Background: The application of Convoluted Neural Networks in AI algorithms for skin lesion classification have reported accuracy on par, and even outperformed expert dermatologists in experimental settings. While reported algorithm performance is promising, they do not represent a real-world clinical approach, and often don’t consider translation to clinical application. To date, majority of lesion classification algorithms have been trained on dermoscopy images of single lesions and fail to incorporate any clinical background information into the decision process. As the use of total body photography (TBP) in dermatology clinics increases, the exploration of applying AI technology to skin monitoring software offers opportunities in providing intra-patient assessment as opposed to single lesion analysis.
Objective: To review the current AI models for skin cancer analysis and describe opportunities and limitations in the application AI models for TBP.
Conclusions: Applying AI models for image analysis of TBP presents a clinically relevant opportunity for a more holistic patient risk assessment. Training datasets derived using standardised image acquisition process by TBP systems, and DICOM labelling of patient-level metadata would be advantageous for machine learning models. Additionally, TBP offers the opportunity to provide a skin phenotype assessment to evaluate risk factors of skin colour, UV damage and naevi distribution. Challenges exist in protecting patient privacy in algorithm development, and improving explainable AI (XAI) methods to increase both patient and clinician trust.