A retrospective dual-center investigation has demonstrated that a non-contrast computed tomography framework powered by deep learning can accurately predict severity in patients diagnosed with acute pancreatitis by combining three-dimensional volumetric pancreatic imaging, two-dimensional body composition features, and routine clinical laboratory markers. Evaluating abdominal scans obtained within seventy-two hours of symptom onset, researchers discovered that this multi-source architecture achieved an area under the receiver operating characteristic curve of 0.897 in the training cohort, 0.909 in an internal test cohort, and 0.910 in an independent external test cohort. Rather than assessing the inflamed organ in isolation, the computational strategy simultaneously captures local parenchymal injury alongside systemic host vulnerability through opportunistic evaluation of visceral adipose tissue and skeletal muscle at the third lumbar vertebral level. The resulting predictive output demonstrated a consistent stepwise elevation from mild cases to moderately severe and severe disease classifications, offering an automated avenue for early risk stratification.
Acute pancreatitis represents a frequent and clinically volatile gastrointestinal inflammatory disorder characterized by unpredictable patient trajectories during the initial phases of hospital admission. Although the vast majority of individuals experience a mild and self-limiting disease course that resolves with supportive care, approximately fifteen to twenty percent of patients progress into moderately severe acute pancreatitis or severe acute pancreatitis. These advanced disease stages carry substantial clinical risks, including the development of persistent organ failure, extensive local necrotic collections, secondary systemic complications, prolonged hospital stays, and elevated mortality rates. Because therapeutic interventions and intensive monitoring strategies are most effective when initiated before systemic decompensation becomes irreversible, identifying individuals destined for severe complications within the earliest days of admission remains an urgent clinical priority. However, differentiating between individuals who will recover uneventfully and those who will deteriorate rapidly continues to present substantial diagnostic hurdles in acute triage environments.
Clinicians have traditionally evaluated early disease trajectory using bedside risk stratification tools such as the Ranson criteria, the Acute Physiology and Chronic Health Evaluation II, and the Bedside Index for Severity in Acute Pancreatitis. While these clinical scoring instruments remain standard reference points, they frequently necessitate the collation of numerous disparate physiological variables and require complex calculations that hinder immediate clinical utility during initial emergency department encounters. Furthermore, established imaging-based rubrics, notably the conventional Computed Tomography Severity Index and the Modified Computed Tomography Severity Index, depend almost entirely upon visual assessments of parenchymal necrosis and localized fluid collections visualized on contrast-enhanced scans. Because hallmark morphologic indicators of tissue necrosis often require up to seventy-two hours following initial symptom onset to become fully established on standard radiological examinations, conventional computed tomography evaluations performed during the acute presentation window frequently yield ambiguous prognostic assessments.
To overcome the limitations of isolated morphologic inspection, researchers have increasingly recognized that acute pancreatitis involves both localized glandular parenchymal destruction and broader systemic metabolic disruption reflected throughout host tissues. Body composition parameters derived opportunistically from abdominal computed tomography examinations provide quantifiable indicators of baseline physiological reserve and inflammatory susceptibility without requiring additional imaging procedures. Visceral adipose tissue actively participates in systemic inflammatory cascades by releasing pro-inflammatory cytokines, adipokines, and toxic unesterified free fatty acids that can exacerbate pancreatic tissue injury and vascular permeability. Conversely, skeletal muscle mass and muscular attenuation reflect underlying nutritional reserves and physical metabolic tolerance against acute systemic catabolic stress. While previous retrospective investigations documented correlations between visceral adiposity, sarcopenia, and adverse clinical outcomes, earlier methodologies depended on labor-intensive manual tracings or rudimentary cross-sectional area and density measurements that failed to capture complex spatial heterogeneity.
Addressing these methodological constraints, the investigators structured a dual-center retrospective study involving five hundred and thirty-six consecutive adult patients who underwent non-contrast computed tomography within seventy-two hours of initial symptom onset. The primary research dataset, accrued at Xuanwu Hospital Capital Medical University between January 2021 and September 2025, encompassed four hundred and twenty-six eligible individuals randomly allocated at the patient level into a training cohort of two hundred and ninety-eight individuals and an internal validation cohort of one hundred and twenty-eight individuals. An additional external validation cohort consisting of one hundred and ten consecutive patients was compiled at Chongqing General Hospital between January 2023 and September 2024. Across the entire analyzed study population, the mean participant age was fifty years with a standard deviation of sixteen years, comprising three hundred and sixty male patients and one hundred and seventy-six female patients.
Rigorous screening across both medical centers excluded patients presenting with trauma-induced or post-surgical pancreatitis, concurrent neoplastic lesions, incomplete clinical or laboratory documentation, and uninterpretable or low-quality radiological scans. All abdominal examinations were performed using non-contrast computed tomography protocols, eliminating confounding factors associated with intravenous iodinated contrast timing. For each enrolled subject, the entire pancreatic parenchymal volume was delineated manually on axial computed tomography slices using ITK-SNAP software, after which cropped volumetric regions of interest were standardized to uniform dimensions of 128 by 128 by 32 voxels. Simultaneously, body composition profiling was accomplished by selecting the single axial computed tomography slice corresponding to the third lumbar vertebra. Visceral adipose tissue and skeletal muscle compartments were segmented automatically utilizing an open-source neural network tool followed by blinded manual refinement, with resultant two-dimensional regions resized to 224 by 224 pixels.
