ResNet50 two-stage fine-tuning boosts melanoma detection to 88% accuracy

A new research paper from Aryan Bhagat introduces a two-stage fine-tuning strategy for ResNet50 aimed at high-sensitivity melanoma detection from dermoscopic images. The core challenge addressed is class imbalance and suboptimal transfer learning common in single-stage fine-tuning. After stratified splitting and applying random oversampling to the training set to achieve a 1:1 class balance, Stage 1 trains only the classification head with the ResNet50 base frozen. Stage 2 then fine-tunes all layers jointly at a low learning rate of 1e-5 to prevent catastrophic forgetting of learned visual features. On an independent test set of 3,826 images, the model achieves an AUC-ROC of 0.9559, accuracy of 88.34%, sensitivity of 87.56%, specificity of 89.13%, and F1-score of 88.29%. An ablation study confirms the two-stage protocol significantly outperforms single-stage fine-tuning, with sensitivity gains of over 4%. Grad-CAM visualizations further demonstrate correct lesion localization, adding interpretability.

This work is notable not only for its performance metrics but also for its practical deployability. The author provides a fully functional Streamlit detection application alongside all training code, making it easy for dermatologists and researchers to test the model on new images without deep learning expertise. The sensitivity of 87.56% is particularly important because early detection of melanoma dramatically improves survival rates—up to 99% five-year survival when caught early. While the model is not yet clinically certified, it represents a significant step toward AI-assisted skin cancer screening. The two-stage fine-tuning methodology could also be adapted to other medical imaging tasks suffering from class imbalance, offering a reproducible template for transfer learning in healthcare AI.