An EfficientNetB4-Based CNN Model for Skin Lesion Classification

Abstract

Skin cancer, with over three million new cases per year worldwide, is a significant public health issue, with the most lethal form being melanoma. Early detection of skin cancer is crucial for higher survival rates, but the visual similarity of lesions and class imbalance in image datasets such as HAM10000 complicate diagnosis. This study presents a cutting-edge deep learning model based on EfficientNetB4, reinforced by a Soft Attention block to improve the classification of skin lesions. Using the HAM10000 dataset, class imbalance is addressed through data augmentation, such as random rotation, flip, and MixUp, to obtain an equal representation of the diagnostic classes. The proposed architecture yields enhanced performance with a global accuracy of 93.09%, macro F1-score of 0.8352, and ROC-AUC of 0.9901. Particularly noteworthy, precision in melanoma was as high as 0.5874 with a recall of 0.7706, demonstrating strong identification of high-impact cases despite underrepresentation in areas. Discriminability is amplified by the Soft Attention module, indicating diagnostically important image regions and reducing misclassification errors. Compared to state-of-the-art models, the proposed approach has higher stability and generalizability, particularly on imbalanced datasets, offering a rich resource for dermatological clinical practice. This work contributes to AI-based medical imaging by offering an interpretable and rational paradigm for early skin cancer detection that holds promise for supporting high-risk area dermatologists.