AI Tool Flags Six Acute Abdominal Emergencies on CT

 An AI system identified six acute abdominal emergencies on CT scans using a multiwindow Hounsfield unit (HU) encoding approach and localised the detected pathology to the clinically appropriate abdominal region with 99.5% accuracy, a study showed.

METHODOLOGY

• Researchers developed and retrospectively analysed a deep learning system trained and internally validated on a patient-level split of a teleradiology dataset comprising 1274 patients with 42,922 bounding box annotations.
• Each CT slice was encoded into three diagnostic HU windows mapped to separate image channels (centre/width in HU): soft tissue (50/400), bone/stone (500/1600), and angio/liver (60/150).
• A "YOLOv11-Large" model with an added stride-4 head was trained at a high resolution of 1280 × 1280 pixels to preserve fine soft tissue details, evaluating the area under the receiver operating characteristic curve (AUROC) and macro F1 scores.
• The model was trained to detect six acute abdominal emergencies: abdominal aortic aneurysm, acute pancreatitis, acute cholecystitis, kidney and ureter stones, acute diverticulitis, and acute appendicitis.
• The researchers mapped anatomic localisations to a clinical nine-region abdominal grid and validated the system using a radiologist-adjudicated 280-patient Stanford Merlin cohort.

TAKEAWAY

• The system achieved a high macro AUROC of 0.941 across all six conditions, with abdominal aortic aneurysm achieving the highest at 0.998 and acute appendicitis achieving the lowest at 0.880; the macro F1 score was 76.1%.
• On external validation, all the six classes maintained an AUROC ≥ 0.80, and macro F1 was 0.545 at unmodified thresholds, increasing to 0.648 after site-specific recalibration.
• Anatomic localisation to the correct abdominal region was 99.5% accurate among detected cases and 90.9% accurate when factoring in missed detections (199/219 patient-pathology pairs).
• Among 80 target-negative patients, the specificity was 86.2%, with 11 patients incorrectly flagged for at least one acute finding, representing a 13.8% per-patient false-positive rate.

IN PRACTICE

"[The study] findings suggest that a region-aware multi-pathology CT triage system may provide clinically interpretable decision support, although prospective validation integrated with a picture archiving and communication system (PACS) remains necessary before clinical use," the authors wrote.

They added that the tool's "layered output is important for emergency CT triage because the radiologist needs both an indication that an abnormality may be present and a rapid sense of where to look and how the finding relates to the clinical presentation."

SOURCE

The study was led by Hasan Mete Erdoğan, Budapest University of Technology and Economics, Budapest, Hungary. It was published online on July 08, 2026, in the Journal of Imaging Informatics in Medicine.

LIMITATIONS

The internal dataset was obtained from a single national teleradiology network. Retrospective external validation was performed using a single US dataset, and external adjudication was performed by a single radiologist. Furthermore, this study was limited by the absence of external bounding box annotations, the unavailability of contrast phase metadata, a two-dimensional per-slice detector design, and the lack of multireader radiologist comparisons.

https://www.medscape.com/viewarticle/ai-tool-flags-six-acute-abdominal-emergencies-ct-2026a1000nwm