A machine learning algorithm may provide greater objective assessment of psoriasis than traditional tools, such as the Psoriasis Area and Severity Index, and this type of algorithm may play an important adjunctive role in disease evaluation, study results published in the Journal of European Academy of Dermatology and Venereology suggest.
The retrospective single-centered study examined the diagnostic accuracy of 203 standardized photographs of patients with plaque-type psoriasis. Analyses compared lesion detection performed by a neural network trained with an unweighted objective function and a network trained with a penalty factor on false predictions of diseased regions. These were compared with manually marked psoriasis lesions on identical images using accuracy, difference in area, and F1-score. Approximately 80% of the marked photographs were utilized for training and the remaining 20% were used for model validation.
The algorithm trained with different weights (background=1.0, healthy=1.0, psoriasis=2.5) had the best overall F1-score (0.71) for the full resolution. Overall accuracy on a single pixel level was 0.91. There was 90% accuracy of the algorithm in approximately 77% of the images. This accuracy rate differed 5.9% on average from the manually marked psoriasis lesion regions. There was a mean difference of 8.1% (95% CI, 5.2-11.0) between algorithm predicted and photo-based estimated areas by physicians.
Study limitations were the use of only area as a measured outcome, whereas factors such as induration, scaling, and redness were not included.
“An artificial intelligence approach like ours would potentially annul such bias and therefore be a more adequate criterion for treatment decisions and evaluation in pharmaceutical studies,” the researchers wrote.
Disclosure: This clinical trial was supported by Novartis. Please see the original reference for a full list of authors’ disclosures.
Meienberger N, Anzengruber F, Amruthalingam L, et al. Observer-independent assessment of psoriasis affected area using machine learning [published online October 8, 2019]. J Eur Acad Dermatol Venereol. doi:10.1111/jdv.16002