AI Body Composition Scans Hold Steady Across CT Reconstruction Algorithms, Study Finds

Computed tomography has quietly become one of the most powerful tools in precision medicine, not just for spotting tumors or fractures, but for turning a routine abdominal scan into a detailed map of a patient’s health. By measuring muscle mass, fat distribution, and even bone density at specific vertebral levels, radiologists can extract quantitative biomarkers that predict surgical outcomes, track chronic disease, and flag frailty long before symptoms appear. Yet a nagging question has lingered over this booming field: do the mathematical algorithms used to reconstruct CT images from raw scanner data subtly change these measurements? A new study published in BMC Medical Imaging tackles that question head-on, and its answer is both reassuring and cautionary in equal measure.

The research, led by Chuluunbaatar Otgonbaatar of the Medical Imaging AI Research Center at Canon Medical Systems Korea and colleagues at Yonsei University Wonju College of Medicine and the Mongolian National University of Medical Sciences, took advantage of a uniquely controlled setup. The team enrolled 202 adults, with a mean age of 65.5 years and 112 men among them, who underwent abdominopelvic CT between April and December 2025. Critically, every scan in the study was reconstructed from the exact same raw projection data using three different algorithms: filtered back projection, the classic workhorse of CT; hybrid iterative reconstruction, which refines images through repeated correction cycles; and deep learning reconstruction, the newest entrant that uses neural networks to suppress noise while preserving anatomical detail. Because the underlying x-ray data were identical, any differences in the final measurements could be attributed squarely to the reconstruction method rather than to variations in radiation dose, patient positioning, or biology.

To read out body composition values, the researchers deployed a fully automated deep learning segmentation tool, removing the human variability that has traditionally plagued quantitative imaging. The software measured adipose tissue density and area, muscle density and area, organ attenuation and volume at the third lumbar vertebral level, a standard anatomical landmark for body composition analysis, and calculated bone mineral density at the first and second lumbar vertebral bodies. Agreement across the three reconstruction algorithms was then evaluated using Bland-Altman analysis, which quantifies systematic bias and the 95 percent limits of agreement, together with intraclass correlation coefficients, a statistical measure of consistency that approaches 1.0 when measurements are essentially interchangeable.

The headline result is that the measurements clinicians rely on most heavily proved remarkably robust. Subcutaneous fat area, the fat layer just beneath the skin, showed a bias of just 0.05 square centimeters between filtered back projection and deep learning reconstruction, with 95 percent limits of agreement spanning a razor-thin range from minus 1.51 to 1.61 square centimeters. Skeletal muscle area, arguably the single most important biomarker in CT body composition analysis, differed by only minus 0.32 square centimeters between filtered back projection and hybrid iterative reconstruction, with limits of agreement from minus 1.42 to 0.76 square centimeters. Visceral fat area, the metabolically dangerous fat wrapped around internal organs, showed a bias of minus 0.56 square centimeters with limits from minus 2.04 to 0.91 square centimeters. Bone mineral density measurements were similarly stable, differing by just 0.12 milligrams per cubic centimeter between hybrid iterative and deep learning reconstruction. Across the evaluated measurements, intraclass correlation coefficients ranged from 0.953 to a perfect 1.000.

But the study also uncovered a genuine weak spot. While raw areas of muscle and fat barely budged across algorithms, the parameters that depend on attenuation thresholds proved far more sensitive. These measurements classify muscle tissue into low-attenuation and normal-attenuation categories based on Hounsfield unit cutoffs, a technique used to quantify fatty infiltration of muscle, itself a powerful marker of poor prognosis in cancer, liver disease, and aging. Because reconstruction algorithms alter the noise texture and local attenuation values of an image, pixels hovering near a classification threshold can flip categories, causing low-attenuation and normal-attenuation muscle areas to vary substantially between reconstructions even though the total muscle area stays constant. The lowest intraclass correlation coefficient in the study, 0.953, was observed precisely for low-attenuation muscle area between filtered back projection and deep learning reconstruction, and most area-based measures other than these threshold-dependent ones showed small absolute differences of only one to two square centimeters.

