AI Diagnosing Skin Diseases Narrows Skin-Tone Gap
AI TIMES ·
✦ AI Summary
According to AI TIMES, a research team at Asan Medical Center in Seoul developed a generative AI that changes only skin tone while preservin…
According to AI TIMES, a research team at Asan Medical Center in Seoul developed a generative AI that changes only skin tone while preserving the shape of photos of lesions on light skin, narrowing the performance gap in skin disease diagnostic AI that had widened by skin tone. In the study, introduced on October 8, a model trained with additional data on darker skin tones achieved an AUC of 0.66, improving on the existing model. The generated images were difficult to distinguish from real ones, showing potential as a way to supplement scarce clinical data on darker skin. The team said it focused on preserving shape information, which is important for diagnosis, by changing only skin tone based on actual patient photos instead of creating lesions from scratch as conventional generative AI does. As a result, it said, the model improved performance while also maintaining diagnostic performance on light skin. The study is meaningful in that it presented a practical way to address concerns that medical AI could work to the disadvantage of specific groups.
Perspective
The key point in this case is that competition in medical AI performance is now shifting from average accuracy alone to whether it works similarly for everyone. In particular, the idea that technology can fill data gaps in groups where real-world data is scarce suggests room to improve both deployment speed and fairness. Going forward, as much as building better models, how well training data are acquired and supplemented evenly is likely to become an important benchmark of medical AI competitiveness.
This perspective is BizCrush's own commentary and is not part of the reporting by AI TIMES.
This article was produced with the help of an automated content generation algorithm.
Source: AI TIMES
View originalThis article was summarized and organized by BizCrush based on the original article from AI TIMES. For exact quotations and full details, please refer to the original article.