Design of a feature-optimized hybrid CNN–Transformer model for melanoma-vs.-non-melanoma classification

Introduction Accurate melanoma recognition from dermoscopic images is challenging because lesion appearance varies substantially across acquisition devices, skin backgrounds, and diagnostic categories. Methods This study presents Lesion-aware Hybrid CNN-Transformer Network (LHCT-Net), a lesion-aware hybrid CNN-Transformer framework for binary melanoma-vs.-non-melanoma classification on the 25,331-image ISIC 2019 training collection. The endpoint is defined explicitly as melanoma (positive class) vs. all remaining ISIC 2019 diagnostic categories (negative class); consequently, the reported results are not interpreted as benign-vs.-malignant or general skin-cancer detection performance. The pipeline combines fixed U-Net-based lesion localization, training-only mask-guided lesion augmentation, a boundary-aware EfficientNet-B0 branch for local morphology, a Swin Transformer branch for global context, attention-based local-global fusion, a Lightweight Feature Optimization Block (LFOB), and qualitative Grad-CAM++ visualization. The dataset was evaluated using a fixed 70:15:15 stratified image-level holdout split, with augmentation and hyperparameter selection restricted to the training and validation partitions and the test partition used only for final evaluation. Results On the locked test set, the confusion matrix contained 3,096 true negatives, 26 false positives, 29 false negatives, and 650 true positives, corresponding to 98.55% accuracy, 96.15% melanoma precision, 95.73% melanoma sensitivity, 99.17% specificity, 95.94% F1-score, 97.45% balanced accuracy, and an MCC of 0.951. The reported ROC-AUC was 0.98. Discussion These results indicate strong discrimination for the defined melanoma-vs.-non-melanoma task, while external validation, subgroup analysis, calibration, and expert validation of explanations remain necessary before clinical use.