A foundation model for acute abdomen diagnosis stratification and triage on noncontrast computed tomography Authors: Chao Zhu, Ruipeng Zhang, Xinyu Song, Kai Shang, Yu Lu, Wenjuan Wu, Jian Ma, Yixiao Tang, Zhongzheng Cao, Li Shen, Jianyong Wei, Lisong Dai, Ping Wang, Dan Wang, Xiaoer Wei, Qing Lu, Lei Zhang, Tianle Wang, Yuehua Li Published: 2026 ABSTRACT Accurate and timely diagnosis of acute abdominal emergencies remains challenging. Non-contrast computed tomography (NCCT) is often used as an initial imaging modality because of atypical presentations, contraindications to contrast, or resource constraints. Here we present AbdomenNet, a multi-task AI system built on a self-supervised foundation model that detects 11 acute abdominal conditions and performs three risk-stratification subtasks from NCCT images. AbdomenNet is pre-trained on 103,989 NCCT examinations and fine-tuned on 5816 annotated cases. We assess generalizability in 2528 patients from three independent external cohorts, evaluate radiologist performance in a multi-reader multi-case crossover study, and estimate workflow impact using retrospective reconstruction. In external validation, AbdomenNet achieves a macro-average AUROC of 0.919 for five emergent conditions. AI assistance increases radiologists’ mean AUROC from 0.812 to 0.924 and reduces median reading time by 52.5 seconds per case. Workflow reconstruction indicates that AI-driven prioritization could shorten median report turnaround time by 37 minutes. Introduction Acute abdomen is among the most formidable challenges in emergency medicine, where the immediacy and accuracy of diagnosis directly affect patient outcomes1. This condition encompasses a wide spectrum of etiologies, while its clinical presentations are often non‑specific, making early and precise differential diagnosis difficult2. For a spectrum of highly time‑sensitive and life‑threatening conditions such as acute abdominal aortic dissection, any delay in diagnosis may result in catastrophic outcomes3. Among patients with acute abdomen undergoing abdominal surgery, the overall mortality rate can reach 12% and increases with each hour of delay4. While contrast-enhanced computed tomography (CECT) and computed tomography angiography (CTA) remain guideline-recommended modalities for the definitive diagnosis of many acute abdominal emergencies5, immediate access may be limited by patient-specific contraindications, such as renal dysfunction or severe contrast allergy, as well as by logistical delays in establishing venous access and resource constraints in overburdened emergency departments6. In China and some resource-constrained emergency environments, considerations of cost-effectiveness, resource allocation, workflow efficiency, and avoidance of unnecessary contrast-enhanced phases have led to a common clinical pathway in which non-contrast computed tomography (NCCT) is used as a rapid and readily accessible initial assessment tool for suspected acute abdomen, with CECT or CTA reserved for patients with high radiologic suspicion or elevated clinical risk7,8. Nationwide survey data indicate that between January and June 2022, tertiary hospitals in China performed a median of 16,512 emergency computed tomography (CT) examinations across all body regions (chest, head, abdomen, and others), corresponding to approximately 92 cases per day; abdominal CT represents a subset of this emergency imaging workload9. In this context, NCCT is often used to support the initial evaluation and triage of acute abdominal emergencies in China, despite its limited sensitivity compared with CECT6. Although NCCT is not the definitive imaging modality for many acute abdominal conditions, it often serves as the initial imaging examination used to inform diagnostic and triage decisions. This poses substantial diagnostic challenges even for experienced radiologists and clinicians. In high-volume emergency settings, this diagnostic limitation, compounded by long reading queues, can undermine the speed advantage of NCCT and delay recognition of atypical yet life-threatening presentations10. Enhancing the diagnostic performance of NCCT is an urgent priority in high-volume and resource-limited emergency settings, essential for reducing diagnostic delays and ensuring timely, equitable care across diverse healthcare systems. Artificial intelligence (AI) offers a promising avenue to augment radiologist performance in this high-stakes environment. NCCT-based AI applications have achieved high diagnostic accuracy in identifying specific acute emergencies, such as acute cholecystitis and acute aortic dissection11,12. However, existing studies, largely confined to single-condition detection, lack comprehensive evaluation of their performance and clinical applicability in complex, multi-condition emergency settings, with limited integration of robust, surgery-oriented risk stratification and quantitative assessment of workflow optimization13,14. Self-supervised visual foundation models15–17, such as DINOv218, learn generalizable representations from large-scale unlabeled datasets and offer a viable path to creating a comprehensive NCCT model for clinical decision support19,20, yet their utility for acute abdomen diagnosis, risk stratification, and emergency triage remains systematically unvalidated. We developed AbdomenNet, a multi-task AI system leveraging a self-supervised foundation model to analyze NCCT. The system detects 11 acute abdominal conditions and provides clinically actionable risk stratification for surgical emergencies. We rigorously validated its diagnostic performance, clinical utility, and potential workflow impact through multicenter external validation, a human-AI reader study, and retrospective triage workflow reconstruction. These findings demonstrate AbdomenNet’s potential translational value for enhancing detection, stratification, and triage in acute abdomen care. Results Patient characteristics A total of 103,989 abdominal NCCT examinations from 66,975 patients (mean age, 56 ± 21 years) were used for self-supervised pre-training. The internal fine-tuning cohort comprised 5,816 patients with acute abdomen from Center 1 (mean age, 54 ± 19 years; Fig. 1). External validation included 2528 patients from three independent tertiary medical centers (mean age, 53 ± 17 years). Detailed demographic and disease characteristics of the fine-tuning and external validation cohorts are summarized in Table 1. The overall AbdomenNet workflow and architecture are shown in Fig. 2.Fig. 1Model development and validation workflow.The model was developed, validated, and internally tested using the internal dataset, and its diagnostic performance and generalizability were subsequently evaluated across three independent external test cohorts. Arrows indicate patient selection and dataset partitioning. NCCT non-contrast computed tomography, n number of patients.Fig. 2Design of AbdomenNet.A Schematic diagram of the AbdomenNet system. Diagnosing acute abdominal pain is challenging due to non-specific symptoms. Non-contrast CT (NCCT) is often used for initial triage, but the lack of contrast makes subtle critical findings difficult to detect. AbdomenNet was designed to rapidly identify emergent conditions from NCCT scans and assist clinical decision-making by reprioritizing high-risk cases. B Distribution of the internal fine-tuning cohort. The 11 acute abdominal conditions are categorized by anatomical location, with each region indicating the disease name and corresponding patient count. Diagnostic conditions were not mutually exclusive; therefore, counts across conditions may exceed the total number of patients. C Development and validation of the AI system. The AI system was developed using an internal dataset and its generalization performance was subsequently evaluated across three independent external cohorts. D Pre-training and fine-tuning stages of AbdomenNet. For a given CT volume, AbdomenNet first encodes each slice to extract two-dimensional features, which are then fed into a classifier consisting of a three‑layer Transformer decoder. This architecture captures global inter-slice dependencies to perform scan-level prediction. Red icons patients with emergent conditions, blue icons patients without emergent conditions or remaining patients in the reporting queue, AI artificial intelligence, CT computed tomography, NCCT non-contrast computed tomography, ViT Vision Transformer, [CLS] token classification token, n number of patients.Table 1The demographic and diagnostic characteristics across included datasetsModel developmentExternal validationCharacteristicTraining setValidation setTesting setExternal Set1External Set2External Set3No. of patients41705441102878837813Age (y)54 ± 1955 ± 1955 ± 1952 ± 1755 ± 1753 ± 18Recorded sex Female1604/4170 (38.5)214/544 (39.3)407/1102 (36.9)335/878 (38.1)302/837 (36.1)322/813 (39.6)Male2566/4170 (61.5)330/544 (60.7)695/1102 (63.1)543/878 (61.9)535/837 (63.9)491/813 (60.4)Acute abdomen Acute abdominal aortic dissection233/4170 (5.6)28/544 (5.1)56/1102 (5.1)24/878 (2.7)23/837 (2.7)21/813 (2.6) Acute mesenteric ischemia91/4170 (2.2)13/544 (2.4)29/1102 (2.6)15/878 (1.7)14/837 (1.7)14/813 (1.7) Acute gastrointestinal perforation310/4170 (7.4)44/544 (8.1)86/1102 (7.8)66/878 (7.5)59/837 (7.0)52/813 (6.4) Acute intestinal obstruction542/4170 (13.0)70/544 (12.9)128/1102 (11.6)94/878 (10.7)96/837 (11.5)101/813 (12.4) Acute visceral injury282/4170 (6.8)44/544 (8.1)85/1102 (7.7)33/878 (3.8)31/837 (3.7)30/813 (3.7) Acute cholecystitis585/4170 (14.0)79/544 (14.5)162/1102 (14.7)92/878 (10.5)93/837 (11.1)98/813 (12.1) Acute choledocholithiasis298/4170 (7.1)40/544 (7.4)86/1102 (7.8)50/878 (5.7)46/837 (5.5)40/813 (4.9) Acute pancreatitis406/4170 (9.7)45/544 (8.3)108/1102 (9.8)83/878 (9.5)79/837 (9.4)86/813 (10.6) Acute appendicitis491/4170 (11.8)62/544 (11.4)121/1102 (11.0)103/878 (11.7)99/837 (11.8)94/813 (11.6) Acute diverticulitis350/4170 (8.4)44/544 (8.1)91/1102 (8.3)60/878 (6.8)58/837 (6.9)54/813 (6.6) Acute ureteral stone434/4170 (10.4)58/544 (10.7)110/1102 (10.0)99/878 (11.3)93/837 (11.1)88/813 (10.8)Control group432/4170 (10.4)61/544 (11.2)125/1102 (11.3)195/878 (22.2)189/837 (22.6)185/813 (22.8)Data are numbers of patients with percentages in parentheses unless otherwise indicated. Percentages were calculated using the total number of patients in each dataset. Diagnostic conditions were not mutually exclusive and therefore do not sum to 100%. Age is mean ± SD. Recorded sex was extracted from electronic medical records.SD standard deviation. Model development and validation workflow. The model was developed, validated, and internally tested using the internal dataset, and its diagnostic performance and generalizability were subsequently evaluated across three independent external test cohorts. Arrows indicate patient selection and dataset partitioning. NCCT non-contrast computed tomography, n number of patients. Design of AbdomenNet. A Schematic diagram of the AbdomenNet system. Diagnosing acute abdominal pain is challenging due to non-specific symptoms. Non-contrast CT (NCCT) is often used for initial triage, but the lack of contrast makes subtle critical findings difficult to detect. AbdomenNet was designed to rapidly identify emergent conditions from NCCT scans and assist clinical decision-making by reprioritizing high-risk cases. B Distribution of the internal fine-tuning cohort. The 11 acute abdominal conditions are categorized by anatomical location, with each region indicating the disease name and corresponding patient count. Diagnostic conditions were not mutually exclusive; therefore, counts across conditions may exceed the total number of patients. C Development and validation of the AI system. The AI system was developed using an internal dataset and its generalization performance was subsequently evaluated across three independent external cohorts. D Pre-training and fine-tuning stages of AbdomenNet. For a given CT volume, AbdomenNet first encodes each slice to extract two-dimensional features, which are then fed into a classifier consisting of a three‑layer Transformer decoder. This architecture captures global inter-slice dependencies to perform scan-level prediction. Red icons patients with emergent conditions, blue icons patients without emergent conditions or remaining patients in the reporting queue, AI artificial intelligence, CT computed tomography, NCCT non-contrast computed tomography, ViT Vision Transformer, [CLS] token classification token, n number of patients. Table 1The demographic and diagnostic characteristics across included datasetsModel developmentExternal validationCharacteristicTraining setValidation setTesting setExternal Set1External Set2External Set3No. of patients41705441102878837813Age (y)54 ± 1955 ± 1955 ± 1952 ± 1755 ± 1753 ± 18Recorded sex Female1604/4170 (38.5)214/544 (39.3)407/1102 (36.9)335/878 (38.1)302/837 (36.1)322/813 (39.6)Male2566/4170 (61.5)330/544 (60.7)695/1102 (63.1)543/878 (61.9)535/837 (63.9)491/813 (60.4)Acute abdomen Acute abdominal aortic dissection233/4170 (5.6)28/544 (5.1)56/1102 (5.1)24/878 (2.7)23/837 (2.7)21/813 (2.6) Acute mesenteric ischemia91/4170 (2.2)13/544 (2.4)29/1102 (2.6)15/878 (1.7)14/837 (1.7)14/813 (1.7) Acute gastrointestinal perforation310/4170 (7.4)44/544 (8.1)86/1102 (7.8)66/878 (7.5)59/837 (7.0)52/813 (6.4) Acute intestinal obstruction542/4170 (13.0)70/544 (12.9)128/1102 (11.6)94/878 (10.7)96/837 (11.5)101/813 (12.4) Acute visceral injury282/4170 (6.8)44/544 (8.1)85/1102 (7.7)33/878 (3.8)31/837 (3.7)30/813 (3.7) Acute cholecystitis585/4170 (14.0)79/544 (14.5)162/1102 (14.7)92/878 (10.5)93/837 (11.1)98/813 (12.1) Acute choledocholithiasis298/4170 (7.1)40/544 (7.4)86/1102 (7.8)50/878 (5.7)46/837 (5.5)40/813 (4.9) Acute pancreatitis406/4170 (9.7)45/544 (8.3)108/1102 (9.8)83/878 (9.5)79/837 (9.4)86/813 (10.6) Acute appendicitis491/4170 (11.8)62/544 (11.4)121/1102 (11.0)103/878 (11.7)99/837 (11.8)94/813 (11.6) Acute diverticulitis350/4170 (8.4)44/544 (8.1)91/1102 (8.3)60/878 (6.8)58/837 (6.9)54/813 (6.6) Acute ureteral stone434/4170 (10.4)58/544 (10.7)110/1102 (10.0)99/878 (11.3)93/837 (11.1)88/813 (10.8)Control group432/4170 (10.4)61/544 (11.2)125/1102 (11.3)195/878 (22.2)189/837 (22.6)185/813 (22.8)Data are numbers of patients with percentages in parentheses unless otherwise indicated. Percentages were calculated using the total number of patients in each dataset. Diagnostic conditions were not mutually exclusive and therefore do not sum to 100%. Age is mean ± SD. Recorded sex was extracted from electronic medical records.SD standard deviation. The demographic and diagnostic characteristics across included datasets Data are numbers of patients with percentages in parentheses unless otherwise indicated. Percentages were calculated using the total number of patients in each dataset. Diagnostic conditions were not mutually exclusive and therefore do not sum to 100%. Age is mean ± SD. Recorded sex was extracted from electronic medical records. SD standard deviation. Performance of AbdomenNet on primary diagnostic tasks In the internal test set, AbdomenNet achieved a macro-average area under the receiver operating characteristic curve (AUROC) of 0.908 (95% confidence interval [CI], 0.887–0.933) across the 11 primary diagnostic tasks (Table 2, Fig. 3A). The corresponding ROC curves and F1-score for all 11 primary diagnostic tasks are shown in Fig. S1 and Table S1. For the five emergent conditions, the macro-average AUROC reached 0.937 (95% CI, 0.903–0.971).Fig. 3Diagnostic performance of AbdomenNet.A AUROC comparison of AbdomenNet with DINOv2, Merlin, CT-FM, TAP-CT, Curia, and radiomics across the 11 primary diagnostic tasks on the internal test set. B Performance of the 11 primary diagnostic tasks in terms of macro-average AUROC, sensitivity, specificity, and accuracy across the internal test set and three external test sets. Bars indicate macro-average values across the 11 primary diagnostic tasks, dots indicate condition-level values and are horizontally jittered for visibility, and error bars indicate 95% confidence intervals calculated using 10,000 bootstrap iterations. For each dataset and performance metric, n = 11 primary diagnostic tasks. C AUROC performance of three risk-stratification subtasks across the internal test set and three external test sets. Points indicate AUROC estimates, and error bars indicate 95% confidence intervals calculated using 10,000 bootstrap iterations. The unit of analysis was the patient, and all samples were independent patient-level observations. Sample sizes for the internal set and external sets 1–3 were n = 82, 33, 31, and 30 patients for identifying high-grade visceral injury, respectively; n = 96, 20, 21, and 23 patients for identifying strangulated intestinal obstruction, respectively; and n = 68, 50, 47, and 42 patients for identifying upper gastrointestinal perforation, respectively. D, E Representative NCCT images and corresponding Grad-CAM heatmaps for abdominal aortic dissection and acute appendicitis, respectively. The paired contrast-enhanced CT image in (D) is provided for illustrative comparison. Cyan boxes and connector lines indicate magnified regions of interest, and warmer colors in Grad-CAM heatmaps indicate higher model activation. AUROC area under the receiver operating characteristic curve, CI confidence interval, CT computed tomography, CT-FM CT foundation model, Grad-CAM gradient-weighted class activation mapping, NCCT non-contrast computed tomography, TAP-CT task-agnostic pre-training for CT. Colors in (A) indicate different models as shown in the in-panel legend; colors in (B, C) indicate datasets.Table 2Diagnostic performance of AbdomenNet for emergent and urgent acute abdominal conditionsDiagnostic conditionAUROCSensitivitySpecificityAccuracyEmergent conditions Acute abdominal aortic dissection0.987 (0.978, 0.994)0.965 (0.907, 1.000)0.955 (0.943, 0.968)0.956 (0.944, 0.967) Acute mesenteric ischemia0.896 (0.866, 0.923)0.838 (0.800, 0.933)0.827 (0.787, 0.865)0.827 (0.789, 0.864) Acute gastrointestinal perforation0.921 (0.897, 0.943)0.896 (0.828, 0.955)0.782 (0.756, 0.808)0.791 (0.766, 0.816) Acute intestinal obstruction0.902 (0.879, 0.922)0.858 (0.850, 0.875)0.786 (0.729, 0.849)0.795 (0.744, 0.850) Acute visceral injury0.981 (0.969, 0.990)0.977 (0.939, 1.000)0.945 (0.931, 0.958)0.947 (0.935, 0.960)Macro-average emergent conditions0.937 (0.903, 0.971)0.907 (0.858, 0.956)0.859 (0.793, 0.925)0.863 (0.800, 0.927)Urgent conditions Acute cholecystitis0.885 (0.862, 0.908)0.841 (0.801, 0.891)0.766 (0.736, 0.794)0.777 (0.749, 0.802) Acute choledocholithiasis0.887 (0.857, 0.916)0.843 (0.800, 0.910)0.763 (0.714, 0.791)0.769 (0.724, 0.796) Acute pancreatitis0.875 (0.848, 0.902)0.871 (0.811, 0.933)0.722 (0.693, 0.749)0.736 (0.710, 0.762) Acute appendicitis0.862 (0.838, 0.884)0.934 (0.887, 0.975)0.718 (0.690, 0.746)0.742 (0.715, 0.768) Acute diverticulitis0.908 (0.875, 0.937)0.842 (0.800, 0.907)0.712 (0.680, 0.742)0.723 (0.692, 0.751) Acute ureteral stone0.880 (0.852, 0.907)0.808 (0.800, 0.828)0.762 (0.708, 0.818)0.767 (0.719, 0.818)Macro-average urgent conditions0.883 (0.872, 0.893)0.857 (0.821, 0.894)0.741 (0.727, 0.759)0.752 (0.729, 0.775)Overall macro-average0.908 (0.887, 0.933)0.879 (0.856, 0.929)0.794 (0.722, 0.836)0.803 (0.740, 0.844)Data in parentheses are 95% confidence intervals. Macro-average values were calculated as unweighted averages across the corresponding diagnostic conditions.AUROC area under the receiver operating characteristic curve. Diagnostic performance of AbdomenNet. A AUROC comparison of AbdomenNet with DINOv2, Merlin, CT-FM, TAP-CT, Curia, and radiomics across the 11 primary diagnostic tasks on the internal test set. B Performance of the 11 primary diagnostic tasks in terms of macro-average AUROC, sensitivity, specificity, and accuracy across the internal test set and three external test sets. Bars indicate macro-average values across the 11 primary diagnostic tasks, dots indicate condition-level values and are horizontally jittered for visibility, and error bars indicate 95% confidence intervals calculated using 10,000 bootstrap iterations. For each dataset and performance metric, n = 11 primary diagnostic tasks. C AUROC performance of three risk-stratification subtasks across the internal test set and three external test sets. Points indicate AUROC estimates, and error bars indicate 95% confidence intervals calculated using 10,000 bootstrap iterations. The unit of analysis was the patient, and all samples were independent patient-level observations. Sample sizes for the internal set and external sets 1–3 were n = 82, 33, 31, and 30 patients for identifying high-grade visceral injury, respectively; n = 96, 20, 21, and 23 patients for identifying strangulated intestinal obstruction, respectively; and n = 68, 50, 47, and 42 patients for identifying upper gastrointestinal perforation, respectively. D, E Representative NCCT images and corresponding Grad-CAM heatmaps for abdominal aortic dissection and acute appendicitis, respectively. The paired contrast-enhanced CT image in (D) is provided for illustrative comparison. Cyan boxes and connector lines indicate magnified regions of interest, and warmer colors in Grad-CAM heatmaps indicate higher model activation. AUROC area under the receiver operating characteristic curve, CI confidence interval, CT computed tomography, CT-FM CT foundation model, Grad-CAM gradient-weighted class activation mapping, NCCT non-contrast computed tomography, TAP-CT task-agnostic pre-training for CT. Colors in (A) indicate different models as shown in the in-panel legend; colors in (B, C) indicate datasets. Table 2Diagnostic performance of AbdomenNet for emergent and urgent acute abdominal conditionsDiagnostic conditionAUROCSensitivitySpecificityAccuracyEmergent conditions Acute abdominal aortic dissection0.987 (0.978, 0.994)0.965 (0.907, 1.000)0.955 (0.943, 0.968)0.956 (0.944, 0.967) Acute mesenteric ischemia0.896 (0.866, 0.923)0.838 (0.800, 0.933)0.827 (0.787, 0.865)0.827 (0.789, 0.864) Acute gastrointestinal perforation0.921 (0.897, 0.943)0.896 (0.828, 0.955)0.782 (0.756, 0.808)0.791 (0.766, 0.816) Acute intestinal obstruction0.902 (0.879, 0.922)0.858 (0.850, 0.875)0.786 (0.729, 0.849)0.795 (0.744, 0.850) Acute visceral injury0.981 (0.969, 0.990)0.977 (0.939, 1.000)0.945 (0.931, 0.958)0.947 (0.935, 0.960)Macro-average emergent conditions0.937 (0.903, 0.971)0.907 (0.858, 0.956)0.859 (0.793, 0.925)0.863 (0.800, 0.927)Urgent conditions Acute cholecystitis0.885 (0.862, 0.908)0.841 (0.801, 0.891)0.766 (0.736, 0.794)0.777 (0.749, 0.802) Acute choledocholithiasis0.887 (0.857, 0.916)0.843 (0.800, 0.910)0.763 (0.714, 0.791)0.769 (0.724, 0.796) Acute pancreatitis0.875 (0.848, 0.902)0.871 (0.811, 0.933)0.722 (0.693, 0.749)0.736 (0.710, 0.762) Acute appendicitis0.862 (0.838, 0.884)0.934 (0.887, 0.975)0.718 (0.690, 0.746)0.742 (0.715, 0.768) Acute diverticulitis0.908 (0.875, 0.937)0.842 (0.800, 0.907)0.712 (0.680, 0.742)0.723 (0.692, 0.751) Acute ureteral stone0.880 (0.852, 0.907)0.808 (0.800, 0.828)0.762 (0.708, 0.818)0.767 (0.719, 0.818)Macro-average urgent conditions0.883 (0.872, 0.893)0.857 (0.821, 0.894)0.741 (0.727, 0.759)0.752 (0.729, 0.775)Overall macro-average0.908 (0.887, 0.933)0.879 (0.856, 0.929)0.794 (0.722, 0.836)0.803 (0.740, 0.844)Data in parentheses are 95% confidence intervals. Macro-average values were calculated as unweighted averages across the corresponding diagnostic conditions.AUROC area under the receiver operating characteristic curve. Diagnostic performance of AbdomenNet for emergent and urgent acute abdominal conditions Data in parentheses are 95% confidence intervals. Macro-average values were calculated as unweighted averages across the corresponding diagnostic conditions. AUROC area under the receiver operating characteristic curve. As illustrated in Fig. 3D, E, the resulting heatmaps generally highlighted clinically relevant pathological regions corresponding to the ground truth diagnosis. For instance, in a case of acute abdominal aortic dissection, the model’s attention correctly localized to the intimal flap and true lumen of the aortic wall, while in a case of acute appendicitis, it focused on the inflamed and thickened appendix. The expanded Gradient-weighted Class Activation Mapping (Grad-CAM) visualization included 33 representative cases across the 11 conditions, comprising two well-localized examples and one discordant example per condition (Fig. S2). Discordant examples were mainly observed in early or subtle presentations, including mesenteric ischemia with subtle or segmental bowel wall changes and acute abdominal aortic dissection with limited visibility of the intimal flap on NCCT. Quantitative localization analysis using TotalSegmentator-derived organ-level regions of interest (ROIs) is summarized in Table S1. Ablation and benchmark comparisons Domain-specific DINOv2 pre-training (AbdomenNet) achieved the highest macro-average AUROC on the internal test set, 0.908 (95% CI, 0.887–0.933), exceeding training from scratch, 0.714 (95% CI, 0.698–0.732), general-domain DINOv2 initialization, 0.822 (95% CI, 0.781–0.865), and alternative self-supervised objectives including SimCLR, 0.794 (95% CI, 0.779–0.808), and MAE, 0.750 (95% CI, 0.733–0.767) (Table S2). Performance increased consistently with pre-training data volume, with no plateau observed within the evaluated range (Fig. S3). AbdomenNet also achieved higher performance than Merlin under both evaluated configurations: publicly released pre-trained weights, 0.678 (95% CI: 0.634–0.730), and re-pre-training on our NCCT corpus with paired radiology reports, 0.833 (95% CI, 0.818–0.846). AbdomenNet further outperformed three publicly available general-purpose CT foundation models, CT-FM, 0.823 (95% CI, 0.808–0.836), TAP-CT, 0.845 (95% CI, 0.834–0.856), and Curia, 0.874 (95% CI, 0.863–0.884), as well as the radiomics pipeline, 0.731 (95% CI, 0.675–0.782), with the 95% CI of AbdomenNet not overlapping those of any comparator (Table S2). Performance of AbdomenNet in risk stratification subtasks For high-grade visceral injury classification in the overall subtask cohort (n = 411), defined as American Association for the Surgery of Trauma (AAST) grade ≥III, AbdomenNet achieved an AUROC of 0.896 (95% CI, 0.816–0.961) on the internal test set, using paired emergency abdominal CECT as the reference standard. For identifying strangulated intestinal obstruction in the overall subtask cohort (n = 202), using intraoperative findings or postoperative pathology as the reference standard, the AUROC was 0.912 (95% CI, 0.799–0.986) on the internal test set. For identifying upper gastrointestinal perforation in the overall subtask cohort (n = 341) among patients with surgically confirmed perforation sites, the AUROC was 0.873 (95% CI, 0.769–0.959) on the internal test set. The performance across these subtasks on the internal test set is presented in Fig. 3C. Generalization performance across external centers To assess generalization capability across different centers, AbdomenNet was validated on three independent external cohorts. Across all 11 primary diagnostic tasks, the macro-average AUROCs for the three external centers were 0.898 (95% CI, 0.877–0.923), 0.889 (95% CI, 0.872–0.912), and 0.896 (95% CI, 0.875–0.920), respectively (Fig. 3B). Detailed condition-specific sensitivity and specificity results for each external cohort are summarized in Table S3. Across the five emergent-condition categories, the macro-average AUROC was 0.919 (95% CI, 0.889–0.948). For the external risk-stratification subtasks, AbdomenNet achieved AUROCs of 0.888 (95% CI, 0.817–0.949) for identifying high-grade visceral injury (AAST grade ≥III), 0.889 (95% CI, 0.804–0.957) for identifying strangulated intestinal obstruction, and 0.875 (95% CI, 0.811–0.930) for identifying upper gastrointestinal perforation (Fig. S4). Radiologist performance with and without AI assistance An overview of the human-AI comparative experiment is presented in Fig. 4A. Regarding diagnostic accuracy, a generalized linear mixed-effects (GLMM) model analysis revealed a strong association between AI assistance and