Digital Twin-Inspired Deep Learning for Skin Cancer Detection and Longitudinal Risk Monitoring

Abstract

Skin cancer, especially melanoma, remains one of the most virulent cancers worldwide, with incidence rising sharply and placing increasing burden on healthcare systems. Traditional diagnosis relies on subjective visual inspection and dermoscopy without longitudinal lesion tracking. We propose DT-EfficientNetB0, a CNN-integrated Digital Twin-Inspired framework combining a

two-phase fine-tuned EfficientNetB0 classifier with a clinicallyinformed risk simulation module. Trained on HAM10000 with class-weighted loss and clinical metadata (age, sex, localization, history), the model achieves 94.6% accuracy, weighted F1-score of 0.940, and macro ROC-AUC of 0.989, outperforming CNN and ResNet50 baselines with p < 0.001. The Digital Twin module contributes an additional 1.2% accuracy gain through clinical context integration, and generates simulated longitudinal risk trajectories pending real longitudinal validation. This work bridges high-performance image classification with interpretable, clinically-informed risk assessment for proactive clinical decision support.