Abstract / Summary
Melanoma accounts for a small fraction of skin cancer cases but for the majority of skin-cancer-related deaths, making accurate early detection essential. Although deep learning-based dermoscopic image analysis has shown strong diagnostic potential, its performance depends strongly on training hyperparameters and loss-function weighting, which are commonly tuned manually or through computationally expensive grid search. To address this, Adaptive Combined Artificial Protozoa–Starfish Optimization (AC-APSO) is introduced as a bio-inspired hyperparameter optimization strategy that dynamically regulates the contributions of the Artificial Protozoa Optimizer (APO) and the Starfish Optimization Algorithm (SFOA) at each iteration according to population diversity, fitness improvement, and optimization progress. AC-APSO configures RM-LBAT-Net, a hybrid CNN–Transformer architecture that integrates Res2Net-50-based multi-scale feature extraction, MaxViT-Tiny-based global contextual learning, a Deformable Lesion Boundary Attention Module (DLBAM), and Cross-Scale Adaptive Semantic Fusion (CSASF). The optimized framework achieved an accuracy of 98.85%, with precision, recall, and F1-score of 98.70%, 99.00%, and 98.85%, respectively. Controlled ablation studies showed that each architectural component contributed under both fixed and optimized configurations and that adaptive coordination outperformed APO alone, SFOA alone, and a fixed-weight APO–SFOA combination. In a head-to-head comparison under an identical protocol, the framework outperformed contemporary architectures, including ConvNeXtV2-Tiny and Swin Transformer V2-Tiny. Five-fold cross-validation and external validation on the ISIC 2016 dataset (97.36% accuracy) further indicated stable performance and promising cross-dataset generalization.