Automated Skin Cancer Classification from Dermoscopic Images Using Fine-Tuned CNN Models

Abstract

Skin cancer mostly melanoma cause a significant health risk as the dieses spread quickly and high risk of death if not detected early. The early detection of skin disease is essential but depending only on visual examination can be difficult for dermatologists and it may lead to delayed diagnoses. The deep learning based diagnostic systems can support clinicians by classifying skin lesions in a consistent manner and reducing variability in visual assessment. To automatically classify dermoscopic images we introduce a machine learning model that combines transfer learning with pre-trained CNNs. Three architectures VGG16, MobileNetV2 and EfficientNetB0 were evaluated on the MED-NODE dataset which contains images of melanoma and naevus that limited and imbalance. The methodology integrates the methods such as image pre-processing, extensive data augmentation, selective fine-tuning, hierarchical feature extraction and class weighting to improve model generalization and handle dataset limitations. The experimental results show that VGG16 achieved the highest performance with 82.5% accuracy, 84.67% weighted precision, 82.22% F1-score and 91.25% AUC. MobileNetV2 achieved moderate performance, whereas EfficientNetB0 performed poorly. This highlights the importance of deeper pretrained features and careful adaptation for small and imbalanced medical datasets. Analysis of per class performance, confusion matrices and training/validation curves shows that VGG16 extracts low-level features including edges, textures and also identifies higher-level patterns like asymmetry, borders and lesion shape which helps distinguish melanoma from naevus more reliably. The results show that transfer learning with deep CNN models improves automated skin lesion classification. The performance further improves with data pre-processing, data augmentation, class weighting and fine-tuning strategies. We proposed a machine learning framework for melanoma detection that helps clinicians to achieve more reliable diagnostic results in clinical practice. The framework can be evaluated on larger datasets using hybrid models to improve both performance and interpretability.