A new melanoma-detection framework doesn't just look at one mole - it looks at all of them, plus where they sit and who they belong to.
Researchers built a modular system, called MelDD, that diagnoses potential melanoma using three tiers of evidence: the lesion itself, the patient's full set of moles, and population-level demographic data. Instead of judging a single skin lesion in isolation, it uses what the team calls an anatomic-site aware masked transformer to weigh every lesion on a patient - however many there are - along with where each one sits on the body. It also folds in patient metadata, like age and sex, through learnable demographic embeddings. Tested on the SIIM-ISIC 2020 dataset, adding lesion location data raised specificity by 17.15%, and adding demographic metadata added another 7.14%, while also improving overall balanced accuracy.
Most skin cancer AI tools get graded on sensitivity - catching every possible melanoma - which tends to flag plenty of harmless moles along with it. This work is notable for targeting specificity instead, the metric that decides how many patients get sent for an unnecessary biopsy. Modeling a patient's whole mole map, rather than a cropped photo of one spot, is also a closer match to how dermatologists actually work: they compare lesions against each other, not in a vacuum.
The code is public on GitHub, but public code isn't the same as clinical validation - this is one dataset, not a green light to replace a dermatologist's judgment.