To the Editor: Alopecia areata (AA) is a complex autoimmune condition that causes nonscaring hair loss. A systemic review of the epidemiology of AA indicated a similar worldwide lifetime incidence of around 2%.1 The Severity of Alopecia Tool (SALT) was developed to provide a standardized scale for assessing the severity of individual AA cases and is used for clinical practice and research.2 Existing methods for assigning SALT scores are cumbersome, time-consuming, and inconsistent among raters.3
We have developed an artificial intelligence (AI) system for objectively quantifying AA and assigning SALT scores. The AI system computes SALT scores by analyzing 4 images (1 for each quadrant) of a person’s head. The system has 2 main parts. Part 1 uses a deep learning model (Unet with EfficientNetB4 as base model) whose input is a head image and whose outputs are a scalp segmentation (with outputs of 0 or 1) and a heat map (the result of a regression task) showing the percentage of hair loss on that image. The model was trained on 823 head images labelled by trained annotators and verified by doctors using our annotation system. The severity of hair loss is divided into 4 levels: 0; 0.33; 0.66 and 1. The annotation of hair loss was done by area. In part 2, the system compiles the results of the 4 images to compute a single SALT score (the percentage of hair loss) for the entire head. Details are in the Supplementary File, available via Mendeley at https://data.mendeley.com/datasets/kv4658vwgs/1.
We evaluated 188 images containing 47 sets of 4 views (left, right, top, back) for each subject. All subjects are adults, have alopecia aerata and are from the Philadelphia region. The images were taken during clinic visits using an iPad. Each image was independently scored by a human investigator (a veteran U.S. dermatologist) and by the AI system, with the score representing the percentage of hair loss in that image (see Fig 1). We then used intraclass correlation coefficient (ICC) to measure accuracy between scores by the doctor and the AI. We note that based on a 95% confidence interval, ICC scores between 0.75 and 0.9 indicate good correlation, and scores greater than 0.90 indicate excellent correlation (see Fig 2).4 We obtained an ICC score of 0.97 for SALT scoring, and ICC scores of 0.96, 0.97, 0.95, and 0.92 for left, right, top, and back view images.
Fig 1.
Dermatologist-estimated hair loss is 33%. AI analysis: Calculated heatmap (whiter - more severe), scalp segmentation (green) and identified quadrant (blue). AI-estimated hair loss is 15%. AI, Artificial intelligence.
Fig 2.
SALT scores for patients by dermatologist (in blue) and by Al (red). The horizontal axis represents individual patients. AI, Artificial intelligence; SALT, Severity of Alopecia Tool.
Results were not affected by skin tone, skin color, or hair length. Most cases where the doctor and AI scores differed involved severe alopecia or patients with new, white and short hair (doctor scored 100% hair loss but AI scored less, error is about 10%).
Our data supports a potential role for AI techniques in assigning SALT scores. For the next step, we are working on a 3D model for scalp area which will give a much more accurate hair loss percentage.
Conflicts of interest
Dr Hang Nguyen, Lea Gazeau and Dr Wolfe are shareholders of BelleTorus.
Footnotes
Funding sources: The project was supported by BelleTorus Corp. and Torus AI.
Patient consent: Consent for the publication of recognizable patient photographs or other identifiable material was obtained by the authors and included at the time of article submission to the journal stating that all patients gave consent with the understanding that this information may be publicly available.
IRB approval status: Not applicable.
References
- 1.Villasante Fricke A.C., Miteva M. Epidemiology and burden of alopecia areata: a systematic review. Clin Cosmet Investig Dermatol. 2015;8:397–403. doi: 10.2147/CCID.S53985. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Olsen E.A., Hordinsky M.K., Price V.H., et al. Alopecia areata investigational assessment guidelines – part II. National alopecia areata foundation. J Am Acad Dermatol. 2004;51:440–447. doi: 10.1016/j.jaad.2003.09.032. [DOI] [PubMed] [Google Scholar]
- 3.Bernardis E., Castelo-Soccio L. Quantifying alopecia areata via texture analysis to automate the SALT score computation. J Invest Dermatol Symp Proc. 2018;19(1):S34–S40. doi: 10.1016/j.jisp.2017.10.010. [DOI] [PubMed] [Google Scholar]
- 4.Koo T.K., Li M.Y. A guideline of selecting and reporting intraclass correlation coefficients for reliability research. J Chiropr Med. 2016;15(2):155–163. doi: 10.1016/j.jcm.2016.02.012. Erratum in: J Chiropr Med, 16(4), 346. PMID: 27330520; PMCID: PMC4913118. [DOI] [PMC free article] [PubMed] [Google Scholar]