An Optimized Balanced-Learning Framework for Malignant Skin Lesion Triage Using Compound-Scaled Neural Networks

Abstract

Skin cancer represents a prevalent global health challenge, and early detection is very important to reduce mortality risk. Manual dermoscopic diagnosis risks human bias, making deep learning classification a vital research topic. While several previous studies utilizing the ISIC 2019 dataset have demonstrated high diagnostic capabilities, they primarily focus on complex multi-class classification. However, in real-world clinical workflows, the primary necessity is a swift, dependable triage system that can confidently distinguish dangerous lesions from non-threatening ones. Furthermore, many existing models require substantial computational overhead yet still suffer from imbalanced accuracy when dealing with minority malignant classes. The novelty of this study lies in addressing these gaps by developing a streamlined, clinically practical binary screening framework optimized specifically for malignant-versus-benign triage. The original multi-class labels were transformed into binary classes where malignant lesions consist of melanoma (MEL), basal cell carcinoma (BCC), and squamous cell carcinoma (SCC), while benign lesions consist of nevus (NV), benign keratosis (BKL), dermatofibroma (DF), and vascular lesions (VASC). The experiment applied transfer learning with ImageNet-pretrained weights, data augmentation, class weighting, and fourfold stratified cross-validation. Unlike prior works that rely on resource-heavy architectures, we leverage the compound-scaled EfficientNet-B4 backbone—delivering superior feature representational power with significantly fewer parameters evaluate on a large-scale cohort of 25,331 dermoscopic images. Experimental results show that the proposed model achieved an average accuracy of 89.77% and an average ROC AUC of 96.16%. The best fold obtained 91.49% accuracy with ROC AUC of 97.19%. Simultaneously, the framework maintained an average F1-score of 89.20%