DERMAEDGE-X: AN EFFICIENCY-ORIENTED AND CLASS-IMBALANCE-AWARE DEEP LEARNING FRAMEWORK FOR MULTICLASS SKIN LESION CLASSIFICATION

Abstract

Automated dermoscopic image classification remains challenging because of severe class imbalance, inter-class visual similarity and the computational demands of contemporary deep learning models. This study presents DermaEdge-X, an efficiency-oriented framework for seven-class skin lesion classification using lesion-grouped data partitioning, transfer learning, class-weighted optimisation, label smoothing, image augmentation, validation-guided checkpoint selection and test-time augmentation. The framework was evaluated on a lesion-disjoint held-out test set using predictive, calibration, robustness and computational metrics. DermaEdge-X achieved 88.19% accuracy, 73.99% balanced accuracy, a macro F1-score of 76.98%, a macro receiver operating characteristic area under the curve of 0.9695 and a macro precision–recall area under the curve of 0.8610. The model contained 51.19 million parameters, required 51.04 GFLOPs per image and achieved a median inference latency of 23.85 ms per image on an NVIDIA Tesla T4 GPU. The proposed framework provides a reproducible basis for developing class-imbalance-aware and deployment-conscious skin lesion classification systems for computer-assisted dermatological screening.