Google Tests PhotoScan for Insulin-Resistance Risk From Smartphone Images
The research prototype approaches a DXA-based classifier in one 132-person validation cohort, but it is not a diagnostic product or evidence of clinical readiness.
Google Research published an August 17 report on PhotoScan, a deep-learning method that estimates body composition from standard smartphone images. The research team then combined those estimates with basic demographic information to classify insulin resistance in a retrospective clinical-research cohort.
The work matters to medical-AI and digital-health researchers because body-fat distribution can carry information that body mass index misses, while DXA scans require specialised equipment. The boundary is equally important: Google describes PhotoScan as a research prototype. This is not a released diagnostic product, a medical recommendation or evidence that the method is ready for patient decisions.
From two-dimensional photos to body composition
PhotoScan processes frontal and lateral body images and estimates three measures: total body-fat percentage, the ratio of abdominal to hip-and-thigh fat, and the ratio of visceral to subcutaneous abdominal fat. The research uses those predicted features alongside age, sex and BMI rather than attempting to infer insulin resistance from a photograph alone.
The model was pre-trained on 35,323 UK Biobank participant records, where the team had MRI-derived body information and DXA ground truth. It was then fine-tuned on a newly recruited PhotoBIA cohort of 677 adults using smartphone photos paired with DXA measurements.
A separate MetabolicMosaic cohort supplied the final validation set. The paper reports complete data for 132 individuals in a 30-week San Francisco study, including DXA, PhotoScan, blood tests and other measurements. In that sense, the validation cohort was separate from the large pre-training and fine-tuning groups, although the evaluation still came from the same research programme and authors.
The reported classification result
The team's demographics-only insulin-resistance classifier reached an AUROC of 0.692. Adding PhotoScan-derived body-composition features raised that figure to 0.760, with a reported DeLong-test p-value of 0.002 and a net reclassification index of 0.593.
For comparison, adding DXA-derived body-composition data to the same demographic baseline produced an AUROC of 0.773 and a net reclassification index of 0.748. The useful finding is therefore narrower than replacing a clinical scan: in this cohort, image-derived body-composition estimates brought the classifier close to the result obtained from DXA-derived features.
AUROC measures discrimination within an evaluated dataset; it does not establish that a model is calibrated for clinical decisions, that it will work equally across populations, or that using it improves patient outcomes. The numbers also come from the research team and preprint rather than an independent replication.
Why the evidence is not clinical proof
The MetabolicMosaic analysis involved 132 people, of whom the paper lists 32 as insulin resistant. The cohort came from one San Francisco study and is small compared with the UK Biobank pre-training set. That makes the independent-cohort result useful for feasibility while leaving substantial uncertainty about performance across healthcare systems, camera conditions, body types and demographic groups.
PhotoScan also relies on body images plus demographic inputs. Any future real-world implementation would need explicit consent, secure handling of sensitive images, clear retention rules and evidence that errors do not fall unevenly across populations. Those product, privacy and governance questions were not resolved by the retrospective performance comparison.
The next verifiable checkpoints are external multi-site replication, prospective studies, subgroup and calibration results, and a defined clinical-use pathway. Regulatory review and evidence that the method improves care would be separate requirements if PhotoScan ever moved from research into a medical product.
Status
Learning, with medium internal confidence. The study design and reported results are available in Google's research article and the team's preprint, but the evidence is author-produced, retrospective and not independently reproduced here.
Sources
Update note: Last reviewed 2026-08-18. We will revise this post if independent validation, prospective clinical evidence, product availability or regulatory information becomes available.
Sources
- Google Research — Seeing beyond BMI with smartphone imagery — official
- PhotoScan research-team preprint — research
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.