Performance of transformer-convolutional neural network ensemble for melanoma diagnosis on segmented 3D total body photography data: Cross-Validation stratified K-fold

Skin cancer is one of the most common cancers worldwide, and its early diagnosis is critical for the survival of cancer patients. Many artificial intelligence models have been developed to support early detection; however, limited progress has been made towards a reliable pre-diagnosis of skin cancer types. We developed a deep learning ensemble integrating ConvNeXt-Base, ResNet-50, and Swin Transformer-Base architectures to detect melanoma using images from 3D total body photography (3D-TBP) data. By combining the complementary strengths of individual models coupled with the Quadruple Stratified Leak-Free 5-Fold Cross-Validation (QSLF-KF-CV) approach, the ensemble achieved an Area Under the Curve (AUC) of 0.9208, which is greater than that of individual models. Our strategy demonstrates potential for accurate melanoma detection, as well as broader applications in medical image diagnostics. • An ensemble of deep-learning methods can improve diagnosis of skin cancer type from total body data images. • Use of ensemble of ConvNeXt, ResNet and Swin Transformer achieves Area Under the Curve (AUC) to 0.9208. • Individual models AUC for ConvNeXt, ResNet and Swin Transformer remains was 0.8722, 0.8551 and 0.8763, which is less than ensemble model. • The QSLF-KF-CV approach ensured that the evaluation of the model was fair, generalizable, and statistically reliable.