DeepSkinNet: An EfficientNetV2B0-Based Framework for Binary Skin Lesion Classification

Abstract

The classification of skin lesions with deep learning is still difficult when there is a limited amount of labeled dermoscopic data. In this work, the authors propose an EfficientNetV2B0 framework for binary classification of actinic keratosis (AKIEC) and vascular lesions (VASC), which is referred to as DeepSkinNet. Images that were verified for this study were 270 images, which included 128 AKIEC images and 142 VASC images from the HAM10000 dataset. To avoid information leakage from multiple images in one dataset belonging to the same lesion, the datasets were split on a lesion-by-lesion basis using the HAM10000 lesion identifiers with the aim that no specific lesion would be present across training, validation, and test sets. The framework features two-stage selective fine-tuning, online data augmentation, class weighting, Test-Time Augmentation (TTA), and a decision threshold optimized for the validation set, as well as transfer learning with ImageNet. The optimized threshold (0.59) is solely based on the validation set and further applied to the independent test set. DeepSkinNet's performance on the lesion independent test set (44 images) was 93.18% accuracy, 92.86% balanced accuracy, 93.88% F1-score and 99.79% ROC-AUC. The findings in this study show that discrimination between AKIEC and VASC was promising under the protocol for the evaluation of leakage, and further validation on larger and external datasets is needed.