Medical AI · asia
Acute Abdominal Imaging Queues Should Not Be Strictly First Come, First Served: AbdomenNet Uses Non-Contrast CT to Help Identify Emergencies
The self-supervised model maintained strong identification performance across three external hospital cohorts and helped radiologists interpret scans faster and more accurately; however, the 37-minute reduction in reporting time still comes from a retrospective simulation and has not yet been shown to improve real-world patient outcomes.
Behind abdominal pain in the emergency department could be something as simple as a ureteral stone, or a time-critical condition such as mesenteric ischemia or aortic dissection. When large volumes of images enter the worklist at the same time, the most critical cases may not be seen first. A study published in *Nature Communications* presents AbdomenNet, which aims to identify patients requiring priority interpretation from non-contrast abdominal computed tomography scans.
Using self-supervised learning, the research team first trained the model to learn imaging features from 103,989 unlabeled non-contrast abdominal CT examinations, then fine-tuned it on 5,816 diagnostically labeled cases. The system can simultaneously identify 11 acute abdominal diseases and further assess risks relevant to surgical decision-making, including high-grade organ injury, strangulated bowel obstruction, and upper gastrointestinal perforation.
In external validation involving a total of 2,528 patients at three independent tertiary hospitals in China, AbdomenNet achieved an overall mean AUROC of 0.919 for five categories of emergency conditions. When all 11 diagnoses were combined, the mean AUROC across centers ranged from 0.889 to 0.898. These cases came from scanners made by different manufacturers and used varying scan settings, suggesting that the model is not limited to a single machine or one hospital’s data. However, all validation centers were still large hospitals in the same country, and its ability to generalize across healthcare systems has not yet been established.
The study also had four radiologists interpret 380 cases using a crossover design. After AI-provided disease probabilities, a top prediction, and heatmaps were added, the radiologists’ mean AUROC rose from 0.812 to 0.924, while the median interpretation time per case fell from 197.5 seconds to 145 seconds. However, the experiment only required identification of 11 prespecified conditions and did not include integration of complete clinical data or report writing, so the 52.5-second difference cannot be directly regarded as the actual time saved in routine practice.
A question closer to real-world emergency care is whether AI can change the order of imaging queues. Researchers reconstructed the workflow using existing time records from 18,398 emergency CT examinations, 1,761 of which were non-contrast abdominal CT scans. The simulation showed that prioritizing emergency cases flagged by the model could reduce the median report turnaround time for truly urgent cases from 49.3 minutes to 12.3 minutes. However, the high-sensitivity threshold also produced many false positives: of 767 priority flags, 203 were true emergencies and 564 were false positives.
The findings are therefore better understood as supporting a “radiologist-supervised triage tool” rather than an autonomous diagnostic system. It also cannot replace contrast-enhanced CT or CT angiography, which remain the confirmatory examinations for many acute abdominal conditions. The study was retrospective, the 37-minute benefit came from workflow reconstruction, and the image-interpretation experiment also used internal cases. The next step is prospective validation in consecutive, unselected patients, measuring missed diagnoses, alert fatigue, cross-hospital calibration, and patient outcomes before it will be possible to determine whether the system is suitable for clinical use and regulatory review.