AI Tool Extracts Body Composition Data from Routine Chest X-Rays
Multimodal deep learning model could expand access to clinically meaningful body composition metrics
Body composition metrics such as muscle mass and visceral fat are increasingly recognized as powerful predictors of disease risk and treatment outcomes. Yet in routine clinical practice, these measures are rarely obtained because they typically require specialized testing.
A study published in Radiology Advances suggests an AI model could estimate body composition from routine chest X-rays, potentially expanding access to these metrics in everyday care.
“Clinicians frequently make decisions influenced by body composition such as chemotherapy dosing, kidney function, frailty risk and weight-loss treatment,” said Nicholas Stucky, MD, PhD, a physician-scientist with Providence Health & Services in Washington State, and a senior author of the study. “They often do so without actually knowing a patient’s muscle and fat composition, because we lack accessible ways to measure it directly.”
“In the absence of direct measurement, we rely heavily on body mass index (BMI) because it’s simple and accessible—but BMI cannot distinguish muscle from fat and can be misleading at the individual level,” he added.
More precise techniques such as CT, MRI and dual energy X-ray (DEXA) imaging can provide accurate measurements but are considered impractical for routine screening because of cost, radiation exposure or limited availability.
By comparison, chest radiographs are more widely accessible. “Because chest X-rays are one of the most common imaging studies in medicine, we asked whether we could unlock more information from something we already obtain millions of times each year,” Dr. Stucky said.
That question led Dr. Stucky and colleagues to build XComposition, a multimodal deep learning model that estimates body composition from chest radiographs combined with basic clinical data. The approach could enable opportunistic assessment of body composition metrics using imaging at scale.
Listen as Drs. Alipour and Stucky discuss the importance of body composition and how XComposition could expand screening and improve clinical decision-making.
Unlocking Information Hidden in Routine Imaging
XComposition combines chest radiograph data with four clinical variables (age, sex, height and weight) to estimate multiple body composition measures, including visceral fat, subcutaneous fat and skeletal muscle.
The model was trained on imaging and clinical data from 1,118 patients across 30 U.S. health systems and showed strong agreement with CT-derived body composition measurements.
“Each modality looks at the patient from a different angle,” said Ehsan Alipour, MD, PhD, MPH, a clinician-scientist at the University of Washington in Seattle and the study’s lead author. “While each alone can estimate body composition with limited precision, combining them enables the model to find relationships between features and achieve higher performance.”
The multimodal model outperformed imaging-only and clinical-only models, showing strong agreement with CT-based standards. It achieved a correlation of 0.85 for subcutaneous fat and 0.76 for visceral fat, indicating strong concordance.
“These results demonstrate that routinely available imaging contains far more information than we typically extract,” Dr. Alipour said.“Each modality looks at the patient from a different angle. While each alone can estimate body composition with limited precision, combining them enables the model to find relationships between features and achieve higher performance.”
— EHSAN ALIPOUR, MD, PHD
Expanding Access to Body Composition Data
According to Dr. Alipour, the team’s findings demonstrate that the wide availability and relatively low cost of chest radiography can help extend body composition analysis to much larger populations than previously possible.
“We democratize access to these metrics for both researchers and clinicians by enabling body composition calculation for anyone who has ever had a chest radiograph,” he said.
The authors emphasize that body composition metrics are linked to outcomes in multiple conditions, including cardiometabolic disease, cancer prognosis and frailty. They point to oncology, nephrology and geriatric risk assessment as near-term use cases.
“In oncology, muscle mass has been linked to chemotherapy toxicity and outcomes,” Dr. Stucky said. “More precise estimates could help guide dosing or identify patients at higher risk.”
He added that muscle mass also affects interpretation of kidney function tests and could improve risk assessment in aging populations or neurologic disease.
Dr. Alipour noted that opportunistic screening is another promising use.
“By applying this model to routine imaging, we can flag individuals with abnormal values for metrics like visceral adipose tissue,” he said. “Changes could also be monitored over time if patients undergo repeated imaging.”
Saliency maps for feature importance in prediction of the top performing body composition metrics. For the first 3, occlusion-sensitivity saliency maps are depicted, whereas for the vertebral body volume, the saliency map is based on the integrated gradients method. The white box highlights the region of increased importance for vertebral body volume calculation that coincides with the region containing the lower thoracic and upper lumbar vertebral bodies. The images are aggregates of individual chest radiographs and their respective saliency maps in the hold-out test cohort.
https://doi.org/10.1093/radadv/umaf035 ©RSNA 2025
Workflow Integration and Future Steps
The researchers envision the tool operating automatically within radiology systems.
“A good use case would be the integration into PACS so the model runs whenever a chest radiograph is performed,” Dr. Alipour said. “The resulting metrics could be stored in the electronic health record and accessed if clinicians need them.”
Such integration could allow opportunistic analysis without additional radiation, imaging time or cost.
Dr. Stucky emphasized that further validation is needed before clinical deployment. “Prospective studies demonstrating improved decision making or patient outcomes will be essential,” he said. “Regulatory review, workflow integration and clinician education will also be important.”
Beyond integrating this model into PACS and the EHR, the study underscores the growing role of multimodal AI approaches in radiology.
“One big takeaway is that routinely performed imaging contains much more information than we currently use,” Dr. Alipour said. “Combining imaging with clinical data can uncover insights that weren’t visible before.”
Dr. Stucky agreed, noting that such tools may enhance the value of imaging without increasing complexity.
“I hope radiologists see this as an example of how AI can expand the value of imaging without requiring more expensive modalities,” he said. “It shows how we can extract more meaningful information from studies we already perform every day.”
For More Information
Access the Radiology Advances study, “XComposition: multimodal deep learning model to measure body composition using chest radiographs and clinical data.”
Read previous RSNA News stories on opportunistic imaging: