An Effective Hybrid EfficientNet-B3 Deep Learning Approach for Multi-Class Skin Disease Classification

Access to proper medical care is one of the most fundamental human needs. Since the cancer diseases are spreading day by day and among them Skin cancer and other dermatological diseases are highly prevalent worldwide, making precise diagnosis is critically important. However sometimes a traditional model may not accurately detect skin cancer at early. Therefore, this research developed and proposed a hybrid Deep Learning method based on EfficientNet-B3 architecture to automate the accurate detection of skin cancer on a new dataset. The proposed method can effectively extract fine grained dermatological features through compound scaling and depth-wise separable convolutions. These extracted features make it suitable for the task of skin cancer classification. To assess the performance of the proposed method a realistic data set comprising 48,229 dermoscopic images across 15 classes of cancerous, precancerous, and non-cancerous lesions are used for experiments. To obtain satisfactory performance, the proposed method proceeds with some preprocessing including hair removal via black-hat filtering and dataset balancing. Besides these, oversampling and data augmentation were performed to enhance image quality and model generalization. Therefore, the proposed method achieved an impressive 94% classification accuracy, with precision, recall, and F1-scores all averaging 0.94 across classes. Comparative analysis demonstrates that the proposed model outperforms the existing methods despite handling a more diverse set of skin conditions. Therefore the proposed method recommends the early detection of skin cancer with it to save many precious lives.