A snapshot can rank a mouse's age: within-session ordering hits 65%, and mice that look older are frailer with 14% higher death hazard
Synopsis
A laboratory mouse's age can be read from an ordinary photograph, but what transfers is relative age rather than absolute age. Across 2,846 photographs of 1,292 mice from seven facilities, a frozen DINOv2 vision model with ridge regression predicted age; on unseen facilities the median absolute error was 138 days, yet within a session the older of two mice was ranked older in 65% of pairs, rising above 95% for age gaps over 500 days. Mice predicted older than their age were frailer, with each 100 days of photographic age gap linked to 0.12 SD more frailty and a 14% higher hazard of death.
Interpretation
An ordinary photograph carries a relative age signal in mice: among mice photographed in the same session, the model ranks the older one higher, and this ordering holds at facilities it has never seen. Biological-age readouts in mice previously relied on the lifespan endpoint, molecular clocks needing tissue or blood, or frailty indices scored by trained staff; photographs taken as illustrations were never reliably quantified. Under leave-one-facility-out validation, within-session pair concordance was 0.65 (95% CI 0.62–0.68) for unseen facilities and 0.69 (0.66–0.72) for unseen mice; concordance was about 0.97–0.98 for age differences above 500 days but only about 0.54–0.65 below 100 days.
The photographic age gap carries biological information: mice that look older than their age are frailer and die sooner, and fourth-generation telomerase-deficient mice look older than the wild-type cohort photographed alongside them. Absolute age transferred poorly between facilities, with a mean absolute error of 162 days and a median of 138 days, mostly from facility-level offsets, so the study turned to offset-free ranking and to a calibrated age gap. Two facilities scored frailty with a vitality score (296 mice) and a frailty index (98 mice); each 100 days of age gap corresponded to 0.12 SD more frailty (0.06–0.18, P<0.001). Among 296 control mice, adjusting for age and sex, each 100 days of gap raised the hazard of death (hazard ratio 1.14, 1.06–1.23, P=0.001); TERC⁻/⁻ G4 mice looked 151 days older than wild-type controls (84–219 days, P<0.001, 6 versus 19 mice).
The model crops the mouse out of the photograph, replaces the background, and turns the image into 768 features with a frozen self-supervised vision model, which a ridge regression maps to age. The network was not fine-tuned, because end-to-end training and a conservative version keeping early layers frozen both failed to match the frozen model on held-out facilities, and end-to-end training collapsed to near-constant predictions in the largest facility. A YOLO detector locates and crops the mouse, Grounding DINO and SAM trace its outline and the background is set to flat grey; DINOv2 ViT-B/14, pretrained by self-supervision on 142 million images and not trained on mice, produces the features, averaged over the image and its horizontal mirror before ridge regression.
Perspective
The findings rest on 2,846 photographs of 1,292 mice of four strains aged 1 to 34 months from seven facilities, with all main predictions evaluated under leave-one-facility-out splits; frailty scores came from two facilities and survival was analysed in the one facility with 296 control mice. The authors propose a standard pose, background and lighting, plus photographing the same mice at each routine handling, as a way to reduce facility offsets and turn the within-mouse signal into a longitudinal readout.
Facilities differed simultaneously in strain, sex, age range and photographing protocol, so these effects cannot be fully separated; the fourth-generation telomerase-deficient group had only 6 mice from one facility; and only 3 human raters were tested on 100 photographs. Adding the vitality score weakened the age-gap association with death (hazard ratio 1.08, 0.99–1.17), and using only each mouse's first photograph gave 1.09 (1.00–1.20).