Age-associated radiomic phenotyping in mice using multi-energy photon-counting and preclinical CT: an exploratory study

To investigate whether CT-derived radiomic features capture imaging patterns associated with chronological age groups in mice and to assess the influence of imaging platform and photon-counting CT reconstruction strategy on classification performance. Female C57BL/6NRj mice from three chronological age groups (28, 88, and 99 weeks; n = 11) underwent clinical photon-counting CT (PCCT) and preclinical Albira II CT. Heart, liver, and spleen were segmented and radiomic features were extracted. The individual mouse was defined as the independent biological unit. Model development and evaluation used nested leave-one-mouse-out cross-validation with all preprocessing, feature selection, model selection, and, where applicable, reconstruction-energy selection restricted to the respective training folds. Organ-level class probabilities were averaged to generate one mouse-level prediction. Monoenergetic PCCT reconstructions from 40 to 180 keV were evaluated, including a fully nested energy-and-pipeline selection analysis. A grouped mouse-level permutation analysis with 200 permutations was performed for the primary monoenergetic workflow. Albira II correctly classified 5/11 mice (45.5%; 95% CI, 21.3-72.0%), whereas PCCT-SPP correctly classified 10/11 mice (90.9%; 95% CI, 62.3-98.4%). The fully nested monoenergetic PCCT workflow correctly classified 9/11 mice (81.8%; 95% CI, 52.3-94.9%; balanced accuracy, 83.3%; macro-F1, 83.3%). Energy-specific performance varied considerably: 110 keV showed the highest accuracy (81.8%) and was selected in 7/11 outer folds, whereas the highest macro one-vs-rest AUC occurred at 130 keV (0.867). None of the 200 permuted analyses reached the observed fully nested monoenergetic accuracy (plus-one corrected Monte-Carlo p = 0.005). CT-derived radiomic features showed chronological age-group-associated patterns in this exploratory murine cohort, with classification performance varying substantially across imaging platforms and reconstruction strategies. PCCT-SPP showed the numerically highest performance, while monoenergetic PCCT demonstrated pronounced energy dependence without evidence for a universally optimal reconstruction energy. These findings require validation in larger independent cohorts and correlation with independent biological-aging reference measures.