Pancreatic cancer diagnosis on unenhanced CT with deep learning for opportunistic diagnosis - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Radiol Adv . 2026 Mar 24;3(2):umag017. doi: 10.1093/radadv/umag017 Search in PMC Search in PubMed View in NLM Catalog Add to search Pancreatic cancer diagnosis on unenhanced CT with deep learning for opportunistic diagnosis Po-Ting Chen Po-Ting Chen , MD 1 Department of Medical Imaging, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing - original draft Find articles by Po-Ting Chen 1, # , Dawei Chang Dawei Chang , MSc 2 Data Science Degree Program, National Taiwan University and Academia Sinica, Taipei, 106319, Taiwan Data curation, Formal analysis, Methodology, Software, Validation, Writing - original draft Find articles by Dawei Chang 2, # , Yenjia Chen Yenjia Chen , BSc 3 Data Science Degree Program, National Taiwan University and Academia Sinica, Taipei, 106319, Taiwan Data curation, Formal analysis, Methodology, Software, Validation Find articles by Yenjia Chen 3 , Pochuan Wang Pochuan Wang , MSc 4 Department of Computer Science and Information Engineering, National Taiwan University, Taipei, 106319, Taiwan Data curation, Methodology, Software, Validation Find articles by Pochuan Wang 4 , Andre Yanchen Yeh Andre Yanchen Yeh , MD 5 Department of Medicine, National Taiwan University College of Medicine, Taipei, 100233, Taiwan Data curation, Methodology, Software, Validation Find articles by Andre Yanchen Yeh 5 , Kao-Lang Liu Kao-Lang Liu , MD 6 Department of Medical Imaging, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan Investigation, Supervision, Validation Find articles by Kao-Lang Liu 6 , Ming-Shiang Wu Ming-Shiang Wu , MD, PhD 7 Division of Gastroenterology and Hepatology, Department of Internal Medicine, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan 8 Internal Medicine, College of Medicine, National Taiwan University, Taipei, 100233, Taiwan Investigation, Resources, Supervision, Validation Find articles by Ming-Shiang Wu 7, 8 , Wei-Chih Liao Wei-Chih Liao , MD, PhD 9 Division of Gastroenterology and Hepatology, Department of Internal Medicine, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan 10 Internal Medicine, College of Medicine, National Taiwan University, Taipei, 100233, Taiwan Funding acquisition, Methodology, Project administration, Supervision, Visualization, Writing - review & editing Find articles by Wei-Chih Liao 9, 10, 3, ✉ , Weichung Wang Weichung Wang 11 Institute of Applied Mathematical Sciences, National Taiwan University, Taipei, 106319, Taiwan Funding acquisition, Project administration, Resources, Supervision, Validation, Visualization, Writing - review & editing Find articles by Weichung Wang 11, 3, ✉ Author information Article notes Copyright and License information 1 Department of Medical Imaging, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan 2 Data Science Degree Program, National Taiwan University and Academia Sinica, Taipei, 106319, Taiwan 3 Data Science Degree Program, National Taiwan University and Academia Sinica, Taipei, 106319, Taiwan 4 Department of Computer Science and Information Engineering, National Taiwan University, Taipei, 106319, Taiwan 5 Department of Medicine, National Taiwan University College of Medicine, Taipei, 100233, Taiwan 6 Department of Medical Imaging, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan 7 Division of Gastroenterology and Hepatology, Department of Internal Medicine, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan 8 Internal Medicine, College of Medicine, National Taiwan University, Taipei, 100233, Taiwan 9 Division of Gastroenterology and Hepatology, Department of Internal Medicine, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan 10 Internal Medicine, College of Medicine, National Taiwan University, Taipei, 100233, Taiwan 11 Institute of Applied Mathematical Sciences, National Taiwan University, Taipei, 106319, Taiwan ✉ Corresponding authors: Wei-Chih Liao, MD, PhD, Department of Internal Medicine, National Taiwan University Hospital, No. 7, Chung-Shan South Road, Taipei 100225, Taiwan ( [email protected] ); Weichung Wang, PhD, Institute of Applied Mathematical Sciences, National Taiwan University, No. 1, Section 4, Roosevelt Road, Taipei 106319, Taiwan ( [email protected] ) # Po-Ting Chen and Dawei Chang contributed equally to this work. 3 Wei-Chih Liao and Weichung Wang authors are co-senior authors. Roles Po-Ting Chen : MD , Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing - original draft Dawei Chang : MSc , Data curation, Formal analysis, Methodology, Software, Validation, Writing - original draft Yenjia Chen : BSc , Data curation, Formal analysis, Methodology, Software, Validation Pochuan Wang : MSc , Data curation, Methodology, Software, Validation Andre Yanchen Yeh : MD , Data curation, Methodology, Software, Validation Kao-Lang Liu : MD , Investigation, Supervision, Validation