improved diagnostic correctness after controlling for the random effects of radiologists and cases (Odds Ratio [OR] = 2.29, 95% CI, 2.11–2.49, P < 0.001). In a disease-specific analysis, AI assistance was significantly associated with higher diagnostic accuracy for each of the 11 acute conditions (Fig. 4B). For the diagnosis of the five emergent conditions, the AUROC for senior radiologists significantly increased from 0.876 to 0.968 with AI assistance (ΔAUROC = 0.092, P < 0.001) (Table 3). The performance of radiologists with varying seniority on the six urgent conditions with AI assistance is detailed in Fig. 4D, E. Furthermore, AI assistance significantly reduced the median reading time for radiologists from 197.5 seconds [interquartile range (IQR) 75.0] to 145.0 seconds [IQR 40.0] (P < 0.001) (Fig. 4C).Fig. 4Overview of the human-AI comparative experiment.A Multi-reader, multi-case crossover design. In Reading Session 1, four radiologists reviewed Set A with AI assistance and Set B without AI assistance; after a 3-month washout period, they reviewed Set A without AI assistance and Set B with AI assistance in Reading Session 2. B Disease-specific association between AI assistance and diagnostic accuracy across the 11 acute abdominal conditions. Dots represent odds ratio point estimates, and horizontal error bars indicate 95% confidence intervals. The dashed vertical line indicates an odds ratio of 1.0, and the gray shaded region indicates odds ratios greater than 1.0, favoring AI assistance. C Reading times of senior and junior radiologists with and without AI assistance. Box plots show reader-case-level reading-time distributions; the center line indicates the median, box bounds indicate the first and third quartiles, and whiskers extend to the minimum and maximum observed values. The unit of analysis was the reader-case observation, with n = 760 observations per group. Statistical comparisons used paired two-sided Wilcoxon signed-rank tests; no adjustment for multiple comparisons was applied. Exact two-sided P values for senior and junior radiologists were 5.190 × 10−121 and 4.587 × 10−126, respectively. D, E Comparison of AUROC and sensitivity, respectively, with and without AI assistance. Bars indicate mean radiologist-level values, dots indicate individual radiologists, and numeric labels indicate group means. Solid bars indicate emergent conditions, and hatched bars indicate urgent conditions. The unit of analysis was the radiologist, with n = 2 radiologists per bar. Statistical comparisons used paired two-sided DeLong tests for (D) and two-sided exact McNemar tests for (E); no adjustment for multiple comparisons was applied. Exact two-sided P values for senior-emergent, senior-urgent, junior-emergent, and junior-urgent comparisons were 5.469 × 10−17, 8.061 × 10−11, 2.089 × 10−27, and 3.098 × 10−50, respectively, in (D) and 5.996 × 10−17, 3.638 × 10−12, 8.758 × 10−23, and 5.781 × 10−35, respectively, in (E). *** indicates P < 0.001. AI artificial intelligence, AUROC area under the receiver operating characteristic curve, CI confidence interval, MRMC multi-reader multi-case, OR odds ratio. Pink without AI assistance; blue with AI assistance.Table 3Radiologist performance with and without AbdomenNet assistanceParameterWithout AIWith AIDifferenceP valueSenior radiologists AUROC0.876 (0.852, 0.898)0.968 (0.954, 0.979)0.092<0.001 (DeLong) AUPRC0.758 (0.712, 0.800)0.925 (0.897, 0.950)0.167<0.001 (Bootstrap) Sensitivity (%)75.4 (70.6, 79.7)93.8 (90.7, 95.9)18.4<0.001 (McNemar) Reading time (s)162.0 (147.0, 176.0)126.0 (113.0, 137.0)−36.0<0.001 (Wilcoxon)Junior radiologists AUROC0.735 (0.708, 0.762)0.874 (0.852, 0.896)0.139<0.001 (DeLong) AUPRC0.328 (0.280, 0.375)0.603 (0.551, 0.654)0.275<0.001 (Bootstrap) Sensitivity (%)51.2 (45.9, 56.5)77.2 (72.5, 81.4)26.0<0.001 (McNemar) Reading time (s)237.0 (222.0, 253.0)166.0 (155.0, 175.0)−71.0<0.001 (Wilcoxon)Data in parentheses are 95% confidence intervals, except for reading time, which is shown as median (Q1, Q3). Differences were calculated as values with AI assistance minus values without AI assistance. Diagnostic metrics were calculated on the 380-case MRMC set with n = 2 radiologists per seniority group; reading-time analyses included n = 760 reader-case observations per group. Two-sided DeLong, bootstrap, McNemar, and Wilcoxon signed-rank tests were used for AUROC, AUPRC, sensitivity, and reading time, respectively. No adjustment for multiple comparisons was applied. For the emergent-condition analyses shown in this table, exact two-sided P values for senior/junior radiologists were 5.469 × 10−17/2.089 × 10−27 for AUROC, <2.000 × 10−5/<2.000 × 10−5 for AUPRC, and 5.996 × 10−17/8.758 × 10−23 for sensitivity. Exact two-sided P values for reading time were 5.190 × 10−121/4.587 × 10−126 for senior/junior radiologists. AUPRC P values are reported at the two-sided resolution limit of 100,000 paired-bootstrap resamples.AI artificial intelligence, AUROC area under the receiver operating characteristic curve, AUPRC area under the precision-recall curve, MRMC multi-reader multi-case. Overview of the human-AI comparative experiment. A Multi-reader, multi-case crossover design. In Reading Session 1, four radiologists reviewed Set A with AI assistance and Set B without AI assistance; after a 3-month washout period, they reviewed Set A without AI assistance and Set B with AI assistance in Reading Session 2. B Disease-specific association between AI assistance and diagnostic accuracy across the 11 acute abdominal conditions. Dots represent odds ratio point estimates, and horizontal error bars indicate 95% confidence intervals. The dashed vertical line indicates an odds ratio of 1.0, and the gray shaded region indicates odds ratios greater than 1.0, favoring AI assistance. C Reading times of senior and junior radiologists with and without AI assistance. Box plots show reader-case-level reading-time distributions; the center line indicates the median, box bounds indicate the first and third quartiles, and whiskers extend to the minimum and maximum observed values. The unit of analysis was the reader-case observation, with n = 760 observations per group. Statistical comparisons used paired two-sided Wilcoxon signed-rank tests; no adjustment for multiple comparisons was applied. Exact two-sided P values for senior and junior radiologists were 5.190 × 10−121 and 4.587 × 10−126, respectively. D, E Comparison of AUROC and sensitivity, respectively, with and without AI assistance. Bars indicate mean radiologist-level values, dots indicate individual radiologists, and numeric labels indicate group means. Solid bars indicate emergent conditions, and hatched bars indicate urgent conditions. The unit of analysis was the radiologist, with n = 2 radiologists per bar. Statistical comparisons used paired two-sided DeLong tests for (D) and two-sided exact McNemar tests for (E); no adjustment for multiple comparisons was applied. Exact two-sided P values for senior-emergent, senior-urgent, junior-emergent, and junior-urgent comparisons were 5.469 × 10−17, 8.061 × 10−11, 2.089 × 10−27, and 3.098 × 10−50, respectively, in (D) and 5.996 × 10−17, 3.638 × 10−12, 8.758 × 10−23, and 5.781 × 10−35, respectively, in (E). *** indicates P < 0.001. AI artificial intelligence, AUROC area under the receiver operating characteristic curve, CI confidence interval, MRMC multi-reader multi-case, OR odds ratio. Pink without AI assistance; blue with AI assistance. Table 3Radiologist performance with and without AbdomenNet assistanceParameterWithout AIWith AIDifferenceP valueSenior radiologists AUROC0.876 (0.852, 0.898)0.968 (0.954, 0.979)0.092<0.001 (DeLong) AUPRC0.758 (0.712, 0.800)0.925 (0.897, 0.950)0.167<0.001 (Bootstrap) Sensitivity (%)75.4 (70.6, 79.7)93.8 (90.7, 95.9)18.4<0.001 (McNemar) Reading time (s)162.0 (147.0, 176.0)126.0 (113.0, 137.0)−36.0<0.001 (Wilcoxon)Junior radiologists AUROC0.735 (0.708, 0.762)0.874 (0.852, 0.896)0.139<0.001 (DeLong) AUPRC0.328 (0.280, 0.375)0.603 (0.551, 0.654)0.275<0.001 (Bootstrap) Sensitivity (%)51.2 (45.9, 56.5)77.2 (72.5, 81.4)26.0<0.001 (McNemar) Reading time (s)237.0 (222.0, 253.0)166.0 (155.0, 175.0)−71.0<0.001 (Wilcoxon)Data in parentheses are 95% confidence intervals, except for reading time, which is shown as median (Q1, Q3). Differences were calculated as values with AI assistance minus values without AI assistance. Diagnostic metrics were calculated on the 380-case MRMC set with n = 2 radiologists per seniority group; reading-time analyses included n = 760 reader-case observations per group. Two-sided DeLong, bootstrap, McNemar, and Wilcoxon signed-rank tests were used for AUROC, AUPRC, sensitivity, and reading time, respectively. No adjustment for multiple comparisons was applied. For the emergent-condition analyses shown in this table, exact two-sided P values for senior/junior radiologists were 5.469 × 10−17/2.089 × 10−27 for AUROC, <2.000 × 10−5/<2.000 × 10−5 for AUPRC, and 5.996 × 10−17/8.758 × 10−23 for sensitivity. Exact two-sided P values for reading time were 5.190 × 10−121/4.587 × 10−126 for senior/junior radiologists. AUPRC P values are reported at the two-sided resolution limit of 100,000 paired-bootstrap resamples.AI artificial intelligence, AUROC area under the receiver operating characteristic curve, AUPRC area under the precision-recall curve, MRMC multi-reader multi-case. Radiologist performance with