Dermatology Image Quality Assessment (DIQA): Artificial Intelligence to ensure the clinical utility of images for remote consultations and clinical trials

Authors

Ignacio Hernández Montilla, Taig Mac Carthy, Andy Aguilar, Alfonso Medela

What the study evaluated

The authors assembled 934 clinical and dermoscopic images captured with smartphones, digital cameras, and dermatoscopes. Forty non-expert observers rated each image from 1 to 10 using a protocol that considered lighting, focus, and distance. A neural network was then trained to predict the resulting mean opinion score.

On the dermatology validation set, the model trained with both general-domain and dermatology images produced a mean absolute error of 0.472, a linear correlation of 0.737, and a Spearman correlation of 0.734.

Limits

The ratings measured perceived image quality, not whether an image was clinically sufficient for a particular diagnosis or severity assessment. The observers were not dermatologists, and the paper states that clinical-meaning evaluation by dermatologists was future work. DIQA should therefore be read as an early image-quality study, not as evidence that the model guarantees clinical utility or patient safety.

Read full text

Available at https://doi.org/10.1016/j.jaad.2022.11.002.