A Novel Transfer Learning Approach for Skin Cancer Classification on ISIC 2024 3D Total Body Photographs

Abstract

Skin cancer, and melanoma in particular, is a significant public health issue in the modern era because of the exponential death rate. Previous research has used 2D data to detect skin cancer, and the present methods, such as biopsies, are arduous. Therefore, we need new, more effective models and tools to tackle current problems quickly. The main objective of the work is to improve the 3D ResNet50 model for skin cancer classification by transfer learning. Trained on the ISIC 2024 3D Total Body Photographs (3D-TBP), a Kaggle competition dataset, the model aims to detect five significant types of skin cancer: Melanoma (Mel), Melanocytic nevus (Nev), Basal cell carcinoma (BCC), Actinic keratosis (AK), and Benign keratosis (BK). While fine-tuning achieves peak performance, data augmentation addresses the issue of overfitting. The proposed model outperforms state-of-the-art methods with an overall accuracy of 93.88%. Since the accuracy drops to 85.67% while utilizing 2D data, the substantial contribution becomes apparent when working with 3D data. The model articulates excellent memory and precision with remarkable accuracy. According to the findings, the 3D ResNet50 model improves the diagnostic process and may be rated better than conventional approaches as a noninvasive, accurate, and efficient substitute. The current model is valuable because it can help with a significant clinical application: the early diagnosis of melanoma.