Comprehensive Summary
This study investigates whether machine learning is able to automatically triage potentially malignant skin lesions using 3D total-body photography rather than clinician-selected images. The researchers analyzed results from the ISIC 2024 international machine learning competition, which used over 900,000 lesion tiles extracted from 3D total-body photographs of nearly 1000 patients across multiple institutions. The models were evaluated using metrics focusing on sensitivity to reflect real world scenarios. The winning ISIC 2024 model demonstrated strong abilities to identify skin cancer, achieving high AUC values, while also significantly reducing the number of lesions that needed expert evaluation. Models that were able to also incorporate patient context, such as lesion size and deviation from baseline, had better performance compared to models who only considered images. The findings support the concept that AI-assisted triage using 3D total-body photography is possible in the future.
Outcomes and Implications
Screening skin cancer is limited by time and access to dermatologist care, making identifying the lesions in the proper time frame incredibly important. Automated triage could help clinicians focus attention on the most concerning lesions without an overwhelming increase in workload. These models have the ability to support dermatologists and primary care providers in evaluating more patients, especially in high risk populations. However, this model is not ready for full implementation into clinical practice, but lays the groundwork for AI-assisted screening tools that could identify signs earlier and improve access in dermatology care.