Ming-Shiang Wu : MD, PhD , Investigation, Resources, Supervision, Validation Wei-Chih Liao : MD, PhD , Funding acquisition, Methodology, Project administration, Supervision, Visualization, Writing - review & editing Weichung Wang : Funding acquisition, Project administration, Resources, Supervision, Validation, Visualization, Writing - review & editing Received 2025 Nov 19; Revised 2026 Feb 25; Accepted 2026 Mar 9; Collection date 2026 Mar. © The Author(s) 2026. Published by Oxford University Press on behalf of the Radiological Society of North America. This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License ( https://creativecommons.org/licenses/by-nc/4.0/ ), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] for reprints and translation rights for reprints. All other permissions can be obtained through our RightsLink service via the Permissions link on the article page on our site—for further information please contact [email protected]. PMC Copyright notice PMCID: PMC13092298 PMID: 42011325 Abstract Background Pancreatic cancer (PC) is frequently missed on unenhanced CT examinations performed for unrelated clinical indications, where the pancreas is included incidentally and clinical suspicion is low. Purpose To develop and validate a deep learning-based tool for PC diagnosis and risk stratification on unenhanced CT. Materials and Methods This retrospective study included 3080 unenhanced CT studies of Taiwanese patients with PC, other pancreatic diseases and normal pancreas between 2004 and 2019 from a tertiary hospital, randomly divided into training, validation, and internal test sets. Unenhanced CT studies from United States institutions were used for external testing. A hybrid convolutional neural network–transformer model was trained for PC diagnosis and risk stratification. Performance was evaluated using sensitivity, specificity, and area under the curve (AUC), with comparisons to 2 radiologists by McNemar’s test and exploratory decision curve analysis. Results The internal dataset included 713 PCs (mean age, 64.6 ± 12.0 years; 384 men), 1661 normal pancreas and 706 other pancreatic diseases. In an exploratory comparison restricted to unenhanced CT (29 PCs, 31 controls), the sensitivity of computer-aided diagnosis (CAD) tool (89.7%, 72.6-97.8) seemed comparable with that of 1 radiologist (86.2%, 68.3-96.1) and higher than another (41.4%, 23.5-61.1); but wide confidence intervals and inter-radiologist variability warrant cautious interpretation. In the internal test set (142 PCs, 474 controls), sensitivity was 90.8% (84.9-95.0) and specificity 93.0% (90.4-95.2) (AUC: 0.98), with sensitivity comparable to radiologist reports based on enhanced and unenhanced CT (95.4%, 90.2-98.3; P = .21). In the external set (42 PCs, 22 controls), sensitivity was 76.2% (60.5-87.9) and specificity 86.4% (65.1-97.1) (AUC: 0.89). The tool stratified cases into 7 risk levels with likelihood ratios ranging from <0.01 to 173.46. Exploratory decision curve analysis suggested potential net benefit across threshold probabilities. Conclusion This tool may assist in the opportunistic detection and risk stratification of PC on unenhanced CT. Keywords: pancreatic cancer, deep learning, unenhanced computed tomography, risk stratification Summary A deep learning tool diagnosed pancreatic cancer and flagged key diagnostic images on unenhanced CT, achieving sensitivity comparable to radiologists interpreting both enhanced and unenhanced CT images. Key Results In a Taiwanese internal test set, a tool combining a convolutional neural network and transformer diagnosed pancreatic cancer on unenhanced CT with 90.8% sensitivity and 93.0% specificity, comparable to radiologists’ performance. In an external test set from United States institutions using unenhanced CT, the tool demonstrated 76.2% sensitivity and 86.4% specificity. For tumors smaller than 2 cm, the tool achieved sensitivities of 87.5% internally and 77.8% externally on unenhanced CT. Introduction Pancreatic cancer (PC) is the third leading cause of cancer deaths in the United States, with a 5-year survival rate of 12%. 1 Contrast-enhanced CT, the primary imaging modality for the detection and diagnosis of PC, is highly sensitive in detecting PC except for tumors smaller than 2 cm. 2 In routine clinical practice, abdominal CT examinations are frequently performed without intravenous contrast for indications unrelated to suspected pancreatic disease, such as lung cancer screening, renal colic, or nonspecific abdominal pain. 3 , 4 In these examinations, the pancreas is included within the scan range but is not the primary organ of interest, making PC detectable only opportunistically. Under these circumstances, the sensitivity for detecting PC is substantially reduced, largely due to the absence of contrast enhancement. 