Research October 28, 2026

The Indication Drives the Search. That's the Problem.

By Aldo Ruffolo, DO, MBA

What a six-week SMA occlusion study says about search patterns, safety nets, and AI-drafted reports

Every missed superior mesenteric artery occlusion in a recent CT study had one thing in common. Nobody asked about the SMA.

The study is Harris et al., published in Clinical Imaging (volume 134, article 110810) and covered in a press release this week [1, 2]. It's worth reading past the headline number, because the headline number isn't the most useful finding.

What the study did

The authors trained a 3D convolutional neural network to localize SMA occlusion on post-contrast abdominal CT. Training cases came from reports that a natural language processing model identified as positive, which were then manually annotated [1].

Then they ran it prospectively. Over six weeks the model analyzed 79,163 post-contrast CT exams with abdominal coverage [1]. When the model saw an occlusion and the signed report didn't mention one, the exam went to quality-assurance review by a second radiologist [2].

The numbers from the abstract [1]:

  • Prospective sensitivity was 67.6% and specificity was 99.6%, with an AUC of 0.917.
  • The model flagged 237 exams for QA review. 83 (35.0%) were confirmed missed SMA occlusions.
  • Those 83 represent 40.7% of the SMA occlusion cases identified during the period.
  • Incidence was 0.26%.
  • Exams with a positive model result reached a radiologist faster, with a median delay of 5.1 minutes versus 27.9 minutes for model-negative exams.

The finding that matters

The most useful line in the abstract is about why the misses happened.

None of the QA-detected cases mentioned SMA occlusion or mesenteric ischemia in the clinical indication. Among the cases diagnosed on the first read, 17.4% did [1]. And the missed cases were rarely CT angiograms: 2.4% of QA-detected cases were CTAs, versus 46.3% of the cases caught initially [1, 2]. The authors conclude that search pattern, more than an inability to recognize the finding once the vessel is carefully assessed, is the main driver [2].

That matches what we see in the reading room. The order says right lower quadrant pain. We look hard at the appendix, the terminal ileum, the adnexa. The SMA origin gets a glance on the way through, if that. A fully occluded SMA can have attenuation close to the surrounding mesentery, and early on the rest of the abdomen can look unremarkable [2]. Nothing pulls the eye there.

This isn't a story about careless radiologists. It's a story about how search works. The indication sets the priors, and the priors set where attention goes. On a CTA for mesenteric ischemia, the SMA is the first thing anyone checks. On a routine portal venous abdomen for vague pain, it competes with everything else.

Two caveats before anyone quotes 40.7%

First, the authors are affiliated with Virtual Radiologic, the teleradiology practice that built and runs the model [1]. That doesn't make the work wrong, and they published it in a peer-reviewed journal. It does mean we'd like to see the same approach replicated somewhere with no stake in the model.

Second, the denominator. The 40.7% is the share of SMA occlusions that were identified during the study, either on the first read or by the AI-triggered QA review. With 67.6% sensitivity, some occlusions were likely missed by both the original reader and the model. Those cases aren't in the count at all. So the true miss rate on the first read isn't known from this design, and it could be higher. We'd rather say that plainly than repeat a number that sounds more precise than it is.

We've only read the abstract and the press release. We haven't seen the full text, so we can't comment on how missed cases were adjudicated or on outcomes after the QA call.

The same pattern shows up in AI performance

Earlier this month, a separate study in npj Digital Medicine looked at two FDA-cleared pulmonary embolism algorithms running across a 17-facility academic health system [3]. The PE model on dedicated CTPA reached 86.8% sensitivity. The incidental PE model on routine contrast-enhanced CT reached 73.5% [3]. The authors note that performance fell off for peripheral and non-acute emboli in a way that mirrors known human reader limitations [3].

Different finding, different design, same shape. When the exam wasn't ordered to look for the thing, the thing gets found less often. That holds for us and, apparently, for at least some of the models.

