed strong generalization without retraining, demonstrating domain robustness. The work highlights how careful architectural fusion and loss design can push the state of the art in computational pathology.