The modeling framework leveraged fine-tuned ResNet18 convolutional neural network architectures to extract deep quantitative imaging representations from the preprocessed anatomical regions. A three-dimensional ResNet18 network extracted volumetric feature representations directly from the segmented pancreatic volumes, while dedicated two-dimensional ResNet18 networks processed the L3-level visceral adipose tissue and skeletal muscle regions, drawing feature vectors from penultimate average pooling layers. Extracted imaging attributes were standardized via z-score normalization and filtered through least absolute shrinkage and selection operator regression to isolate the most robust predictors. Concurrently, multivariable logistic regression evaluated admission clinical parameters, identifying C-reactive protein with an odds ratio of 1.005 and serum triglycerides with an odds ratio of 1.041 as independent laboratory markers of severe progression. These analytical pathways culminated in six comparative models: clinical, pancreas, visceral fat, skeletal muscle, pancreas-clinical, and the comprehensive multi-source combined model.
Comparative discriminative evaluations demonstrated the clear prognostic superiority of the multi-source design over models relying on isolated data streams. The standalone clinical model yielded modest discriminative power, achieving areas under the receiver operating characteristic curve of 0.729 in the training cohort, 0.722 in the internal test cohort, and 0.712 in the external validation cohort. Incorporating three-dimensional pancreatic imaging features noticeably improved discrimination, with the combined model recording areas under the curve of 0.897, 0.909, and 0.910 across the respective cohorts. Formal DeLong testing revealed that the combined architecture significantly outperformed the intermediate pancreas-clinical model in both the primary training set with a p-value of 0.002 and the external test set with a p-value of 0.020, while maintaining equivalent performance in the internal validation cohort where both models reached an area under the curve of 0.909.
The incremental diagnostic benefit attributable to computed tomography-derived body composition metrics was further substantiated through comprehensive reclassification and decision curve analyses. In the internal validation cohort, integrating body composition features yielded a net reclassification improvement of 0.791 with a p-value below 0.001 and an integrated discrimination improvement of 0.072 with a p-value of 0.003. In the external testing cohort, net reclassification improvement reached 0.777 with a p-value below 0.001, accompanied by an integrated discrimination improvement of 0.108 with a p-value below 0.001. When benchmarked against the standard Bedside Index for Severity in Acute Pancreatitis, the combined model demonstrated statistically superior discrimination across all three datasets, easily surpassing the clinical scoring tool which achieved areas under the curve of only 0.749, 0.745, and 0.772. Furthermore, decision curve analysis verified higher net clinical benefit across relevant threshold probabilities.
To demystify algorithmic outputs, the researchers implemented complementary interpretability frameworks using Gradient-weighted Class Activation Mapping and Shapley Additive Explanations. Spatial visualization demonstrated that the neural network concentrated attention heavily within inflamed pancreatic tissue and surrounding visceral adipose tissue compartments in severe cases, whereas mild cases exhibited localized, restrained activation patterns. Feature attribution plots confirmed that admission C-reactive protein, serum triglycerides, and high-dimensional deep learning features derived from visceral fat and pancreatic parenchyma served as primary drivers of individual risk scores. Despite these encouraging results, the authors highlighted critical operational limitations, including the retrospective observational design, geographical recruitment confined to institutions within a single country, cross-center variations in scanner slice thickness, and reliance on manual pancreatic boundary delineations. Prospective multi-center clinical trials utilizing fully automated segmentation pipelines remain essential before clinical deployment.
Subject of Research: Deep learning-based computed tomography assessment of acute pancreatitis severity
Article Title: CT-based deep learning integrating pancreatic imaging and body composition for severity prediction in acute pancreatitis
Article References: Zeng, D., He, M., Lin, B., Liu, Y., Wang, X., & Lu, J. (2026). CT-based deep learning integrating pancreatic imaging and body composition for severity prediction in acute pancreatitis. Insights into Imaging, 17(1), Article 266. https://doi.org/10.1186/s13244-026-02443-w
Image Credits: AI Generated
DOI: 10.1186/s13244-026-02443-w
Keywords: acute pancreatitis, computed tomography, deep learning, body composition, visceral adipose tissue, skeletal muscle, disease severity prediction, artificial intelligence, C-reactive protein, hypertriglyceridemia, Atlanta classification, ResNet18, receiver operating characteristic, decision curve analysis, BISAP score, medical imaging, model interpretability
Cite Scienmag News
Blake Davidson. (November 19, 2026). CT Deep Learning Model Integrates Body Composition to Predict Pancreatitis Severity. Scienmag. https://scienmag.com/ct-deep-learning-model-integrates-body-composition-to-predict-pancreatitis-severity/
Blake Davidson. "CT Deep Learning Model Integrates Body Composition to Predict Pancreatitis Severity." Scienmag, 19 November 2026, https://scienmag.com/ct-deep-learning-model-integrates-body-composition-to-predict-pancreatitis-severity/. Accessed 19 November 2026.
Blake Davidson. "CT Deep Learning Model Integrates Body Composition to Predict Pancreatitis Severity." Scienmag. November 19, 2026. https://scienmag.com/ct-deep-learning-model-integrates-body-composition-to-predict-pancreatitis-severity/








