EfficientNetV2-Based Deep Learning Classifier for Skin Cancer Lesion on Imbalanced Dataset

Skin cancer remains a prevalent global concern, with its incidence steadily increasing each year. Timely detection and accurate classification are pivotal in enhancing recovery prospects and minimizing mortality rates. However, discerning between skin cancer lesions and benign counterparts bearing similar appearances poses a significant challenge in lesion classification. This research endeavors to preprocess skin lesion data and construct an effective model for classifying skin cancer lesions utilizing the fine-tuned EfficientNetV2-B2 architecture. The study meticulously examines various hyperparameter configurations in the HAM10000 dataset, containing 10,015 dermatoscopic images of skin cancer lesions. Optimal settings include the Adamax optimizer, a learning rate of 0.001, a batch size of 64, 50 epochs, an image size of 260×260, and the fine-tuned EfficientNetV2-B2 model. During the preprocessing stage, image augmentation techniques such as vertical and horizontal flipping are applied to data training to bolster model performance. Resizing and normalization techniques are also implemented to prepare the dataset. Through this meticulously designed configuration, the model attains impressive metrics, including an accuracy of 89%, a precision of 81%, a recall of 80%, and an F1-score of 80%. These results underscore the efficacy of employing EfficientNetV2-B2 with tailored settings for enhancing skin cancer lesion classification on imbalanced datasets, thereby making significant strides in early skin cancer detection.