AHF-RBFNet: Attention-Guided Hierarchical Fusion U-Net with Learnable Fuzzy Membership Kernels for Medical Image Segmentation
Project Abstract & Overview
A primary challenge in automated skin lesion and boundary segmentation is the semantic uncertainty caused by blurry boundaries and low tissue contrast. To dynamically filter this uncertainty, we engineered AHF-RBFNet by integrating a learnable Radial Basis Function (RBF) fuzzy membership kernel directly into the skip-connections of an Attention-Guided Hierarchical Fusion U-Net. This integration enables the network to explicitly model fuzzy spatial boundaries. The architecture achieves state-of-the-art performance on the ISIC 2016 dataset, yielding 84.55% IoU and 91.97% Dice metrics, strictly outperforming 11 established segmentation baselines.
Key Methodologies & Contributions
- Learnable Fuzzy Kernels: Successfully combined fuzzy logic uncertainty filtering with deep representation learning, creating an explicit mathematical mechanism to handle spatial margin ambiguity in medical imaging.
- Massive Hyperparameter Optimization: Orchestrated large-scale hyperparameter searches utilizing Optuna to converge on robust kernel configurations.
- Rigorous Multi-Dataset Evaluation: Validated generalizability through extensive testing across multiple clinical datasets (ISIC, BUSI, PH2) supported by detailed 5-variant systematic ablation studies.