Class-Balanced Dermoscopic Lesion Segmentation

This package provides the complete implementation of a class-balanced dermoscopic skin lesion segmentation framework, integrating:

Advanced Preprocessing Pipeline for illumination correction, lesion morphology enhancement, and artifact-free normalization;

MoG-LISA (Morphology-Guided Latent Interpolation and Synthesis for Lesion Augmentation) for generating realistic, clinically valid synthetic melanoma samples, addressing class imbalance at the dataset level; and

CB-SwinGMO (Class-Balanced Swin-UNet Optimization Using GM-FDEF), a segmentation architecture enhanced with windowed self-attention and trained using a Geometric Mean-Based Feedback-Driven Evolutionary Optimization Strategy, ensuring robust convergence and superior segmentation accuracy.

The pipeline is engineered to function end-to-end, from raw dermoscopic image preprocessing, through morphology-aware latent space augmentation, to final lesion segmentation with Dice, IoU, and boundary-based overlay outputs. The model is designed to preserve lesion structure, color patterns, border irregularity, and clinical interpretability, making it suitable for both research and clinical decision-support workflows.