and without AbdomenNet assistance Data in parentheses are 95% confidence intervals, except for reading time, which is shown as median (Q1, Q3). Differences were calculated as values with AI assistance minus values without AI assistance. Diagnostic metrics were calculated on the 380-case MRMC set with n = 2 radiologists per seniority group; reading-time analyses included n = 760 reader-case observations per group. Two-sided DeLong, bootstrap, McNemar, and Wilcoxon signed-rank tests were used for AUROC, AUPRC, sensitivity, and reading time, respectively. No adjustment for multiple comparisons was applied. For the emergent-condition analyses shown in this table, exact two-sided P values for senior/junior radiologists were 5.469 × 10−17/2.089 × 10−27 for AUROC, <2.000 × 10−5/<2.000 × 10−5 for AUPRC, and 5.996 × 10−17/8.758 × 10−23 for sensitivity. Exact two-sided P values for reading time were 5.190 × 10−121/4.587 × 10−126 for senior/junior radiologists. AUPRC P values are reported at the two-sided resolution limit of 100,000 paired-bootstrap resamples. AI artificial intelligence, AUROC area under the receiver operating characteristic curve, AUPRC area under the precision-recall curve, MRMC multi-reader multi-case. Workflow reconstruction and triage impact The schematic diagram of the emergency triage study is presented in Fig. 5A. Table 4 presents the performance of AbdomenNet on the internal test set at high-sensitivity operating thresholds selected on the internal validation set. For the seven critical conditions, Fig. S5 shows ROC curves with validation-selected operating points and the corresponding confusion matrices after applying these thresholds to the internal test set. Our retrospective workflow reconstruction was performed on a cohort of 18,398 emergency CT scans (including 1,761 abdominal CT scans). The median report turnaround time (TAT) for the emergent conditions group was significantly reduced by 37.0 min (12.3 min [IQR 5.3] post-triage vs. 49.3 min [IQR 40.5] pre-triage, P < 0.001). Crucially, this prioritization did not result in a statistically significant delay for the remaining cohort of emergency CT scans (57.9 min [IQR 46.1] post-triage vs. 57.1 min [IQR 45.7] pre-triage, P = 0.170). Under the triage thresholds calibrated to achieve sensitivity ≥0.95, the model flagged 767 of the 1761 abdominal NCCT cases as P1; these included 203 true emergent cases and 564 false positives, yielding a P1 flag rate of 4.2% relative to the total emergency CT queue (767/18,398). The absence of a significant delay in the remaining emergency CT cohort suggested no evidence of queue saturation from false-positive alerts in this retrospective reconstruction. Detailed information on queue lengths, TAT, and related metrics under the reconstructed AI triage workflow, stratified by daytime and nighttime shifts, is provided in Fig. 5 and Table S4.Fig. 5Retrospective workflow reconstruction of AI-prioritized emergency CT reporting.A Schematic of the retrospective workflow reconstruction. The historically observed first-come, first-served workflow was compared with a reconstructed AI-triage workflow based on real-world historical data. P1-flagged cases were prioritized using high-sensitivity thresholds selected for five emergent conditions and two risk-stratification subtasks. B–E TAT distributions stratified by daytime and nighttime shifts. Box plots show the median (center line), first and third quartiles (box bounds), and minimum and maximum observed values (whiskers). For condition-specific analyses, the unit of analysis was the abdominal NCCT examination; n values for SIO, AGP, HVI, AMI, and AAAD were 15, 45, 17, 14, and 19 in (B), and 13, 37, 16, 13, and 14 in (D), respectively. For group-level queue analyses, each examination contributed paired TAT values under both workflows; n = 110 and 13,050 for the emergent and remaining groups in (C), and n = 93 and 5145 in (E), respectively. No technical replicates were used. Statistical comparisons used paired two-sided Wilcoxon signed-rank tests; no adjustment for multiple comparisons was applied. Exact two-sided P values for SIO, AGP, HVI, AMI, and AAAD were 1.22 × 10−4, 1.14 × 10−13, 3.05 × 10−5, 2.44 × 10−4, and 7.63 × 10−6, respectively, in (B), and 2.44 × 10−4, 2.91 × 10−11, 6.10 × 10−5, 4.88 × 10−4, and 1.22 × 10−4, respectively, in (D). For the overall five-disease emergent group, exact two-sided P values were 4.93 × 10−32 for daytime (C) and 1.62 × 10−27 for nighttime (E); for the remaining group, the corresponding P values were 0.225 and 0.620, respectively. *** indicates P < 0.001. AI artificial intelligence, CT computed tomography, NCCT non-contrast CT, TAT report turnaround time, P1 emergent priority, SIO strangulated intestinal obstruction, AGP acute gastrointestinal perforation, HVI high-grade visceral injury, AMI acute mesenteric ischemia, AAAD acute abdominal aortic dissection. Pink, first-come, first-served workflow; blue, priority-based strategy.Table 4Performance of AbdomenNet on the internal test set at high-sensitivity operating thresholdsEmergent conditionsThresholdSensitivityAUROCAccuracySpecificityAcute abdominal aortic dissection0.5660.964 (0.907, 1.000)0.987 (0.978, 0.994)0.963 (0.952, 0.975)0.964 (0.952, 0.975)Acute mesenteric ischemia0.3230.966 (0.885, 1.000)0.896 (0.866, 0.923)0.805 (0.779, 0.828)0.801 (0.774, 0.823)Acute gastrointestinal perforation0.4380.954 (0.903, 1.000)0.921 (0.897, 0.943)0.752 (0.728, 0.777)0.735 (0.708, 0.762)Acute intestinal obstruction0.6150.953 (0.915, 0.985)0.902 (0.879, 0.922)0.740 (0.712, 0.765)0.711 (0.683, 0.738)Acute visceral injury0.6980.953 (0.901, 0.990)0.981 (0.969, 0.990)0.956 (0.944, 0.967)0.956 (0.942, 0.968)High-grade visceral injury0.3130.963 (0.911, 1.000)0.896 (0.816, 0.961)0.878 (0.805, 0.951)0.714 (0.538, 0.875)Strangulated intestinal obstruction0.4900.962 (0.889, 1.000)0.912 (0.799, 0.986)0.822 (0.625, 0.975)0.771 (0.500, 0.968)Data in parentheses are 95% confidence intervals. Operating thresholds indicate model-score cutoffs selected on the internal validation set to prioritize high sensitivity and fixed before evaluation on the internal test set.AUROC area under the receiver operating characteristic curve. Retrospective workflow reconstruction of AI-prioritized emergency CT reporting. A Schematic of the retrospective workflow reconstruction. The historically observed first-come, first-served workflow was compared with a reconstructed AI-triage workflow based on real-world historical data. P1-flagged cases were prioritized using high-sensitivity thresholds selected for five emergent conditions and two risk-stratification subtasks. B–E TAT distributions stratified by daytime and nighttime shifts. Box plots show the median (center line), first and third quartiles (box bounds), and minimum and maximum observed values (whiskers). For condition-specific analyses, the unit of analysis was the abdominal NCCT examination; n values for SIO, AGP, HVI, AMI, and AAAD were 15, 45, 17, 14, and 19 in (B), and 13, 37, 16, 13, and 14 in (D), respectively. For group-level queue analyses, each examination contributed paired TAT values under both workflows; n = 110 and 13,050 for the emergent and remaining groups in (C), and n = 93 and 5145 in (E), respectively. No technical replicates were used. Statistical comparisons used paired two-sided Wilcoxon signed-rank tests; no adjustment for multiple comparisons was applied. Exact two-sided P values for SIO, AGP, HVI, AMI, and AAAD were 1.22 × 10−4, 1.14 × 10−13, 3.05 × 10−5, 2.44 × 10−4, and 7.63 × 10−6, respectively, in (B), and 2.44 × 10−4, 2.91 × 10−11, 6.10 × 10−5, 4.88 × 10−4, and 1.22 × 10−4, respectively, in (D). For the overall five-disease emergent group, exact two-sided P values were 4.93 × 10−32 for daytime (C) and 1.62 × 10−27 for nighttime (E); for the remaining group, the corresponding P values were 0.225 and 0.620, respectively. *** indicates P < 0.001. AI artificial intelligence, CT computed tomography, NCCT non-contrast CT, TAT report turnaround time, P1 emergent priority, SIO strangulated intestinal obstruction, AGP acute gastrointestinal perforation, HVI high-grade visceral injury, AMI acute mesenteric ischemia, AAAD acute abdominal aortic dissection. Pink, first-come, first-served workflow; blue, priority-based strategy. Table 4Performance of AbdomenNet on the internal test set at high-sensitivity operating thresholdsEmergent conditionsThresholdSensitivityAUROCAccuracySpecificityAcute abdominal aortic dissection0.5660.964 (0.907, 1.000)0.987 (0.978, 0.994)0.963 (0.952, 0.975)0.964 (0.952, 0.975)Acute mesenteric ischemia0.3230.966 (0.885, 1.000)0.896 (0.866, 0.923)0.805 (0.779, 0.828)0.801 (0.774, 0.823)Acute gastrointestinal perforation0.4380.954 (0.903, 1.000)0.921 (0.897, 0.943)0.752 (0.728, 0.777)0.735 (0.708, 0.762)Acute intestinal obstruction0.6150.953 (0.915, 0.985)0.902 (0.879, 0.922)0.740 (0.712, 0.765)0.711 (0.683, 0.738)Acute visceral injury0.6980.953 (0.901, 0.990)0.981 (0.969, 0.990)0.956 (0.944, 0.967)0.956 (0.942, 0.968)High-grade visceral injury0.3130.963 (0.911, 1.000)0.896 (0.816, 0.961)0.878 (0.805, 0.951)0.714 (0.538, 0.875)Strangulated intestinal obstruction0.4900.962 (0.889, 1.000)0.912 (0.799, 0.986)0.822 (0.625, 0.975)0.771 (0.500, 0.968)Data in parentheses are 95% confidence intervals. Operating thresholds indicate model-score cutoffs selected on the internal validation set to