Pancreatic cystic lesions are being found more frequently as cross-sectional imaging becomes more common. Their clinical importance varies widely: some lesions remain harmless, whereas others carry malignant potential and may require surgery. Current assessment relies on features such as cyst walls, septa, mural nodules, solid components, pancreatic-duct changes, and surrounding tissue involvement. Yet these signs often overlap across disease types, and interpretation can vary with reader experience. Conventional radiomics can quantify image patterns, but it depends heavily on segmentation, handcrafted feature selection, and technical choices that may limit reproducibility. Because of these challenges, deeper investigation is needed into noninvasive models that integrate complementary imaging phases and produce reliable preoperative risk estimates.
Researchers from the Department of Radiology at Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, reported (DOI: 10.12290/xhyxzz.2026-0410) the study online on July 13, 2026, in the Medical Journal of Peking Union Medical College Hospital. The team developed and evaluated three-dimensional deep-learning models using arterial- and venous-phase contrast-enhanced computed tomography (CT) scans from patients with pathologically confirmed pancreatic cystic lesions (PCLs). Their goal was to determine whether combining both imaging phases could improve benign-malignant classification and provide an objective aid for clinical decision-making before surgery.
The retrospective, single-center study included 480 patients with 485 lesions collected between June 2014 and May 2023; 206 lesions were malignant and 279 were benign. The researchers divided lesions into training, validation, and independent test sets in a 3:1:1 ratio while keeping all lesions from the same patient in one subset. After image registration and preprocessing, they supplied four three-dimensional channels to five neural-network architectures: arterial-phase CT, venous-phase CT, a pancreatic mask, and a lesion mask. Postoperative pathology served as the reference standard. ResNeXt50 delivered the highest overall point estimates. In the test set, it achieved an area under the receiver operating characteristic curve (AUC) of 0.822, 73.20% accuracy, 82.93% sensitivity, and 66.07% specificity. Its AUC exceeded those of venous-only and arterial-only versions, which reached 0.796 and 0.785, respectively. It also showed higher point estimates than conventional radiomics for AUC, accuracy, and sensitivity. However, the dual-phase gains over single-phase models and the apparent advantages over radiomics were not statistically significant. Model classifications showed good agreement with two readings by one radiologist, but the study was not designed to establish equivalence or superiority to physicians.
The authors said the model's main promise lies in giving clinicians an additional, repeatable view of malignancy risk rather than replacing expert judgment. Arterial and venous images reveal different parts of a lesion's story, they said, from vascularized solid tissue to cyst walls, septa, and relationships with the surrounding pancreas. Combining these signals may help flag lesions that deserve closer attention. They added that specificity and probability calibration still require improvement, and that model outputs should be interpreted alongside clinical findings, laboratory tests, magnetic resonance imaging, endoscopic ultrasound, and radiologist assessment.
The approach could eventually support triage, multidisciplinary review, and follow-up planning for patients with uncertain pancreatic cysts. Its relatively high sensitivity may be useful for identifying malignant or high-risk lesions that should undergo further evaluation, while its standardized analysis could reduce variability between readings. However, the evidence remains preliminary: the dataset came from one hospital, included only surgically treated patients, required manual lesion segmentation, and contained a limited number of cases for deep-learning comparisons. Before clinical deployment, the model will need recalibration and validation across hospitals, scanners, and imaging protocols. Future systems may also combine CT with magnetic resonance imaging, endoscopic ultrasound, cyst-fluid analysis, tumor markers, and clinical symptoms.
###
References
DOI
10.12290/xhyxzz.2026-0410
Original Source URL
https://xhyxzz.pumch.cn/article/doi/10.12290/xhyxzz.2026-0410
Funding information
National Natural Science Foundation of China, grant 82372051; National Key Research and Development Program of China, “Diagnostic and Therapeutic Equipment and Biomedical Materials,” grant SQ2024YFC2400371; National High Level Hospital Clinical Research Funding, grant 2025-PUMCH-D-002; and Beijing Natural Science Foundation Youth Program, grant 7264409.
About Medical Journal of Peking Union Medical College Hospital
Medical Journal of Peking Union Medical College Hospital is a leading clinical medicine publication, supported by the multidisciplinary expertise of Peking Union Medical College Hospital. It features the latest research, advancements, and academic trends in clinical and translational medicine, pharmacy, and related interdisciplinary fields, catering to clinicians and medical students across China. The journal aims to promote the exchange of medical knowledge and serve as a high-quality platform for leading academic discussions and fostering scholarly debate in clinical medicine. The journal is listed in China's Core Journals of Science and Technology (CSTPCD), Chinese Science Citation Database (CSCD), A Guide to the Core Journals of China, and the Chinese Biomedical Literature Database (CMCC). Full-text content is accessible on platforms such as Wanfang Data, CNKI, and Chongqing VIP Database. It is indexed in Scopus (Netherlands), the Directory of Open Access Journals (DOAJ) in Sweden, and the Japan Science and Technology Agency Database (JST).