The physics behind this vulnerability is worth unpacking. Filtered back projection reconstructs images by mathematically inverting the projection process, a fast and elegant method but one that leaves images relatively noisy. Iterative reconstruction repeatedly compares simulated projections with the measured data and refines the image, trading computation time for noise reduction. Deep learning reconstruction goes further still, training neural networks on vast image datasets to distinguish noise from true anatomy, achieving what the study confirmed with noise power spectrum analysis: a progressive reduction in image noise moving from filtered back projection through hybrid iterative reconstruction to deep learning reconstruction. That noise reduction is wonderful for diagnostic image quality, but because computed attenuation values shift subtly as noise characteristics change, any measurement that draws a hard line at a specific attenuation value inherits that instability.

The implications ripple outward across modern medicine. CT body composition analysis is exploding in clinical relevance, guiding oncology dosing, predicting outcomes in surgery and transplantation, characterizing sarcopenia and obesity, and feeding large epidemiological studies that mine millions of existing scans. Automated AI segmentation tools are multiplying rapidly, promising to make these analyses routine rather than labor-intensive. But if a hospital upgrades its scanners and switches from iterative to deep learning reconstruction, could longitudinal measurements of a patient’s muscle quality appear to change even when the patient’s actual muscle has not? The new findings suggest that for simple area and volume measurements, the answer is reassuringly no, at least within a single scanner and protocol, but for attenuation-threshold biomarkers such as low-attenuation muscle area, reconstruction settings genuinely matter and could confound trend analyses.

The study’s authors are careful to frame their conclusions within the limits of their design. The work was conducted under a single-scanner protocol at a single institution, approved by the Institutional Review Board of Wonju Severance Christian Hospital with the requirement for informed consent waived given its retrospective nature. The findings, they emphasize, support the use of consistent reconstruction settings whenever attenuation-dependent biomarkers are involved, and they call for validation across different scanners, scanning protocols, software platforms, and longitudinal examinations before the results can be generalized. Two of the co-authors, Otgonbaatar Chuluunbaatar and Hackjoon Shim, are employees of Canon Medical Systems Korea, a subsidiary of Canon Medical Systems Corporation, a competing interest the journal discloses, though the other authors report no relevant financial interests.

Still, the study delivers a message that will shape how quantitative CT is deployed in the AI era. The marriage of deep learning reconstruction and deep learning segmentation is inevitable, offering cleaner images and fully automated biomarker extraction from scans already acquired for other reasons. This research provides the kind of rigorous, physics-aware validation that must accompany that marriage, demonstrating that the raw anatomical quantities survive the reconstruction pipeline essentially intact, while sounding a precise and actionable warning about derived, threshold-based measures. As hospitals worldwide modernize their CT fleets, the practical prescription is clear: lock in your reconstruction settings, document them, and treat attenuation-threshold biomarkers as sensitive to the imaging pipeline itself. In quantitative imaging, the algorithm behind the picture can matter almost as much as the picture itself.

Subject of Research: Effect of CT image reconstruction algorithms on automated body composition and bone mineral density measurements

Article Title: Agreement of automated computed tomography body composition measurement across image reconstruction algorithms under identical projection data

Article References: Otgonbaatar, C., Kim, H., Kim, J. W., Dagvasumberel, G., Cha, S.-J., Jeon, S.-H., Shim, H., & Ahn, J.-H. (2026). Agreement of automated computed tomography body composition measurement across image reconstruction algorithms under identical projection data. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02938-4

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02938-4

Keywords: computed tomography, body composition, deep learning reconstruction, iterative reconstruction, filtered back projection, skeletal muscle area, adipose tissue, bone mineral density, Bland-Altman analysis, medical imaging AI, sarcopenia, radiology