Optimizing Skin Cancer Classification With ResNet-18: A Scalable Approach With 3D Total Body Photography (3D-TBP)

Abstract

Skin cancer, particularly melanoma, remains a major public health challenge because of its rising incidence and mortality rates. Traditional methods of diagnosis, like dermoscopy and biopsies, are invasive, time-consuming, and highly dependent on clinical experience. Furthermore, previous research has predominantly focused on 2D dermoscopic images, which do not capture important volumetric information required for the proper evaluation of the injury. This work introduces a new deep learning architecture based on the ResNet-18 model, augmented by transfer learning, for binary classification of malignant and benign skin lesions. The model is trained on the ISIC 2024 3D Total Body Photography dataset and uses pre-trained ImageNet weights to enable effective feature extraction. To counter the dataset's natural class imbalance and minimize overfitting, the model uses sophisticated data augmentation and oversampling methods. The suggested model boasts a staggering classification accuracy of 99.82%, surpassing many other 2D-based alternatives. The utilization of 3D-TBP offers a strong diagnostic benefit by allowing volumetric lesion analysis, retaining spatial and depth features usually lost in the conventional 2D images. The findings validate the clinical feasibility of the method, presenting a scalable, noninvasive, and very accurate early detection and diagnosis of melanoma using 3D skin imaging.