LLM Filtering Supports Whole-Body CT Reference Charts

Reference charts can help place routine computed tomography measurements within expected ranges for a person’s age, sex and scan conditions. An analysis led by the Technical University of Munich used 351,915 examinations from one clinical archive and four external CT datasets to build reference distributions across adulthood. Automated software measured the volume and mean attenuation of 106 anatomical structures. Five large language models then screened radiology reports to identify abnormalities that could distort the reference data. Statistical models generated centile charts for organ volume and attenuation, while accounting for contrast enhancement and other acquisition factors.

Must Read: Imaging Biomarker Catalogue Aims to Improve Reuse

Cross-Verification Improves Clinical Data Filtering

Clinical CT archives offer large numbers of scans, broad anatomical coverage and repeated imaging over time. However, they contain many abnormalities because the examinations were performed as part of routine care. Including these findings in reference cohorts can increase variability, shift the limits of expected measurements and make unusually high or low values harder to interpret.

The filtering system grouped the 106 segmented structures into 39 broader anatomical targets. In the first stage, five open-weight language models independently identified abnormal structures in each radiology report. Every proposed finding had to be linked to supporting wording from the report. When the models disagreed, the disputed findings were passed to a second stage for verification using the collected evidence.

This second step increased agreement between the models and improved performance against manually labelled reports. The final MedGemma-based method achieved high recall while improving precision compared with the strongest single-model alternative. It also maintained high recall across different anatomical regions, although performance varied between datasets.

The filtered results were used to remove structure-specific abnormalities before modelling. This created a cohort with less pathology, rather than a completely healthy cohort, because abnormalities not mentioned in the reports could not be detected through text analysis.

Charts Adjust for Biological and Technical Variation

The filtered cohort was used to create cross-sectional reference charts for anatomical volume and mean CT attenuation. The statistical models captured nonlinear changes with age and differences in variability between people. Separate attenuation charts were produced for contrast-enhanced and non-contrast scans because contrast administration strongly changes measured values.

Age patterns differed between organs and were often not linear. In the eight structures presented as examples, typical organ volumes were larger in men, with differences ranging from 18% to 35%. Mean attenuation was higher in women across all example structures. Attenuation also generally fell as tube potential increased. Scanner manufacturer and dataset effects were included because acquisition conditions influenced the measurements.

The models remained stable across adulthood, although uncertainty was greater at ages with fewer available scans. This was particularly important for the outer centiles, which are used to identify unusually high or low measurements.

Filtering pathology changed both the central reference curves and the distribution limits. The largest changes often appeared in the outer centiles. When filtered and unfiltered charts were applied to the same participants, the filtered charts better separated cases from controls for atelectasis, consolidation and mosaic attenuation. Heart-volume centiles were also higher in participants with cardiomegaly, while lung attenuation centiles increased in those with parenchymal abnormalities. The centile scores therefore provide a common way to express both volume and attenuation after adjustment for relevant covariates.

Longitudinal Change Varies by Structure and Contrast

Repeated examinations from the Munich cohort were used to measure within-person change over time. Volume patterns differed by organ and baseline age. The liver, kidney, spleen, heart and gluteus minimus generally became smaller, while the aorta, lung and vertebrae tended to increase. Some changes became stronger or weaker with age. Sex also influenced longitudinal change in selected structures.

Attenuation showed clear differences between contrast conditions. In the liver, kidney, spleen and aorta, attenuation generally decreased over time on non-contrast scans but increased on contrast-enhanced scans. The direction of within-person change did not always match the age patterns seen in the cross-sectional charts. This indicates that comparisons between different people and follow-up measurements in the same person can show different trends.

The charts cover adulthood and do not represent the general population. Most scans came from one institution, although the data included several countries, scanner manufacturers and acquisition settings. Height, weight, body mass index and ethnicity were not consistently available. Remaining pathology, uneven representation by age and segmentation errors may affect the outer centiles most strongly. Use in other languages or reporting environments also requires local validation.

Whole-body CT reference charts can turn organ volume and attenuation measurements into centile scores adjusted for age, sex and scan conditions. Cross-verification between language models helps reduce report-documented pathology in large clinical archives, while flexible statistical models account for nonlinear age patterns and technical variation. Filtering improves the interpretation of extreme measurements, but the resulting cohort is not fully healthy or representative of the wider population. The framework may support quantitative imaging and opportunistic screening research, although further validation is needed before routine clinical use.

Source: arXiv

Image Credit: iStock