prioritize high sensitivity and fixed before evaluation on the internal test set.AUROC area under the receiver operating characteristic curve. Performance of AbdomenNet on the internal test set at high-sensitivity operating thresholds Data in parentheses are 95% confidence intervals. Operating thresholds indicate model-score cutoffs selected on the internal validation set to prioritize high sensitivity and fixed before evaluation on the internal test set. AUROC area under the receiver operating characteristic curve. Discussion In this multicenter study, we developed and systematically validated AbdomenNet, a self-supervised foundation model for NCCT-based acute abdomen diagnosis, risk stratification, and emergency triage. AbdomenNet showed consistent diagnostic performance across independent external cohorts and improved radiologist accuracy and efficiency in the reader study. By addressing multiple stages of the acute abdomen care pathway, including disease detection, surgical risk stratification, and triage prioritization, AbdomenNet provides evidence that AI can enhance the clinical utility of NCCT in real-world emergency settings. In the workflow reconstruction analysis, AI-driven prioritization reduced report TAT for emergent cases without significantly delaying the remaining emergency CT cohort, suggesting its potential to mitigate workflow-related diagnostic delays in high-volume emergency departments. Importantly, our study does not aim to position NCCT as a replacement for CECT or CTA, which remain the guideline-recommended modalities for the definitive evaluation of many acute abdominal conditions5. Rather, AbdomenNet is designed to augment the diagnostic performance of NCCT within its real-world clinical role as a rapid initial assessment and triage tool. By improving NCCT-based detection of emergent findings, AbdomenNet may help identify patients who require urgent escalation to CECT, CTA, or surgical consultation. A key innovation of our work is the use of an image-only foundation model pre-trained on 103,989 unlabeled abdominal NCCT examinations through self-supervised learning. This approach mitigates the manual annotation bottleneck common in medical imaging AI and enables the model to learn transferable imaging representations from large-scale unlabeled data21,22. Notably, although the pre-training and fine-tuning data were obtained from a single institution, they included examinations acquired on multiple scanner models. In addition, the three external validation centers used different scanner models and acquisition configurations. The comparable performance observed across these independent external cohorts supports the generalizability of the learned representations across heterogeneous scanner hardware and acquisition/reconstruction configurations. Beyond simple disease detection, AbdomenNet addresses the emergent clinical need for risk stratification23,24. By integrating cascaded sub-models to identify high-grade visceral injury, strangulated intestinal obstruction, and upper gastrointestinal perforation, the system delivers actionable insights that directly inform surgical decision-making. This two-step process of detection followed by targeted severity assessment mirrors clinical reasoning and represents a more clinically integrated form of AI-assisted decision support than previously described single-task models25–27. In cascaded pathways, overall performance depends on both the detection and risk-stratification steps, because missed detections at the first step may preclude downstream severity assessment. In our triage workflow, high-sensitivity thresholds were therefore applied at both stages to reduce true-positive attrition across the cascade. This design also enables each submodel to be independently optimized for its specific clinical task while providing interpretable, stage-specific outputs that may better support urgent management decisions. Additional ablation and benchmark analyses supported the methodological choices underlying AbdomenNet. Domain-specific image-only self-supervised pretraining on abdominal NCCT demonstrated overall performance gains over training from scratch, general-domain initialization, and alternative self-supervised objectives; AbdomenNet also outperformed the vision-language CT model Merlin, including a configuration re-pretrained on our NCCT corpus with paired radiology reports, as well as publicly available CT foundation models including CT-FM, TAP-CT, and Curia. Although recent radiology vision-language models, including Merlin, have shown the potential of image-text pretraining, our benchmark results suggest that image-only domain-specific pretraining was more effective for emergency NCCT-based acute abdomen diagnosis in this setting. These findings suggest that matching the pretraining data, learning objective, and model architecture to the target emergency NCCT scenario contributes substantially to acute abdomen diagnostic performance. This may be particularly relevant for NCCT-based emergency triage, where many emergent findings are subtle and must be prioritized before a finalized radiology report is available, while emergency reports may be brief, heterogeneous, or influenced by subsequent clinical or contrast-enhanced evaluations. In addition, the 2D backbone with Transformer-based volumetric aggregation balances large-scale representation learning with recovery of inter-slice context, enabling effective feature extraction from variable-length CT volumes. The clinical utility of AbdomenNet was further supported by our multi-reader, multi-case (MRMC) reader study. AI assistance not only improved diagnostic accuracy for both junior and senior radiologists, effectively narrowing the experience gap, but also reduced the median reading time by approximately 27%. The reading times in our MRMC study should be interpreted as task-specific interpretation times rather than full clinical reporting times, because radiologists evaluated only the 11 predefined conditions without comprehensive clinical correlation or report composition. Moreover, scan coverage varied by indication, with some cases comprising an upper or lower abdominal scan rather than a full abdominopelvic scan. The crossover design supports the validity of the relative time reduction attributable to AI assistance as an internal comparison. This dual benefit is particularly relevant in the high-pressure, time-sensitive context of emergency radiology, where AI may act as a vigilant second reader, mitigate perceptual errors for subtle findings on NCCT, and alleviate the burden on strained clinical services28,29. A primary demonstration of AbdomenNet’s translational potential lies in its AI-driven triage workflow. By automatically prioritizing studies, AbdomenNet reduced the median report TAT for true emergent cases by 37 min without significantly delaying the remaining emergency CT cohort in the retrospective workflow reconstruction. For time-sensitive conditions such as acute aortic dissection and acute mesenteric ischemia, earlier radiologist review may be clinically important because diagnostic delays are directly associated with adverse outcomes30,31. This may help allocate expert attention more effectively to the most time-sensitive cases32–34. In clinical deployment, the operating threshold should be tailored to the intended use. For radiologist-supervised AI-assisted reading, a more balanced sensitivity-specificity trade-off may be appropriate, whereas AI-driven worklist prioritization in high-volume or resource-limited settings should prioritize sensitivity to minimize missed emergent diagnoses. Accordingly, we calibrated the triage thresholds to achieve high sensitivity for emergent conditions and risk-stratification subtasks. Although this approach increased false-positive P1 alerts, these cases would be prioritized for earlier radiologist review rather than directly triggering treatment decisions, and no evidence of queue saturation was observed in our retrospective reconstruction. Given that a false-positive alert primarily affects reading priority, whereas a missed emergent diagnosis may cause irreversible patient harm, these findings support AbdomenNet as a radiologist-supervised triage-support tool rather than an autonomous diagnostic system. The Grad-CAM analysis, including anatomical localization quantified using TotalSegmentator-derived organ-level masks, together with the systematic error analysis, further highlighted clinically relevant failure modes that should inform deployment. Although AbdomenNet generally attended to plausible disease-related anatomical regions, discordant attention and false-negative predictions were more often observed in early or subtle presentations, including mesenteric ischemia with subtle or segmental bowel wall changes, acute aortic dissection cases with limited visibility of the intimal flap on NCCT, and partial intestinal obstruction with only mild proximal bowel dilation. These findings suggest that AbdomenNet can support radiologists by prioritizing high-risk cases for earlier review, but its outputs should be interpreted within a supervised triage workflow, particularly for subtle NCCT presentations with direct safety implications. It is important to acknowledge that mesenteric ischemia and visceral injury are inherently difficult to diagnose on NCCT. A recent meta-analysis reported a pooled