Artificial Intelligence‐Based Quantification to Assess the Automatic Psoriasis Area and Severity Index

ABSTRACT Background Psoriasis is a chronic inflammatory skin condition affecting millions globally, traditionally assessed via the Psoriasis Area and Severity Index (PASI). Despite its widespread use, PASI suffers from subjectivity, limited sensitivity for small lesions, and high interobserver variability. Recent advances in artificial intelligence offer a promising route to overcome these limitations. Objectives This study aimed to develop and validate the Automatic Psoriasis Area and Severity Index (APASI), an AI‐driven system for rapid, objective, and standardised assessment of psoriasis severity. Methods We compiled the Legit.Health‐PsO‐PASI dataset, consisting of 2857 images spanning various disease severities and annotated by three expert dermatologists. Two deep learning modules were developed: one for lesion segmentation and another for visual sign intensity classification (erythema, induration and desquamation). Multiple encoder architectures, including ResNet, SE_ResNeXt, Xception, Inceptionv4, EfficientNet, and SegFormer, were evaluated using a five‐fold cross‐validation approach. Performance metrics included accuracy, root mean square error, intersection over union, confusion matrices and Cohen's kappa for assessing interobserver variability. Results The best‐performing model, MiT_b2, achieved human‐comparable performance, with accuracies of 60.6% for erythema, 54.3% for induration, and 61.8% for desquamation. For lesion segmentation, the Xception model outperformed expert dermatologists, achieving an intersection over union of 0.752. Conclusions APASI provides a robust AI‐driven framework for psoriasis severity assessment, delivering rapid, objective, and precise evaluations. Its integration into clinical and research workflows could enhance disease monitoring, improve treatment assessment and reduce evaluation costs.