A Lightweight MobileNetV2 Framework for Four-Class Dermoscopic Skin Lesion Classification using Test-Time Uncertainty Estimation

Automated dermoscopic image classification has potential for skin lesion analysis, but practical use is limited by class imbalance, differences in image quality, model size, limited interpretability, and uncertainty in predictions. This study presents SkinSense AI, a lightweight four-class classifier built on MobileNetV2 and evaluated on a combination of the HAM10000 and ISIC 2019 datasets. The two archives provided 17,132 readable images, of which 1,874 were duplicates; the remaining 15,258 unique images were used. The four classes were benign nevus (NV), melanoma (MEL), basal cell carcinoma (BCC), and squamous cell carcinoma (SCC); the HAM10000 AKIEC class was mapped to SCC. A stratified image-level split gave 10,680 training, 2,289 validation, and 2,289 test images. The model used a MobileNetV2 backbone with a custom classification head, two-stage transfer learning, and training-only augmentation. On the test set it reached 90.74% accuracy, a macro-average F1-score of 0.8705, and a macro-average ROC-AUC of 0.9831. SCC had the lowest F1-score (0.7468), which shows that overall metrics can hide weaker performance on minority classes. A test-time augmentation (TTA) uncertainty score, with a threshold of 0.110 selected on the validation set, flagged 17.78% of test images for review and separated correct from incorrect predictions with an AUC of 0.8636; the TTA-averaged predictions reached 92.00% accuracy. The model has 2,552,836 parameters, occupies 9.74 MB, and ran at 6.90 ms per image on a GPU. These results come from an internal retrospective evaluation on public data; the model is not ready for clinical use without external validation.