Improved Efficientnet Transfer Learning Framework with Data Augmentation For Multiclass Skin Cancer Classification

Abstract

Early detection of skin cancer is vital in enhancing the success of its treatments. This paper introduces a multiclass skin cancer classification method based on EfficientNet employing the HAM10000 dermatoscopic images dataset. The class imbalance of the dataset is very extreme, such that thousands of samples of some types of lesions and less than one hundred samples of others exist. To alleviate this, targeted data augmentation—rotation, flipping, brightness/contrast, zooming, and affine transformation—was performed on minority classes, creating a balanced dataset of 1000 images per class. Transfer learning using EfficientNet was utilized to take advantage of its depth, width, and resolution optimized scaling for efficient feature extraction. The model proposed performed at 99.51% training accuracy, 0.9947 F1-score, and 0.9997 AUC, and at 83.82% validation accuracy, 0.8420 validation F1-score, and 0.9564 validation AUC. Test outputs gave a macro F1-score of 0.8030, with very high recall across various categories of lesions. The findings show that the integration of dataset balancing with EfficientNet-transfer learning prominently enhances classification and minimizes bias against majority classes. This method holds great promise for computer-based skin cancer diagnosis in clinical use.