Skin Cancer Detection and Classification Through Medical Image Analysis Using EfficientNet

Skin cancer is one of the most frequently occurring and life-threatening forms of cancer globally, highlighting the importance of timely and precise diagnosis to enhance treatment success.Automated skin lesion classification has greatly benefited from deep learning methods, especially convolutional neural networks. This study utilizes the EfficientNet-B0 architecture, a lightweight yet robust CNN, to develop a reliable multi-class skin cancer classifier using the HAM10000 dermoscopic image dataset. To ensure compatibility with the pre-trained EfficientNet-B0 model, images were uniformly scaled to 224×224 pixels and normalized according to ImageNet data to achieve consistent dimensions and brightness levels. Addressing class imbalance, minority classes such as actinic keratoses, basal cell carcinoma, dermatofibroma, and vascular lesions were increased to 1,000 images each through augmentation, whereas the majority class, melanocytic nevi (nv), was reduced to 1,300 images.This resulted in a balanced dataset comprising 7512 images distributed evenly across seven classes. Initially, transfer learning was applied by freezing the base layers and fine-tuning the final layer, achieving 77.39% accuracy. Further full-network fine-tuning improved accuracy to 89.36%. Test-time augmentation with flips increased performance to 90.16%, and integrating TTA with Monte Carlo Dropout and additional augmentations boosted the final accuracy to 92.29%. The results highlight the potential of EfficientNet-B0. This research improved classification model for skin lesion detection can aid healthcare professionals in early diagnosis, ultimately enhancing patient care and reducing the burden on healthcare systems.