Researchers Use AI to Establish Body Charts for CT Imaging Across the Adult Lifespan

A large-scale effort to turn routine CT data into clinical benchmarks

BY ALEC BUTTNER June 22, 2026

Wachinger

Chai

CT scans may hold far more clinically useful data than radiologists currently tap in routine reporting, according to a recent study published in Radiology: Artificial Intelligence, which analyzed nearly 8 million CT-derived volume measurements to generate reference charts for 104 anatomic structures across adulthood.

“CT scans contain much more quantitative information than is currently used in routine reporting,” said lead author Christian Wachinger, PhD, a professor in the Department of Diagnostic and Interventional Radiology at the TUM Klinikum School of Medicine and Health, Technical University of Munich.

The study found that most anatomical structures do not follow simple linear trajectories over time. Many continue to grow through adulthood before eventually plateauing or decreasing in volume.

To capture those changes over time, Dr. Wachinger and colleagues analyzed repeated CT exams from the same patients, modeling how organ volumes evolve within individuals across follow-up periods of up to eight years. These within-patient trends largely aligned with the broader population patterns.

The researchers found clear sex-based differences, with males exhibiting consistently larger volumes and greater variation between individuals. Additionally, contrast agents significantly inflated measured volumes in the lungs, gallbladder, kidneys, liver and several vascular structures.

Dr. Wachinger said the findings also point to a broader opportunity for the field. “More broadly, the study shows that existing clinical imaging archives are a very rich resource for studying the human body across adulthood,” he said.

800x500 image of doctor with computer and futuristic AI graphics depicting human body and health care imagery

Turning Routine CT Data Into Reference Charts

Reference charts have long been used to compare individual patients to population norms. Pediatric growth charts and MRI-based brain charts offer those such benchmarks, yet no equivalent existed for other anatomic structures across adulthood.

PACS hold vast amounts of clinical imaging data that reflect a wide range of scanners and imaging protocols used in real-world practice, yet have been largely underutilized beyond routine clinical care—that is, until now.

To address that gap, Dr. Wachinger and colleagues sourced over 200,000 CT scans from patients in Germany, Switzerland and Turkey to map how anatomic structures change with age and to develop reference charts for interpreting individual measurements.

“As automated CT segmentation has matured toward clinical deployment, we can now automatically extract the volume of organs and many other anatomical structures from routine CT scans,” Dr. Wachinger explained. “However, we still lack reference charts that tell us how to interpret these measurements.”

Rather than assuming a linear growth pattern for all structures, the group used flexible statistical modeling to account for both average volume and population variability. “A clinically useful reference chart should describe not only the expected value for a given age and sex, but also the variability around that value,” Dr. Wachinger said.

The researchers carefully accounted for contrast enhancement, which most significantly inflated measured volumes in vascular structures. That effect highlights how imaging protocol can influence volumetric measurements, though the analysis does not break down differences by contrast phase or acquisition technique.

“Clinically, this means that centile scores should account for scan protocol, especially when interpreting vessels or other contrast-sensitive structures,” Dr. Wachinger said.

Even with that variability, the researchers found that clinically meaningful signals could still emerge.

To test the clinical utility of the centile scores, the researchers turned to cardiomegaly as a natural proof of concept. “Heart enlargement is clinically meaningful and directly related to organ volume,” Dr. Wachinger said.

Patients with cardiomegaly tended to fall toward the higher end of the centile range. “This supports the broader idea that centile scores can help translate raw anatomical measurements into clinically interpretable information,” he added.

“As automated CT segmentation has matured toward clinical deployment, we can now automatically extract the volume of organs and many other anatomical structures from routine CT scans. However, we still lack reference charts that tell us how to interpret these measurements.”

— CHRISTIAN WACHINGER, PHD

Promise Amid Limitations

Experts outside the research group also praised the study's potential. “The study is an important conceptual milestone for radiology AI,” said Yuan Chai, PhD, an associate professor in the College of Medicine and Biological Information Engineering at Northeastern University in China. “It points beyond task-specific algorithms and toward a broader infrastructure for quantitative, comparable, and longitudinally trackable radiology.”

Dr. Chai said the findings also offer a glimpse of how AI could fit into future practice. “In five or ten years, CT scans may be automatically segmented in the background, with most normal findings remaining silent and only reliable, actionable abnormalities being flagged for the radiologist,” Dr. Chai said. “AI would function as a quiet quantitative layer behind routine imaging, not as a replacement for radiologists.”

At the same time, Dr. Chai emphasized the importance of differentiating statistical normality and what is biologically healthy. “Building and validating organ-level healthy reference subsets could help reduce the risk of confusing ‘common in clinical data’ with ‘biologically healthy’,” he cautioned.

Both experts agree there are limitations, namely the data’s limited demographic diversity and the fact that the CT scans came from clinical populations rather than health reference cohorts.

— YUAN CHAI, PHD

The study also does not explicitly account for variation in scan acquisition, such as contrast timing, scanner time or slice thickness, which could influence volumetric measurements.

"It is plausible that organ volume trajectories differ across populations that are not fully represented in our current study," Dr. Wachinger acknowledged. “However, our team is actively working to address this, using additional large-scale CT datasets from the United States to make the sample more diverse."

“The harder scientific problem may be the lack of truly healthy CT reference cohorts,” Dr. Chai suggested. “A pragmatic path may be to identify apparently healthy organs within clinical scans, rather than requiring completely healthy patients, but such ‘organ-level healthy references’ would still need careful validation.”

Dr. Wachinger agreed, noting that improving pathology filtering is therefore an important next step.

“Radiology AI should not end with model development or publication. The real value comes when mature AI tools are used to extract clinically meaningful information from routine imaging and translate it into better patient care,” Dr. Chai said. “This study reminds us that AI can be a practical infrastructure for quantitative medicine, not just a research output.”