Dr. Tiff, Elena, Other People

Alopecia is not a single disease, yet clinicians often face the same measurement challenge across diagnoses: determining how much hair has been lost, where that loss is occurring, and whether it is changing over time. A study published in JAMA Dermatology examined whether one image-based measurement framework could quantify hair loss across scarring and nonscarring alopecia rather than requiring a separate computational approach for every condition.

The central insight is clinically important: although established scales such as SALT, Sinclair, and Olsen emphasize different patterns or regions, they may share a measurable foundation – the percentage and spatial distribution of hair loss visible in standardized scalp photographs.

What the Study Examined

The multicenter retrospective study developed and evaluated HairComb, a convolutional neural network designed to calculate hair-loss percentage from a scalp image independent of alopecia subtype. The full development and validation data set included 1,605 images from 404 participants aged 2 years and older, collected through the Children’s Hospital of Philadelphia, the University of Pennsylvania, and a Penn Dermatology web interface.

The investigators also analyzed how three familiar photographic scoring approaches relate to the same underlying image information: the Severity of Alopecia Tool (SALT) used in alopecia areata, the Sinclair scale used in female-pattern hair loss, and the Olsen Top Extent Scale used in central centrifugal cicatricial alopecia (CCCA).

Key Findings

• Different alopecia scales captured closely related information. Olsen and Sinclair scores were strongly correlated with one another (R² = 0.963). Both scales also showed strong logarithmic relationships with the underlying percentage of hair loss, suggesting that diagnosis-specific visual grades may partly reflect a shared quantitative signal.
• Hair loss could be measured locally, not only as one global score. HairComb generated a continuous estimate of hair loss at each pixel of the visible scalp. This produces a spatial map that can describe the distribution, concentration, and average extent of loss rather than reducing the entire image to a single category.
• The model performed across multiple alopecia presentations. On the multiview image set, HairComb achieved 92% segmentation accuracy, a 5% regression error, and a 7% average absolute difference between automated and manually derived percentage-affected-area measurements.
• Automated outputs could be translated back into familiar clinical scales. Using information extracted by HairComb, prediction models showed strong agreement with manual Olsen and Sinclair scores, with correlations of 0.87 and 0.90, respectively, on top-view images not used to train the models.
• The data included clinically meaningful visual diversity. The study included straight, relaxed, wavy, curly, and tightly coiled hair, as well as a range of hair colors and light, medium, and dark skin tones. This is relevant because contrast, styling, texture, and scalp exposure can materially affect image-based hair assessment.

Why These Findings Matter

Dermatologists use different scoring systems because alopecia disorders differ in etiology, pattern, and clinical consequence. The study does not argue that these disease-specific scales are interchangeable. Instead, it suggests that a common quantitative layer could sit beneath them: objective measurement of hair loss across the scalp, followed by interpretation tailored to the condition and clinical question.

That distinction has practical value. A continuous measurement can preserve changes that may be difficult to appreciate within a limited number of visual grades. It can also support comparison across visits, readers, sites, and clinical trials – provided that images are captured consistently and the relevant scalp regions are visible.

For clinical research, this framework points toward image-derived measurements that complement established endpoints rather than simply recreating them. A localized hair-loss map may support analysis of treatment response, regional progression, or patterns that are compressed when represented by a single ordinal score. For routine care, the same approach could strengthen longitudinal documentation by making change more explicit and reproducible.

The broader implication is that standardization begins before the algorithm. Image angle, scalp exposure, hair positioning, and consistency across time points determine what can be measured. Quantitative analysis is most clinically meaningful when the capture protocol and the measurement framework are designed together.

Important Context

These findings should be interpreted in the context of a retrospective computational imaging study. HairComb can analyze only what is visible: hair that obscures an affected region may cause loss to go undetected, and multiple images may be needed to expose patterned thinning or other distributed involvement. The study focused its scoring-system analysis on SALT, Sinclair, and Olsen because each had been clinically validated for grading two-dimensional photographs.

The automated percentage-affected-area difference was 7%, and automated scale predictions did not eliminate disagreement inherent in visual scoring. Additional prospective validation, standardized longitudinal capture, and evaluation in broader clinical settings would be needed before treating an image-derived measure as a validated clinical-trial endpoint or a replacement for clinician assessment.

Read the Full Publication

The full article provides the image-labeling methodology, model architecture, scoring-system analysis, participant characteristics, performance evaluation, and detailed limitations. Read the complete article in JAMA Dermatology.

Reference

1. Gudobba C, Mane T, Bayramova A, et al. Automating hair loss labels for universally scoring alopecia from images: rethinking alopecia scores. JAMA Dermatol. 2023;159(2):143-150. doi:10.1001/jamadermatol.2022.5415

Read the full publication in JAMA Dermatology.