5 , 6 An effective tool that can supplement radiologist interpretation would fulfill an unmet clinical need. 7 , 8 Artificial intelligence (AI) has shown promise in assisting radiologists with detecting PCs on enhanced CT. Deep learning (DL) can detect occult PCs not visible to radiologists. 9 Previous studies showed that DL-based and radiomics-based models could accurately detect PCs on contrast-enhanced CT and validated their generalizability in a nationwide population-based study. 10–13 By contrast, the potential of AI for detecting PCs on unenhanced CT remains underexplored. We developed a DL-based computer-aided diagnosis (CAD) tool to diagnose PCs in unenhanced CT and flag key diagnostic images. The performance of the CAD tool was compared with that of radiologists, and its generalizability was evaluated using an external test set composed of unenhanced CT images from another country and population. Materials and methods This study trained a CAD tool for PC using unenhanced CT images from a tertiary referral center in Taiwan. Its performance was assessed in a hold-out internal test set and an external test set comprising CT images from multiple US institutions. This study was approved by the research ethics committee (201710050RINA, 201904116RINC), which granted waiver of informed consent due to the retrospective design. Computer codes are available online ( https://github.com/davidzan/PCDoUCTwDL ). Datasets The construction of datasets is summarized in Figure 1 . The internal dataset was composed of unenhanced CTs of 3080 patients with pancreatic adenocarcinoma, other pancreatic diseases, such as acute pancreatitis, chronic pancreatitis and pancreatic cystic lesions, and normal/unremarkable pancreas (henceforth referred to as PCs, OPDs, and Normals, respectively) selected from the imaging archive of a tertiary referral center in Taiwan. The diagnosis of PC required confirmation by histology or cytology. The diagnosis of other pancreatic diseases was determined based on radiologist reports, discharge diagnoses, histology/cytology examinations where available, and clinical follow-up. The diagnosis of normal/unremarkable pancreas was based on formal radiologist reports. The Cancer Registry was reviewed to exclude subsequent PC diagnosis within 1 year. Figure 1. Open in a new tab Datasets. The final dataset was randomly divided (3:1:1) with stratification by disease status into a training set (430 PCs, 997 Normals, 424 OPDs) for training DL classification models, a validation set (141 PCs, 331 Normals, 141 OPDs) for model selection and selection of cutoff for binary classification, and a hold-out internal test set (142 PCs, 332 Normals, 141 OPDs) (details in the Supplementary Material ). Of the 3080 patients in the internal dataset, the contrast-enhanced CT images of 1977 patients had been included in previous studies on the diagnosis of PC in contrast-enhanced CT images. 9 , 10 , 12 , 13 In comparison, this study used the unenhanced CT images of those 1977 patients to explore the potential of AI in diagnosing PC on unenhanced CT. The external test set comprised unenhanced CT images of 30 PCs from the publicly available CPTAC-PDA dataset 14 and of patients from 2 US institutions (12 PCs, 18 Normals, 4 OPDs). All CT examinations were obtained in the unenhanced phase. Section thickness and image size were 0.7-1.5 mm and 512 × 512 pixels, respectively. For subsequent interpretation and analysis, all images were reconstructed to a slice thickness of 5 mm. CT examinations were performed using scanners from 6 vendors ( Table S1 ). Overview of the CAD tool Figure 2 summarizes the design of the CAD tool. A rule-based image processing algorithm was used to localize the abdomen in the input non-contrast images for subsequent resampling and windowing. The preprocessed images were then fed into a convolutional neural network (CNN)-transformer hybrid DL model to generate a binary classification indicating the presence or absence of PC. Additionally, the logit value output by the DL classification model was categorized into distinct risk levels, and the likelihood ratio (LR) of PC for each risk level in the validation set was provided to enable estimation of the post-test probability of PC. Moreover, the images on which the model’s binary classification was primarily based were presented to the users as key images for review. Technical details were provided in the Supplementary Methods . Figure 2. Open in a new tab Workflow of the computer-aided diagnosis tool. Comparison between radiologists and CAD The diagnostic performance of CAD on unenhanced CT images alone was compared with real-world radiologists’ performance based on formal radiologist reports in the hold-out internal test set with regard to sensitivity. Formal radiologist reports reflected routine clinical interpretation of both enhanced and unenhanced CT images, whereas the CAD tool was evaluated exclusively on unenhanced CT. This comparison was intended to provide a pragmatic reference reflecting real-world reporting