ACR made a related point in its October comment letter to FDA on generative AI devices. Its proposed list of clinically meaningful failures for imaging includes "omission of a clinically significant incidental finding" alongside laterality errors, fabricated comparisons to priors, and erroneous follow-up recommendations [4].

Why this matters for AI-drafted reports

At ApertureAI we're building Portal, a workflow that places AI-drafted reports in PACS for a radiologist to review, edit and sign. So weigh our bias here. This study changed how we think about one specific risk.

A draft report that takes its cue from the indication will inherit the same blind spot we have. If the order says appendicitis and the draft is organized around the appendix, the radiologist reviewing it is primed twice: once by the order and once by the draft. A clean, well-written draft that says nothing about the SMA reads like a report where someone looked at the SMA. That's the failure we worry about.

Three things we think follow from that:

  1. The draft has to cover what nobody asked about. For a critical finding like SMA occlusion, the check shouldn't depend on the indication. It should run on every eligible exam.
  2. The check should land before the signature, not only after. The QA design in this study caught 83 misses, which is real value. But a QA catch happens after the original report is final. A finding surfaced at draft time can be in the first report instead of the addendum.
  3. The record should show it. If a draft raised a finding and the radiologist removed it, or the draft was silent and the radiologist added one, that should be traceable. That's how a department learns whether its safety net is working, which is the question ACR keeps asking FDA to make sponsors answer [4].

None of this replaces a disciplined search pattern. It adds a second chance at the moment it's most useful.

What we'd want to know next

  • How often does a QA flag for SMA occlusion change management, and how long after the first report does the call happen?
  • What does the false-flag load look like for the reviewer? In this study, 154 of 237 flags were not confirmed misses [1]. Over six weeks that's manageable for a large group. It might not be for a small one.
  • Would the miss pattern look the same in a hospital practice where readers have more clinical context than a teleradiology requisition gives them?

A question for other radiologists

What's your habit for the SMA on a routine portal venous CT? Do you have a hard stop in your search pattern for it, or does it ride along with the aorta?

References

  1. Harris RJ, Baginski SG, Bronstein Y, Schultze D, Lozano KD, Mehta PN, Rethy MC, Sparrman KL, Kim S. Automated detection of superior mesenteric artery occlusion on post-contrast CT using a 3D deep learning model. Clinical Imaging. 2026;134:110810. doi:10.1016/j.clinimag.2026.110810. https://doi.org/10.1016/j.clinimag.2026.110810 (abstract also at https://scholarlyworks.lvhn.org/radiology-diagnostic-medical-imaging/2466/)
  2. vRad. "vRad Study Finds Radiologists Miss 40.7% of Superior Mesenteric Artery Occlusions Across 79,163 CT Exams." Business Wire, October 27, 2026. (Company press release, secondary source; used only for the authors' interpretation and clinical background.) https://www.businesswire.com/news/home/20261027673556/en/vRad-Study-Finds-Radiologists-Miss-40.7-of-Superior-Mesenteric-Artery-Occlusions-Across-79163-CT-Exams
  3. Mansuri A, Chavoshi M, Dapamede T, et al. Multi-site real-world performance of commercial AI for pulmonary and incidental pulmonary embolism detection. npj Digital Medicine. Published October 21, 2026. doi:10.1038/s41746-026-03418-3. https://doi.org/10.1038/s41746-026-03418-3
  4. American College of Radiology. Comments to FDA on "Considerations for the Regulation of Generative AI-Enabled Medical Devices," Docket FDA-2026-N-7874. October 19, 2026. https://edge.sitecorecloud.io/americancoldf5f-acrorgf92a-productioncb02-3650/media/ACR/Files/Advocacy/ACR-Comments-FDA-GenAI-Device-Oversight.pdf (summary: https://www.acr.org/News-and-Publications/2026/acr-comments-genai-device-oversight)