Open-Source CT AI Pushes the Commercial Black Box Toward a Testable Clinical Threshold

An abdominal CT scan often answers more than the question a physician initially asks. The scan images may also contain clues about whether the abdominal aorta is abnormally dilated or whether vertebral bone density is low; but amid busy clinical workflows, such information is not necessarily read out systematically every time. A recent biomedical AI preprint uploaded to arXiv takes this kind of “opportunistic imaging analysis” in a less common direction: it not only reports algorithm performance, but also releases the code, weights, and key validation data for models already cleared by the U.S. FDA through 510(k).

This open-source software package, named Comp2Comp, includes two deep learning pipelines. One is for quantitative abdominal aorta analysis, used to segment the abdominal aorta and estimate its maximum diameter to help assess the size of an abdominal aortic aneurysm. The other is for bone density estimation, which segments vertebrae to estimate trabecular bone density and further determine the risk of low bone mineral density. Both use CT images that have already been acquired, so in theory they do not require additional scanning or increase patients’ radiation exposure.

The research team reports that the abdominal aorta module was validated on CT scans from 258 patients across four external institutions, with the sample deliberately enriched for abdominal aortic aneurysm cases. Compared with radiologist annotations, the model’s estimated maximum aortic diameter had a mean absolute error of 1.57 millimeters, with a 95% confidence interval of 1.38 to 1.80 millimeters. The bone density module was tested on CT scans from 371 patients across four external institutions, using contemporaneous DXA bone density examinations as the reference; for binary classification of low bone density, sensitivity was 81.0% and specificity was 78.4%.

These figures indicate that the models already have support from multicenter data, but they should not be read as infallible clinical answers. Differences of a few millimeters in abdominal aortic diameter may have practical significance near certain follow-up or referral thresholds; bone density classification also still involves false negatives and false positives. More importantly, the paper is currently a preprint and has not yet undergone peer review. The information provided in the study abstract is also insufficient to fully assess population distribution, differences in scanning protocols, patterns of failure cases, and real-world performance when the models operate within different hospital information systems.

What is distinctive about this work lies not only in the models themselves, but in how it touches on the long-standing transparency problem in medical AI. Many commercial radiology AI tools have obtained regulatory clearance, yet outside users have difficulty seeing enough detail, and hospitals often only come to understand their limitations after procurement or clinical trials. If Comp2Comp does indeed release its code and weights as described in the paper, it gives researchers and medical institutions a chance to rerun the models on local data, stress-test them, compare biases, and then decide whether to bring them into clinical workflows.

But open source does not mean automatically usable, nor does it make regulatory responsibility disappear. Even if a model has received FDA 510(k) clearance, hospitals still have to deal with version management, data drift, cybersecurity, imaging quality control, and assignment of responsibility. If researchers modify or retrain the models, new validation and compliance requirements may also arise. Comp2Comp offers a more transparent starting point: it turns biomarkers that may have been dormant in CT images into measurable information, while also reminding the field that the real test for clinical AI often lies not in demonstrating accuracy once, but in whether it can be independently inspected, reliably deployed, and traced when errors occur.