MelaResNet-34: An enhanced transfer learning model for automated and accurate melanoma diagnosis
Liang Wang, Chuantao Cheng, Duwei Dai, Xiaoyi Wang, Zhangjian ZhouBackground
Melanoma is an aggressive skin cancer requiring early diagnosis. However, manual diagnostic procedures utilizing dermoscopes are heavily dependent on clinical expertise and are inherently susceptible to high inter-observer subjectivity. Therefore, developing efficient and objective automated auxiliary diagnostic models remains of great clinical value.
Methods
We proposed MelaResNet-34, a transfer learning-based model for automated melanoma detection. Using the ISIC 2019 dataset, we applied data cleaning and moderate augmentation. The ResNet-34 backbone was optimized via a tailored fine-tuning strategy. Following a strict 7.5:2.5 internal split, performance was evaluated using comprehensive metrics (Accuracy, AUC, F1-score, MCC). Crucially, external validation was conducted on the independent PH2 dataset.
Results
Internally, MelaResNet-34 achieved 97.76% accuracy, 0.987 AUC, and an MCC of 0.955, demonstrating highly competitive performance against both classic baselines and contemporary state-of-the-art architectures. Ablation experiments verified the essential roles of transfer learning and moderate augmentation. Furthermore, external validation on the PH2 dataset confirmed robust generalizability (91.0% accuracy, 0.938 AUC) despite inter-institutional domain shifts.
Conclusion
MelaResNet-34 exhibits superior classification and generalization capabilities, highlighting its potential to serve as an efficient, objective diagnostic tool for clinicians. While current results are promising, future large-scale, multi-center prospective validation is required before routine clinical implementation.