An open dataset and lightweight model for dermatological image quality assessment

Dermatological image quality is important for the reliability of teledermatology and AI-assisted diagnosis, yet it remains an underexplored problem with few publicly available resources. We introduce PAD-DIQA, an open dataset of 2,500 dermatological images collected in real-world clinical settings and annotated with continuous Mean Opinion Scores (MOS) across six perceptual quality criteria. To the best of our knowledge, PAD-DIQA is the first publicly available dataset specifically designed for dermatological image quality assessment following standard IQA annotation protocols. We further evaluate CNN and Transformer-based regression models that predict perceptual image quality and are suitable for on-device deployment. A threshold-based gating mechanism, calibrated at TPR@95%, enables flexible control of the trade-off between false acceptances and false rejections without retraining. As an additional finding, the proposed framework naturally handles out-of-distribution image rejection achieving performance competitive with dedicated OOD detection methods. On PAD-DIQA, all evaluated models achieve PLCC and SROCC above 0.83, while OOD rejection rates exceed 95%. PAD-DIQA is publicly released to support reproducibility and future research in this domain.