This study was approved by the Regional Committee on Health Research Ethics and the Regional Knowledge Centre on Data Protection Compliance (P-2020-663). Data were prospectively collected.

Two main criteria were required for inclusion in the study. First, patients had to be referred to the Department of Diagnostic Radiology on the suspicion of ABI and undergo a venous phase DECT or multiphase ABI protocol including a venous phase DECT and a conventional non-contrast and arterial phase scan. Second, patients had to undergo abdominal surgery no more than 12 h following the image acquisition. Surgical findings were based on a visual and/or palpatory assessment of the bowel, and findings as stated in the surgical report were used as reference. Second look assessments were also included if available. Included patients were considered as independent samples.

Patients referred between October 2020 and August, 2022 were screened for inclusion. The same patient could be included multiple times if inclusion criteria were met. For these patients, statistical independence was assumed. Patient scans were excluded if image quality was severely affected by motion- or metal artifacts. Patient demographics and surgical findings were obtained through the regional electronic journal, Sundhedsplatformen (Epic Systems, Madison, Wisconsin, US). Dose reports including CT dose index volume (CTDIvol) and dose length product (DLP) were obtained within the local picture archiving and communication system (PACS).

All patients were scanned in a second-generation 256-slice CT (Revolution CT; GE Healthcare, Chicago, IL, USA). For the conventional non-contrast and arterial phase scan a kVp of 120 was used with a tube current between 80 and 600 mA (SmartmA) and a noise index of 13 and 14, respectively. For the venous phase acquisition, fast kV-switching (80/140 kV) was applied using Gemstone Spectral Imaging (GSI) Assist (GE Healthcare, Chicago, IL, USA) with a noise index of 14. Rotation speed, pitch, and beam width for all phases were set to 0.5 s, 0.992, and 80 mm, respectively.

Omnipaque 350 mg/mL (Iohexol, GE Healthcare, Chicago, IL, USA) was administered through an 18-gauge plastic cannula in the antecubital vein. Contrast volume was 1.4 mL/kg at a flow rate of 4 mL/s. The arterial and venous phase delay were set to 7 and 45 s following HU threshold triggering within the aorta at 100 HU.

Images were reconstructed using an adaptive statistical iterative reconstruction (ASIR-V) algorithm with a blending factor of 60% [16]. A standard abdominal kernel was applied. The data file for the venous phase scan as well as 0.625 mm sliced images for all scan phases were transferred to the advanced workstation (AW) server for image analysis. DECT images were reconstructed at 74 kiloelectron volt (keV) to resemble 120 kVp images.

Images were evaluated by four observers, two third year residents and two board-certified radiologists with 7 and > 20 years of clinical experience, respectively. Images were assessed on the AW server to blind observers from the list of previous examinations, referral text, radiology report and markers set by the reporting radiologist. All observers evaluated the images independently over two rounds. A gap of 6 weeks was set between the assessment of the last patient in the first round and first patient in the second round to blind observers from their previous assessments.

Round 1 included conventional 120 kVp-like images (venous phase and conventional non-contrast/arterial phase if available) and round 2 included iodine density reconstructions as well as the 120 kVp-like reconstructions presented in the first round. Each observer was given a patient list, which was intra- and inter-reader randomized by simple randomization across the two assessment rounds. A custom workflow within the AW server was setup for the second assessment round. This workflow included a virtual unenhanced (VUE) image overlayed by an iodine (water) map with a rainbow color ramp allowing the observer to freely overlay the iodine density map on top of the VUE image (Fig. 1). The color ramp provides a semiquantitative measure for the iodine concentration, which by default was set to − 12 to 55 \(\mathrm/\mathrm\)3 based on previous experience at our institution. Observers were briefly (1 h) introduced to the AW server and how to navigate and adjust slice thickness, color ramp settings, window level, and multi-planar reconstruction prior to the assessment rounds.

Axial image of the abdomen with an iodine overlay (colored map) on top of a virtual non-contrast image (underlying gray scale image). The color ramp ranging from − 12 to 55 (shown to the far left) represents the varying iodine concentrations in \(\mathrm/\mathrm\)3

For each assessment round, observers were asked whether they suspected ABI to be present or not. If ABI was suspected, the observers were asked to register the affected region(s) defined by the following list: duodenum, jejunum, ileum, small bowel, caecum, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum. Furthermore, observers were asked to rate their diagnostic confidence on a five-point Likert scale (1 = ABI cannot be excluded, 3 = moderate diagnostic confidence for ABI, 5 = high diagnostic confidence for ABI) similar to a previous study by Lourenco et al. [12, 17]. This confidence rating was only registered for patients in which ABI was suspected.

Sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV) were calculated for each observer with and without DECT reconstructions. Furthermore, McNemar’s test was used to assess whether a significant difference was observed between the two assessment rounds, while Wilcoxon signed-rank test was used to evaluate changes in confidence rating for the same observer. A significance level of 0.05 was applied.

Inter-observer agreement was evaluated using kappa statistics (< 0 = poor; 0.0–0.20 = slight agreement; 0.21–0.40 = fair agreement; 0.41–0.60 = moderate agreement; 0.61–0.80 = substantial agreement; and 0.81–1.0 = almost perfect agreement) [18]. Statistical analyses were performed in R version 4.0.1 (R Foundation for Statistical Computing, Vienna, Austria) with RStudio version 1.2.1093 (RStudio, Boston, MA, USA).