Comprehensive Summary
This letter by Wen and Zeng reflects on a study by Kurtansky et al. that explored whether contextual skin images could enhance AI and human diagnosis of melanoma. Although the original study concluded that contextual information did not improve AI performance, Wen and Zeng demonstrated that this finding could be shaped more by methods rather than the irrelevance of patient context. In their letter, they assert two key main issues. First, the competition design did not prompt AI models to use patient-level contextual information, causing most to refer to single-image evaluation. Second, the dataset criteria did not include small benign lesions, which could have introduced selection bias and limited AI’s ability to learn intra-patient patterns.
Outcomes and Implications
This information suggests that contextual patient data may hold untapped value in dermatological AI, and current evaluation frameworks underestimate its importance. Further research should refine competition design by making contextual data use obligatory instead of optional, and also they should broaden datasets to include solitary lesions and a large amount of benign cases. This will in turn counteract bias and allow for more diversity in the study. If these recommendations are taken into account, AI tools for melanoma diagnosis could help real-world clinical reasoning, ultimately making models that are both more accurate and clinically useful.