by Deileta Kamhunga
An artificial intelligence-based psoriasis severity assessment model may be a promising alternative for dermatologists’ assessments, according to a real-world study.
Psoriasis is a chronic, immune-mediated skin disease. The therapeutic management plan for psoriasis is established according to the degree of disease severity. The Psoriasis Area and Severity Index (PASI) is currently the most widely accepted metric for evaluating psoriasis severity.
A study in the Journal of Medical Internet Research developed and validated an image-artificial intelligence (AI)-based system for psoriasis severity assessment to facilitate long-term disease management.
Study Population and Method
A total of 14,096 images from 2367 psoriasis patients were used to train the deep learning system for estimating the PASI score. Of these, 1962 patients were used to train the model and 405 to validate it. The proposed model was then compared with assessments by experienced dermatologists. The proposed model was deployed in a mobile app named SkinTeller to assess psoriasis severity in clinical practice.
Investigating the Performance of the Model With Ablation Studies
The first ablation study examined the number of input lesion images per body part. When the number of input images was increased from one to two, the mean average error (MAE) decreased from 3.47 to 2.25. With the input of three images, the model’s performance reached its peak accuracy, achieving the smallest MAE of 2.05.
The second ablation study evaluated different image fusion strategies to assess the performance of the multiview feature enhancement block. Combining multiple aggregation methods outperformed the methods individually, both in the total score and the four sub-indicators of PASI.
The third study revealed that the combination of the regression header and the classification header with the cross-teacher header outperformed other approaches for output headers. A decrease in the performance of the proposed model was noted when the degree of disease severity increased.
The AI-Based Model Outperformed Dermatologists in Psoriasis Severity Assessment
A total of 43 experienced dermatologists participated in the study. They consisted of 13 professors and 30 attending or resident physicians. The AI-based model outperformed the dermatologists, resulting in a 33.2% improvement in the total PASI score and 23%, 12%, 11%, and 7% accuracy improvements in the four subscores of erythema, area ratio, desquamation, and induration, respectively. The greatest improvement was in erythema, as it has the strongest subjectivity. Using MAE as the main measurement, the proposed model ranked eighth among all dermatologists (dermatologist MAE: 4.67, proposed model MAE: 3.12).
The AI-Based Model Accurately Predicted the Trend of Severity Degree
The model also accurately predicted the direction of severity progress with different ranges of PASI scores. The average trend accuracy was 84.81%. Moreover, the model can predict the severity degree changes between two visits if the score gap is larger than five PASI score points, with an accuracy of 96.09%. Meanwhile, the agreement among dermatologists was poor.
In conclusion, the proposed image-AI-based psoriasis severity assessment model provides an efficient, objective, and accurate automatic estimation of PASI scores. The SkinTeller app may be a promising alternative to dermatologists’ assessments and shows potential for personalized treatment and chronic disease self-management in psoriasis patients.
Source:
Huang, K., Wang, X., Li, Y., Lv, C., Yan, Y., Wu, Z., Zhang, M., Huang, W., Jiang, Z., Hu, K., Li, M., Su, J., Zhu, W., Li, F., Chen, M., Chen, J., Li, Y., Zeng, M., Zhu, J., . . . Zhao, S. (2023). Artificial Intelligence–Based Psoriasis Severity Assessment: Real-world study and Application. Journal of Medical Internet Research, 25, e44932. https://doi.org/10.2196/44932
