A Two-Stage Transfer Learning Framework for Dermoscopic Classification of Facial Pigmented Lesions

Facial pigmented lesions represent a challenging diagnostic setting because benign, premalignant, and malignant entities may show overlapping dermoscopic features, particularly on chronically sun-damaged skin. This study developed and evaluated an image-based deep learning model for the multiclass classification of facial pigmented lesions. A two-step transfer learning strategy was used: a ResNet-152 model was first pretrained on a merged ISIC-derived dataset including 92,044 dermoscopic images across nine diagnostic categories, and subsequently fine-tuned on a facial pigmented lesions dataset comprising 1112 images from seven diagnostic classes. Class imbalance was addressed using weighted sampling and class-weighted focal loss. On the internal test set, the model achieved a macro-AUC of 0.860, averaged recall of 0.697, macro-F1 score of 0.622, accuracy of 0.675. External testing showed preserved discriminative ability, with a macro-AUC of 0.85. These findings suggest that deep learning models based on contrastive learning training technique may support the differen- tial diagnosis of facial pigmented lesions, although further multicenter validation and calibration refinement are required before clinical implementation.