TBPGen: Total Body Photography Generator and AI-Driven Detection of Multiple Skin Lesions

Early detection of new or changing skin lesions in total body images is especially critical for improving skin cancer treatment outcomes. However, tracking skin lesions across the entire body can be challenging for physicians, considering patients with numerous skin lesions and the reliance on subjective human observation. The rising incidence of skin cancer and the limitations of traditional diagnostic methods highlight the need for advanced tools. Artificial intelligence can automate and enhance skin lesions’ detection, tracking, and analysis, but its effectiveness relies on the availability of diverse and realistic datasets. Due to the confidential nature of total body photographs, which are nearly nude images of patients, there is no available dataset for training deep learning models.This study introduces an innovative Total Body Photography Generator (TBPGen) developed to address the lack of diverse and realistic datasets essential for training deep learning models. TBPGen creates a comprehensive dataset by utilizing 3D human models in a virtual reality environment, incorporating skin lesions from the ISIC Archive. This dataset encompasses a wide range of skin tones, lesion types, body types, and postures, providing a robust foundation for research and development. Empirical evaluations indicate that deep learning models trained with the TBPGen-Generated dataset show promising capabilities in detecting and segmenting skin lesions. Thirteen different deep learning models are trained and tested on the TBPGen-Generated dataset. The FPN MB5 and U-Net DN201 models achieved the highest accuracy, both with F1 scores of 0.957 and respective IoU scores of 0.9401 and 0.9416 on the test data.