AMN: An Adaptive Multi-Scale Fusion Network with Boundary and Uncertainty Modeling for Nuclei Segmentation

Researchers introduce AMN, an advanced nuclei segmentation network combining Swin Transformer and ResNet-50 encoders for improved histopathology image analysis. The model achieves state-of-the-art performance on the CoNIC benchmark, outperforming eight existing architectures while demonstrating strong cross-dataset generalization capabilities.

AMN addresses a fundamental challenge in computational pathology: accurately identifying different types of cell nuclei in tissue images. This capability underpins critical clinical workflows including cancer grading, immune response assessment, and patient outcome prediction. The research demonstrates that hybrid architectures combining convolutional and transformer-based approaches can outperform purely specialized designs, suggesting a broader trend toward ensemble methodologies in medical imaging.

The technical innovation centers on adaptive fusion rather than simple concatenation of encoder outputs. By implementing per-channel gating that dynamically weights each encoder's contribution across different scales, AMN optimizes the balance between local texture extraction (CNN strength) and global context understanding (transformer strength). The multi-objective loss function incorporating uncertainty modeling represents another advancement, specifically targeting the problem of overconfident incorrect predictions that plague diagnostic systems.

For the medical AI industry, these results validate that hybrid architectures merit continued investment in production systems. The mean Dice coefficient of 0.82 and particularly the 0.67 F1 score on lymphocytes—historically difficult to classify—suggests practical clinical viability. Cross-dataset evaluation on MoNuSeg without retraining indicates the learned representations generalize across different data distributions and imaging protocols, a critical requirement for real-world deployment.

Future development should focus on extending this architecture to three-dimensional volumetric analysis, reducing computational requirements for integration into clinical workflows, and validating performance on diverse tissue types beyond current benchmark limitations. The uncertainty quantification framework also opens opportunities for active learning systems that identify genuinely ambiguous cases for pathologist review.

• →Hybrid dual-encoder architecture combining Swin Transformer and ResNet-50 achieves superior performance over pure-CNN and pure-transformer baselines
• →Uncertainty-modulated loss function effectively suppresses overconfident predictions in challenging lymphocyte classification tasks
• →Model demonstrates strong cross-dataset generalization on MoNuSeg without retraining, indicating domain robustness for clinical deployment
• →Adaptive per-channel gating mechanism dynamically balances local texture and global context extraction across multiple scales
• →Mean Dice of 0.82 and F1 of 0.68 on CoNIC benchmark represents significant advancement in computational pathology