DeepNAPSI multi-reader nail psoriasis prediction using deep learning.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Nail psoriasis affects approximately 50% of psoriasis patients, often leading to severe nail destruction and indicating a worse disease prognosis.
• Quantifying nail psoriasis severity is challenging due to heterogeneous nail matrix and bed involvement, limiting the clinical utility of the manual Nail Psoriasis Severity Index (NAPSI).

Purpose of the Study:

• To develop and validate a deep learning system for automated quantification of the modified Nail Psoriasis Severity Index (mNAPSI) in patients with psoriasis and psoriatic arthritis.
• To enable user-independent and efficient assessment of nail psoriasis severity for improved clinical practice.

Main Methods:

• Photographs of hands from patients with psoriasis, psoriatic arthritis, and rheumatoid arthritis were collected.
• A dataset of 1154 nail photos was annotated with mNAPSI scores, and nails were automatically extracted using key-point detection.
• A transformer-based neural network (BEiT) was trained to predict mNAPSI scores, with performance evaluated using AUC and PR-AUC metrics.

Main Results:

• High inter-reader agreement (Cronbach's alpha of 94%) was achieved for manual mNAPSI scoring.
• The trained neural network demonstrated strong performance, with an AUC of 88% and a PR-AUC of 63%.
• Patient-level predictions aggregated from the network showed a high positive Pearson correlation (90%) with human annotations.

Conclusions:

• Automated quantification of mNAPSI using neural networks is feasible and accurate.
• The developed system overcomes the limitations of manual NAPSI assessment, offering a potential tool for clinical use.
• Open access to the system facilitates its integration into clinical practice for objective nail psoriasis severity assessment.