AI Is Learning to Read Hair: What Algorithms Can and Can't Do in Scalp Research

By BioLabs Research · October 4, 2026 · 5 min read

Scalp scans are now part of salon visits and smartphone apps, and AI is moving into hair research too. Two systematic reviews published in 2026 show where the algorithms perform well and where the evidence is still thin.

This article summarizes published research and media coverage. It is not medical advice, and it does not describe or refer to any product we supply.

Why hair is a natural target for computer vision

Hair research has long relied on people looking closely at images and counting. Clinicians use trichoscopy, a magnified view of the scalp, along with standard scoring systems. The Severity of Alopecia Tool (SALT) scores alopecia areata, and the Basic and Specific (BASP) classification grades patterned hair loss.

A systematic review in Digital Health explains the problem. Manual reading of trichoscopic images and repeated scoring "remain time-consuming, operator-dependent," and can be thrown off by differences in how images were captured, in hair density and in skin tone (Alanazi et al., Digital Health, Aug 2026). An algorithm that scores the same image the same way every time could, in principle, make those measurements more consistent.

What the newest reviews found

The broad view: 15 studies across scalp conditions. The Digital Health team searched four databases up to December 2025. From 2,256 records, 15 studies met their criteria, and seven of those were published in 2025 or 2026. Thirteen were retrospective image-analysis or model-development studies. Only two were prospective. Seven studies focused on diagnostic classification and six on severity scoring or quantitative measurement. Most used deep learning, usually convolutional neural networks adapted from pretrained models (Alanazi et al., 2026).

The authors say reported performance was "generally high in controlled datasets and structured tasks." The strongest areas were automated SALT estimation and telling psoriasis apart from seborrheic dermatitis. But they conclude that the evidence base "is dominated by retrospective single-source datasets with limited external validation."

The focused view: 17 studies on pattern hair loss. A second systematic review, in Skin Health and Disease (published online July 22, 2026), looked only at machine learning for androgenetic alopecia. It included 15 retrospective cohort studies and two case series. Thirteen used image-based models, and only three were validated in clinical settings. The authors rated the risk of bias as high in 47.1% of the studies (Huang et al., Skin Health and Disease, 2026).

That review also found work beyond images. One study trained models on 117 genetic variants from European men and reported moderate ability to separate any hair loss from none, with an area under the curve (AUC) of 0.702. Another used a hidden Markov model to track disease progression and flagged elevated fasting glucose and hypertension as statistically significant risk factors (Huang et al., 2026).

Counting hairs, one follicle at a time

Hair counting is one of the most concrete uses. The Digital Health review describes a 2022 study from China that used 2,910 trichoscopic images. A two-stage network first detected follicular openings and then estimated counts of vellus, intermediate and terminal hairs. The review reports about 90% cross-validated accuracy for BASP classification, with strong agreement with clinicians (Alanazi et al., 2026).

A smaller example is a 2025 study in the Ewha Medical Journal. Researchers fine-tuned a ResNet-18 network on 318 trichoscopic images, each labeled by board-certified dermatologists using the BASP system. On a hold-out test set of just 20 images, the model reached 0.90 accuracy (95% confidence interval 0.77 to 1.03) and an AUC of 0.97 (Suh et al., Ewha Medical Journal, 2025). The very wide confidence interval is the point: a 20-image test set cannot settle how well a model generalizes.

For research teams, the appeal is measurement. The Digital Health authors suggest that automated, validated scoring could become a scalable outcome measure for cosmetic dermatology studies, alongside established methods such as phototrichogram-based assessment.

General chatbots are not trichologists

Purpose-built models are one thing. General-purpose AI is another. A prospective 2026 study in Diagnostics from Bern University Hospital asked four publicly available multimodal chatbots and 15 dermatologists to read the same set of trichoscopic images. The dermatologists named the correct diagnosis first 58.1% of the time, and trichology experts did best. The chatbots managed 18.2%. Agreement between the AI models and the dermatologists was only "slight to fair." The authors conclude that these general models "currently underperform compared to human experts" in trichology (Signer et al., Diagnostics, 2026).

The bias problem

Both reviews raise the same concern: who the training data represents.

• In the Digital Health review, only one of the 15 studies explicitly validated performance across diverse ethnicities. None reported Fitzpatrick skin phototype or hair texture as formal demographic variables. Three studies used single-ethnicity cohorts, two Korean and one Chinese (Alanazi et al., 2026).
• The Skin Health and Disease authors write that AI training data "historically lack certain ethnoracial groups, which may lead to lower accuracy in these groups." They call for prospective real-world validation and data sets that cover more cases of female-pattern hair loss and a range of skin tones (Huang et al., 2026).

Meanwhile, in beauty tech

At VivaTech 2026, L'Oréal's Kérastase brand premiered a salon scalp-and-hair spa concept. L'Oréal says it starts with a 4K AI-powered diagnostic camera that is "already in use across 9,000 salons" (L'Oréal, updated June 2026).

Smartphone scalp-scan apps are also spreading. A June 2026 Beauty World News explainer notes that these apps "are not medical diagnostic devices in most cases" and mainly identify visual patterns. They struggle to tell temporary shedding from permanent loss and to read poor-quality images (Beauty World News, June 1, 2026).

Our view

AI hair analysis is a real research field, not just a marketing label. On narrow, well-defined tasks like scoring and counting, published models perform well on the data they were built with. What's missing is the step that matters most for science: prospective, multi-site validation across skin tones and hair types. Any question about hair loss belongs with a licensed clinician, not an app.

What we know / What we don't know

What we know

• AI models for trichoscopy and scalp images are an active and growing research area; all 15 studies in the Digital Health review were published between 2020 and 2026 (Alanazi et al., 2026).
• Performance is reported as high on structured tasks such as SALT estimation and BASP classification.

What we don't know

• How these models perform on new populations, devices and lighting conditions, since most studies are retrospective and single-source.
• Whether consumer scalp-scan tools match the accuracy of the models reported in peer-reviewed studies.

Keep reading

More sourced coverage of peptide science, regulation and media trends is on our News page.

Related reading: The Weight-Drug Era Side-Effect Nobody Expected: Why "Hair Loss" Searches Are Spiking

---

Sources are linked inline. This article summarizes published research and media coverage for general information. Company and brand mentions are attributed reporting, not endorsements. It is not medical advice and does not refer to any product we supply.

For research use only. Not for human or veterinary use. Not a drug, food, or cosmetic.

This summary is for research-community awareness only and is not legal advice.