sensitivity of approximately 76% for non-CTA CT protocols in diagnosing acute mesenteric ischemia35. In our study, AbdomenNet achieved a sensitivity of 0.838 for this condition on the internal test set, suggesting that the model may capture subtle NCCT signs, including bowel wall thickening, mesenteric fat stranding, and mesenteric vascular abnormalities. This interpretation is supported by related evidence that deep learning can identify superior mesenteric artery abnormalities on plain CT36. Similarly, prior work has demonstrated clinically meaningful sensitivity for abdominal trauma detection on NCCT using deep learning37. These findings support the role of AI-augmented NCCT as an initial triage aid for identifying patients who may require escalation to contrast-enhanced imaging. Looking ahead, these findings have implications for future clinical care and research. In clinical practice, AbdomenNet could be integrated into the emergency radiology workflow as a radiologist-supervised worklist prioritization tool, prompting earlier review of high-risk NCCT examinations and helping identify patients who may require escalation to CECT, CTA, surgical consultation, or interventional treatment when appropriate. Such a system may be particularly valuable in high-volume emergency departments or during off-hours, where delays in image interpretation can occur. Future research should evaluate this approach prospectively in consecutive, unselected emergency NCCT cohorts and should assess not only diagnostic accuracy but also workflow, safety, and patient-centered endpoints, including report TAT, time to contrast-enhanced imaging, surgical consultation, or therapeutic intervention, missed emergent diagnoses, false-positive alert burden, and clinical outcomes. Additional studies are also needed to optimize deployment-specific thresholds, monitor calibration across institutions and scanner protocols, evaluate alert fatigue and health-economic impact, and extend the framework to CECT, CTA, and rarer life-threatening abdominal emergencies. We acknowledge several limitations. First, NCCT inherently lacks vascular and perfusion information and is less sensitive than CECT for certain acute abdominal conditions. Our study was designed to enhance the diagnostic utility of NCCT within its role as a rapid initial triage tool, rather than to match or replace CECT or CTA. Therefore, our results are specific to NCCT-based emergency triage and should not be extrapolated to contrast-enhanced imaging. Whether a similar AI-based approach can be effectively applied to CECT or CTA warrants further investigation. Second, our emergency triage study was retrospective and based on workflow reconstruction; prospective validation is needed to confirm the observed workflow benefits and calibrate deployment-specific thresholds in live clinical settings. Third, the study cohort has potential selection and spectrum biases. The cohort, drawn from a suspected acute abdomen setting, was enriched for the 11 target acute abdominal conditions, with approximately 22% of patients having none of the evaluated target pathologies, which may have led to overestimation of diagnostic performance relative to fully unselected real-world emergency abdominal CT populations. The exclusion of patients with insufficient or internally inconsistent reference-standard evidence may have introduced additional selection bias. Future validation in consecutive, unselected emergency NCCT cohorts is warranted. Fourth, the MRMC reader study was conducted on an internal test subset to ensure strict crossover control and standardized AI-assisted reading conditions; therefore, the effect of AI assistance on radiologist performance should be further validated in external prospective reader studies. Finally, although AbdomenNet covers a broad range of acute abdominal emergencies, its performance for rarer but life-threatening conditions, such as ruptured abdominal aortic aneurysm, remains to be determined in future studies with larger multicenter cohorts. In conclusion, AbdomenNet, a foundation model-based AI solution for acute abdomen on NCCT, demonstrated value across diagnostic accuracy, radiologist efficiency, and emergency triage. This work provides a robust foundation for prospective validation to improve patient outcomes. Methods Ethical approval This research complies with all relevant ethical regulations. This multicenter retrospective study protocol was approved by the ethics committees of the participating institutions: Shanghai Sixth People’s Hospital (2025-KY-268[K]), Shanghai East Hospital (2025YS-249), Southeast University Affiliated Nantong First People’s Hospital (2025-KT365-01), and Wuxi No.2 People’s Hospital (WXEY-2025-200). Given the retrospective design and use of de-identified data, the ethics committees granted a waiver of written informed consent. All imaging data were anonymized prior to analysis by removing patient identifiers. The study was conducted in accordance with the principles of the Declaration of Helsinki and relevant regulatory requirements. No participant compensation was provided because this study used retrospective de-identified clinical and imaging data. Data acquisition The data acquisition process for this study was designed in three key stages: pre-training, fine-tuning, and external validation. Age and recorded sex were extracted from electronic medical records. Recorded sex was obtained from routine clinical registration and was not collected through study-specific self-report. Gender identity was not available in the retrospective de-identified clinical and imaging records used in this study. Sex and gender were not used as inclusion or exclusion criteria, balancing variables, or prespecified effect modifiers in the study design. No sex- or gender-specific analysis was performed because the study was not designed or powered to evaluate sex- or gender-specific diagnostic performance. Cohort-level age and recorded-sex distributions are reported in Table 1 and provided in the Source Data file. Pre-training Dataset (Unlabeled): We retrospectively collected abdominal NCCT scans performed between January 2016 and December 2020 at Shanghai Sixth People’s Hospital. These scans were acquired for various clinical indications from diverse departments (including emergency, outpatient, and inpatient wards). This dataset was used exclusively for self-supervised pre-training and was not involved in subsequent labeling or model evaluation. Fine-tuning Dataset (Labeled): We retrospectively included adult patients (≥18 years) who underwent abdominal NCCT for acute abdominal pain or suspected acute abdomen at Center 1 between January 2021 and December 2024. Exclusion criteria comprised severe motion artifacts, poor image quality, or inability to establish a definitive reference diagnosis from the available clinical evidence. The latter included patients discharged or transferred before completion of a definitive diagnostic workup, patients with a documented uncertain or provisional final diagnosis, and patients for whom the available clinical, imaging, and follow-up evidence remained insufficient or internally inconsistent to support a definitive adjudicated diagnosis. In total, 139 patients (2.3% of the initially screened cohort) were excluded due to inability to establish a definitive reference diagnosis. We established a multi-label classification task encompassing 11 acute abdominal conditions (Table S5), in which each condition was independently predicted as present or absent, allowing the model to flag multiple co-occurring conditions simultaneously. For each target condition, per-condition AUROC, sensitivity, and specificity were computed in a one-vs-rest manner; patients with the condition under evaluation were treated as positives, whereas patients without that condition, including those with other concurrent target conditions, were treated as negatives for the condition under evaluation. Based on the immediate need for surgical or interventional radiology intervention, we categorized five of these as emergent conditions: acute abdominal aortic dissection, acute mesenteric ischemia, acute gastrointestinal perforation, acute intestinal obstruction (strangulated), and acute visceral injury (high-grade). For acute abdominal aortic dissection, only patients presenting to the emergency department with acute-onset symptoms and a newly diagnosed dissection confirmed on paired CECT/CTA were included; patients with known chronic or stable dissections were excluded. The remaining six were defined as urgent conditions: acute cholecystitis, acute choledocholithiasis, acute pancreatitis, acute ureteral stone, acute diverticulitis, and acute appendicitis. External Validation Datasets: To comprehensively assess the model’s generalizability across different geographical regions and healthcare settings, we retrospectively collected emergency abdominal NCCT cases from three independent tertiary medical centers in China (Shanghai East Hospital, Southeast University Affiliated Nantong First People’s Hospital, and Wuxi No.2 People’s Hospital) between January and April 2025. Scans were acquired using seven scanner models from four manufacturers (GE, Toshiba, Siemens, and Philips), with tube voltages of 100 or 120 kVp, tube current-time products of 80–380 mAs, and reconstructed slice thicknesses of 1 or 5 mm; center-specific