Implementation of Efficientnet and Grad-Cam for Detection and Interpretation Of Skin Cancer Melanoma in Dermoscopic Images

This study aims to implement the EfficientNet architecture for classifying melanoma skin cancer from dermoscopic images. Melanoma, one of the deadliest forms of skin cancer, is often difficult to detect visually, making deep learning technology a promising tool for facilitating faster and more accurate diagnoses. The dataset used consists of dermoscopic images of skin cancer sourced from Kaggle, categorized into two classes: Malignant and Benign. The classification process was carried out using various EfficientNet models (B0, B3, B4, B5, and B7) with transfer learning techniques, where different optimizers, such as Adamax, RMSprop, and SGD, were tested to obtain optimal results. Experimental results showed that the EfficientNetB3 model with the Adamax optimizer achieved the highest accuracy of 95%, making it the best performer in this testing scenario. To enhance the interpretability of the model's predictions, Gradient-weighted Class Activation Mapping (Grad-CAM) was utilized to visualize the areas of the dermoscopic images that contributed most significantly to the classification decision. The resulting model was then integrated into a Flask-based web system to provide real-time predictions and visual explanations from user-uploaded dermoscopic images. This research demonstrates the potential of EfficientNet in improving the accuracy and efficiency of early melanoma detection, while also contributing to the development of technology-based diagnostic tools for medical professionals and the public.