A machine learning technique reaches 99% accuracy at spotting melanoma in photos

Researchers built a new method for detecting melanoma in dermoscopic images by converting skin lesion photos into mathematical graph structures, then training classifiers on those graphs. The approach achieved 99% accuracy on a standard test dataset, suggesting the technique could improve diagnostic screening tools.

Medical AI benchmarks are notoriously easy to game — high accuracy on a curated test set rarely translates to real-world performance on different patient populations or imaging equipment. This paper reports a single-dataset result with no deployment evidence, no cross-validation against different hospitals or patient groups, and no comparison to what dermatologists actually achieve in practice. The 99% figure is mathematically real but contextually hollow without evidence it works outside the lab.