acquisition and reconstruction parameters are summarized in Table S6. Reference standard labeling The reference standard labeling for all abdominal NCCT scans was established through independent review by two senior radiologists (X. Wei, D. Wang), each with over 10 years of experience in abdominal imaging. The reference standard for the final diagnosis was based on a comprehensive review of medical records, including patient history, paired CECT/CTA, surgical reports, laboratory results, pathology reports, radiological follow-up, and the final clinical diagnosis. Any discrepancies in interpretation were resolved by a third senior radiologist (Y. Li), with over 20 years of experience, through adjudication to reach a consensus, thereby improving label reliability. Design of risk stratification subtasks We designed three key risk stratification subtasks to cascade from the primary 11-task multi-label classification. When the primary model detected a condition (e.g., intestinal obstruction or visceral injury), a corresponding sub-model provided further assessment of severity or subtype. Crucially, these outputs are aligned with clinical indicators for surgical decision-making. Specifically, these subtasks included(1): identification of high-grade visceral injury (AAST grade ≥III)38 (2); identification of strangulated intestinal obstruction; and (3) identification of upper gastrointestinal perforation. The reference standards for these subtasks were as follows: the severity of visceral injury was determined using the AAST grading system on paired emergency CECT scans; the labels for strangulated obstruction and upper gastrointestinal perforation were confirmed by surgical findings or subsequent histopathological evidence. Further details of the design of risk stratification subtasks are available in Supplementary Methods. AbdomenNet pre-training, fine-tuning, and interpretability analysis We employed the self-supervised learning framework DINOv218 for pre-training. This method uses self-supervised teacher-student learning with self-distillation to learn robust and generalizable visual representations without manual annotations. Specifically, we used the largest Vision Transformer variant from DINOv2 (ViT-g/14, 1.1 billion parameters, patch size 14 × 14) and pre-trained it on 2D axial CT slices at a resolution of 518 × 518 to create a general-purpose 2D feature extractor. During the fine-tuning stage, this pre-trained extractor served as the backbone for our AbdomenNet. For a given CT volume, AbdomenNet first encodes each slice to extract 2D features. The resulting sequence of slice-level features is then fed into a classifier composed of a three-layer Transformer decoder. This design allows the model to capture global inter-slice dependencies to perform the final diagnostic task, offering a more effective integration of volumetric information compared to a conventional linear classification head (Fig. 2D). To evaluate interpretability, Grad-CAM maps were generated for representative cases, and region-level attention localization was quantified using TotalSegmentator-derived39 organ-level ROIs to assess the anatomical plausibility of model attention. Detailed pre-training, fine-tuning and Grad-CAM procedures are provided in the Supplementary Methods. Baseline models and ablation study To assess the contributions of pre-training source, self-supervised objective, model architecture, and pre-training data scale, we compared a series of baseline and comparison models using identical data splits and harmonized downstream protocols. These models included: (Ⅰ) ViT-g/14 trained from scratch with random initialization; (Ⅱ) ViT-g/14 initialized with publicly available general-domain DINOv2 weights pre-trained on the LVD-142M natural image dataset; (Ⅲ) ViT-g/14 pre-trained on the unlabeled abdominal NCCT corpus using SimCLR40; (Ⅳ) ViT-g/14 pre-trained on the same corpus using MAE41; (Ⅴ) ViT-g/14 pre-trained using our domain-specific DINOv2 framework, AbdomenNet; (Ⅵ) Merlin42, a multimodal model with a 3D ResNet-based CT encoder, evaluated under two configurations: public pre-trained weights fine-tuned on our labeled dataset, and re-pre-training on 103,989 abdominal NCCT examinations with paired radiology reports followed by fine-tuning on the same labeled dataset; (Ⅶ) three publicly available general-purpose CT foundation models, CT-FM43, TAP-CT44, and Curia45; and (Ⅷ) a radiomics pipeline with PyRadiomics feature extraction, LASSO-based feature selection, and conventional machine-learning classifiers. Radiologist evaluation To evaluate the clinical utility of AbdomenNet, we conducted a MRMC study with a crossover design. Four radiologists participated, including two junior radiologists (2 years of experience in abdominal imaging) and two senior radiologists (9 and 10 years of experience). A representative set of 380 cases, covering all 11 primary tasks, was randomly sampled from the internal test set. This case set was then randomly divided into two equal cohorts (Set A, n = 190; Set B, n = 190), and the study proceeded in two reading phases separated by a washout period of 3 months to minimize recall bias. In the first phase (Reading Session 1), all four radiologists independently diagnosed all 380 cases, reviewing Set A with AI assistance (AI-aided) and Set B without AI assistance (unaided). After the washout period, the second phase (Reading Session 2) commenced, during which the same radiologists re-interpreted the cases with the conditions reversed (Set A without AI, Set B with AI). In the AI-assisted condition, radiologists were presented with three components alongside the original NCCT images1: a ranked list of predicted probabilities for all 11 conditions2; the top predicted diagnosis highlighted as the model’s primary recommendation; and ref. 3 Grad-CAM heatmaps for the top predicted condition were overlaid on the axial CT slice with the highest model activation. Radiologists were instructed to use this information as decision support and to make their final diagnostic judgment, with the AI output serving as decision support rather than a directive (Fig. S6). In both phases, all 380 cases were pooled and presented in a single, stratified random order to blind the readers to the case allocation. This crossover design allowed each case to serve as its own control, providing a robust comparison of diagnostic performance with and without AI while controlling for case difficulty and minimizing recall bias. Reconstructed AI-triage workflow We conducted a retrospective simulation-based workflow reconstruction to estimate the potential time savings of AI-based worklist prioritization under predefined priority-queue assumptions. The reconstruction used a real-world cohort of emergency CT examinations performed within the dedicated emergency CT workflow at Shanghai East Hospital from January to June 2025. A high-sensitivity triage strategy was implemented by calibrating operating thresholds on the internal validation set to achieve a diagnostic sensitivity of ≥0.95 for the five emergent conditions and two risk-stratification subtasks. These thresholds were then fixed and applied without further adjustment to the held-out internal test set and to the independent emergency CT cohort used for workflow reconstruction. The reconstruction used real-world scan completion timestamps as the event timeline. The baseline workflow was defined using the historically observed report TAT for all 18,398 emergency CT examinations in the cohort. AbdomenNet was applied to the subset of 1761 abdominal NCCT examinations from patients presenting with acute abdominal symptoms and with definitive reference diagnoses to assign a triage level (P1, emergent priority; P2, routine priority). In the reconstructed AI-triage workflow, P1-flagged cases were inserted after any existing P1 cases but ahead of P2 cases in the reading queue. The radiologist’s currently in-progress examination was not interrupted, and the P1-flagged case became the next eligible case after completion of the ongoing examination. Observed per-case reading durations were preserved during the reconstruction. We analyzed TAT and queue length, with analyses stratified into daytime (08:00–19:59) and nighttime (20:00–07:59) shifts to account for operational variation. Statistical analysis Statistical analyses were performed using Python (version 3.10.14). A two-sided P value less than 0.05 was considered to indicate statistical significance. Model performance metrics (including AUROC, sensitivity, and specificity) were calculated, with 95% CIs determined via 10,000 bootstrap iterations. For the MRMC study, a GLMM model was used to assess the impact of AI on diagnostic accuracy, accounting for reader and case effects. AUROCs from unaided and AI-assisted interpretations were compared using paired two-sided DeLong tests. AUPRCs were compared using two-sided paired bootstrap tests. Sensitivities were compared using two-sided McNemar tests. Paired, non-normally distributed reading times were compared using two-sided Wilcoxon signed-rank tests. For the workflow reconstruction analysis, paired, non-normally distributed report TATs under the FCFS and priority-based workflows were compared using two-sided Wilcoxon signed-rank tests. No adjustment for multiple comparisons was applied. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Source Data