Enhanced Multi-Class Skin Cancer Detection Using EfficientNet-B0 with Test-Time Augmentation and Monte Carlo Dropout

Abstract

Skin cancer is one of the most frequent and potentially fatal malignancies worldwide, emphasizing the significance of early and correct detection. Convolutional neural networks (CNNs), a recent deep learning advancement, have shown promise in automating the classification of skin lesions. This study employed EfficientNet-B0, a lightweight yet very effective CNN architecture, to train a reliable multi-class classification model on the HAM10000 dermoscopic picture dataset. To ensure compatibility with the pre-trained network, all input images were resized to 224 x 224 pixels and normalized using the ImageNet mean and standard deviation values. To rectify the class imbalance, the majority class (melanocytic nevi) was lowered to 1300 samples, while underrepresented categories (actinic keratoses, basal cell carcinoma, dermatofibroma, and vascular lesions) were oversampled with 1000 samples each. This preprocessing resulted in a balanced collection of 7512 pictures organized into seven diagnostic groups. Transfer learning was originally utilized to achieve a 77.39% accuracy by freezing the convolutional basis and training only the final classification layer. After fine-tuning the entire network, the accuracy improved to 89.36%. Test-time augmentation with flips enhanced performance to 90.16%, while combining TTA with Monte Carlo Dropout and additional augmentations increased final accuracy to 92.29%. The results highlight EfficientNet-B0's potential. By assisting medical professionals with early diagnosis, this study's improved classification model for skin lesion detection can improve patient care and reduce strain on healthcare systems.

Keywords: skin cancer; EfficientNet-B0; deep learning; transfer learning; dermoscopic images; test-time augmentation; monte carlo dropout; image processing