The study’s design reflects a growing recognition within the imaging-biomarker community that the unit of statistical analysis matters as much as the imaging technology itself. By treating each mouse as the independent biological unit and nesting organ-level observations within animals, the investigators avoided a common pitfall in preclinical radiomics: inflating sample size by counting heart, liver, and spleen segmentations as if they were independent specimens. With only eleven animals distributed across three age groups, the effective sample size remains small, and the reported confidence intervals are correspondingly wide. The 95% confidence interval around the 90.9% accuracy of the photon-counting standard spectral processing pipeline stretches from roughly 62% to 98%, which means the true performance of the method in the broader population of similar mice could plausibly fall anywhere within that band. This is not a weakness unique to this study but rather an inherent property of exploratory work with limited animal numbers, and it explains why the authors frame their findings as hypothesis-generating rather than definitive.

The choice of three chronological age points—28, 88, and 99 weeks—deserves comment in the context of murine aging biology. A 28-week-old C57BL/6NRj female is a young adult, roughly comparable in developmental terms to a human in early adulthood, while mice at 88 and 99 weeks have entered a phase corresponding to middle and advanced age in the murine life course. Because the C57BL/6N lineage is inbred, genetic variability is minimized, which sharpens the contrast between age groups: any radiomic differences that emerge are more likely to reflect age-associated structural and compositional changes in tissue rather than inter-individual genetic variation. This uniformity is a double-edged sword, however, since inbred strains may not capture the heterogeneity of aging seen in outbred populations, and the same radiomic signatures may behave differently in genetically diverse cohorts or in humans.

The technical distinction between photon-counting detectors and conventional energy-integrating detectors underlies much of the observed performance difference. Photon-counting systems register individual X-ray photons and bin them by energy, which suppresses electronic noise, permits narrower detector pixels, and enables the synthesis of virtual monoenergetic images at arbitrary kiloelectronvolt levels. Monoenergetic reconstructions at higher energy levels reduce beam-hardening artifacts and can improve the conspicuity of soft-tissue interfaces, while lower-energy reconstructions accentuate contrast between materials with different attenuation curves. The finding that classification accuracy peaked at 110 keV in most outer cross-validation folds, yet the highest macro one-versus-rest area under the curve appeared at 130 keV, illustrates that no single reconstruction energy uniformly maximized all performance metrics. This energy dependence is a meaningful practical consideration: any future translation of CT-based aging phenotyping would need to standardize not only scanner hardware but also reconstruction parameters, because the radiomic values themselves shift with the energy spectrum of the reconstructed image.

The radiomic feature set employed here was deliberately comprehensive. More than one thousand features per organ were derived from the original images, from Laplacian-of-Gaussian-filtered images at two smoothing scales, and from eight wavelet decompositions, which together capture shape, intensity distribution, and multi-scale texture information. Laplacian-of-Gaussian filtering emphasizes structures at a chosen spatial scale, so features derived at sigma values of 3 and 5 millimeters probe coarser tissue architecture, while wavelet decompositions separate information into frequency bands that can highlight fine texture, coarse texture, and edge-like patterns. Because the images were resampled in-plane to a common 2-millimeter grid before extraction, features computed from the two platforms were at least spatially comparable, even though the through-plane resolution differed substantially between the clinical scanner and the preclinical instrument. The retention of native slice thickness meant that the volumetric sampling differed between platforms, a factor that may have contributed to the performance gap.

The absence of intensity normalization is noteworthy. CT attenuation values in Hounsfield units are physically calibrated, which is one reason CT is often considered attractive for radiomics: unlike magnetic resonance imaging, where signal intensities are not standardized across scanners, Hounsfield units should in principle be comparable across systems. Yet in practice, differences in detector technology, beam quality, reconstruction kernels, and scatter correction mean that identical tissues can yield slightly different attenuation values on different scanners. The Qr44 kernel used for the photon-counting reconstructions and the soft-tissue kernel used for segmentation on the preclinical system each impose their own spatial-frequency filtering, which alters texture features even when the underlying anatomy is unchanged. This kernel dependence is a well-recognized challenge in multi-scanner radiomics and likely accounts for part of the observed platform effect.

The nested cross-validation scheme deserves emphasis as a methodological strength. In leave-one-mouse-out validation, each fold holds out all data from a single animal, and every decision about preprocessing, feature selection, model choice, and even reconstruction energy was made using only the training animals within that fold. This fully nested design prevents information from the test animal from leaking into any stage of model construction, which is essential when the number of candidate pipelines is large. The permutation testing framework provides an additional safeguard: by randomly shuffling age-group labels two hundred times while preserving the grouped structure of the data, the authors estimated how often a classifier could achieve the observed accuracy by chance alone. That none of the two hundred permuted analyses matched the real result, yielding a Monte-Carlo p-value of 0.005 after plus-one correction, indicates that the association between radiomic patterns and chronological age groups is unlikely to be an artifact of the modeling procedure, even in this small cohort.

The organs selected for analysis are biologically informative in the context of aging. The heart undergoes measurable changes with age, including alterations in myocardial composition and chamber geometry. The liver accumulates age-related changes in parenchymal texture and fat content, and the spleen is a well-established sentinel of immunosenescence, often shrinking or changing in density as lymphoid tissue declines. Radiomic features sensitive to texture and density could plausibly capture such changes, and averaging organ-level class probabilities into a single mouse-level prediction implicitly weighted information from all three organs, allowing the classifier to integrate complementary aging signals across organ systems.

It is important to situate these findings against the broader goal of imaging-based aging assessment. Chronological age is easy to measure but biologically imprecise; two individuals of the same age can differ substantially in their physiological reserve and disease risk. Imaging biomarkers offer a route to non-invasive, repeatable, and spatially resolved assessment of organ-level aging, and computed tomography is particularly attractive because it is widely available and quantitatively calibrated. The present study stops short of estimating biological age itself, instead asking whether radiomic signatures differ across defined chronological groups, which is a necessary precursor step. Before such approaches could inform human studies, several gaps would need to be bridged: validation in larger and more heterogeneous cohorts, correlation with independent biological-aging reference measures such as molecular or functional benchmarks, and demonstration of robustness across scanners, reconstruction protocols, and imaging conditions. The pronounced platform dependence observed here—where the preclinical scanner achieved chance-level classification while the clinical photon-counting system reached over ninety percent accuracy—underscores that radiomic aging signatures are not yet scanner-agnostic, and that protocol standardization will be as important as algorithm development in any future translation.

Subject of Research: …

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

Article References: Age-associated radiomic phenotyping in mice using multi-energy photon-counting and preclinical CT: an exploratory study. (n.d.). https://doi.org/10.1007/s10522-026-10524-y

Image Credits: AI Generated

DOI: 10.1007/s10522-026-10524-y

Keywords: Age-associated, radiomic, phenotyping, mice, multi-energy, photon-counting, preclinical, exploratory, scientific research