practice rather than a head-to-head comparison under identical imaging conditions. Because the selection of controls for the internal dataset was based on the absence of PC according to the radiologists’ report, the specificity of radiologists could not be assessed in the internal test set. To further compare with radiologists when only unenhanced CT images were available, 2 board-certified radiologists (radiologist 1: general and chest radiologist, 6 years of experience; radiologist 2: general and abdominal radiologist, 1 year of experience) were requested to interpret the unenhanced CT images of a random subsample of the internal test set and the external test set, and the sensitivity and specificity for PC were compared with those of CAD (details in the Supplementary Material ). The sensitivity of the 2 radiologists for PCs missed by CAD was also evaluated. Evaluation of CAD-indicated key images To evaluate whether the key images could facilitate detection, 2 experienced abdominal radiologists (P.T.C. and K.L.L.) conducted a consensus review to evaluate if the key images determined by the CAD tool included the tumor or secondary signs of PC (dilated pancreatic duct, dilated bile duct). 15 Statistical analysis The performance of the CAD tool in distinguishing PCs from Normals and OPDs was assessed in test sets using a receiver operating characteristic (ROC) curve, as well as sensitivity, specificity, and accuracy with respective exact 95% confidence intervals (CIs). ROC curves were constructed by plotting the sensitivity against the false-positive rate, and the area under the curve (AUC) and the asymptotic 95% CI were calculated. LR was calculated as the probability of having the test result in PCs divided by the probability of having the test result in those without PC. 16 Comparison of sensitivity between CAD for unenhanced CT and radiologist report for unenhanced and enhanced CT was performed using the exact McNemar test. 17 with stratification by tumor size and stage. The primary analysis was the overall comparison of CAD vs radiologist sensitivity in the internal test set. Analyses stratified by tumor size, stage, dataset, and reader were considered exploratory; P -values for these subgroup analyses were treated as descriptive and were not adjusted for multiple comparisons. Categorical and continuous variables were compared between groups using Fisher exact test and Mann-Whitney U test, respectively. Statistical analysis was implemented with R (4.1.3) and Python (3.10.13). P values < .05 were considered statistically significant. Results Dataset characteristics In the internal datasets ( n = 3080), PCs were older than Normals and OPDs with a mean age of 64.6, 55.3, and 61.7 years, respectively. The proportion of males was higher in OPDs (60.6%) compared with Normals (53.9%) and PCs (52.0%) ( Table 1 ). In the external test set ( n = 64), PCs were also older than Normals and OPDs with a mean age of 65.7, 52.5, and 64.2, respectively. The male-to-female ratio was higher in PCs (59.5%) compared with Normals (44.4%) and OPDs (0%). Table 1. Patient characteristics. Training set Validation set Internal test set External test set Variable Cancer Normal OPD Cancer Normal OPD Cancer Normal OPD Cancer Normal OPD No. of Individuals 430 997 424 141 331 141 142 333 141 42 18 4 Age (years) 64.2 ± 12.2 54.8 ± 16.4 61.4 ± 15.6 65.2 ± 11.7 54.9 ± 17.3 62.4 ± 17.5 65.3 ± 11.6 57.1 ± 15.6 61.8 ± 16.2 65.7 ± 8.7 52.5 ± 14.1 64.2 ± 12.9 Sex, n (%) Male 237 (55.1) 515 (51.7) 265 (62.5) 70 (49.6) 178 (53.8) 80 (56.7) 77 (54.2) 170 (51.1) 83 (58.9) 25 (59.5) 8 (44.4) 0 (0.0) Female 193 (44.9) 482 (48.3) 159 (37.5) 71 (50.4) 153 (46.2) 61 (43.3) 65 (45.8) 163 (48.9) 58 (41.1) 17 (40.5) 10 (55.6) 4 (100.0) Stage, n (%) I 26 (6.0) NA NA 8 (5.7) NA NA 9 (6.3) NA NA 15 (35.7) NA NA II 151 (35.1) NA NA 45 (31.9) NA NA 50 (35.2) NA NA 15 (35.7) NA NA III 52 (12.1) NA NA 26 (18.4) NA NA 18 (12.7) NA NA 3 (7.1) NA NA IV 201 (46.7) NA NA 62 (44.0) NA NA 65 (45.8) NA NA 9 (21.4) NA NA Tumor size (cm) 3.2 ± 1.8 NA NA 3.3 ± 1.6 NA NA 3.2 ± 1.7 NA NA 3.3 ± 1.5 NA NA Reference standard Histology/cytology 430 (100.0) 0 5 (1.2) 141 (100.0) 0 2 (1.4) 142 (100.0) 0 0 42 (100.0) 0 4 (100.0) Clinical follow-up 0 997(100.0) 424(100.0) 0 331(100.0) 141(100.0) 0 333(100.0) 141(100.0) 0 18 (100.0) a 0 Follow-up days NA 1460.8 ± 1043.3 1321.2 ± 938.6 NA 1430.3 ± 1030.1 1334.7 ± 900.7 NA 1579.2 ± 1012.1 1387.1 ± 989.6 NA NA >365 b Open in a new tab Continuous variables shown as mean ± SD; categorical variables are shown as numbers with percentages in parentheses. Stage and tumor size are only available in patients with PC. Abbreviations: NA, not applicable; OPD, other pancreatic diseases. a Confirmed by independent image review by 2 additional radiologists. b At least 1 year of follow-up was required to confirm that PC did not develop. Comparison with radiologist interpretation of unenhanced CT images This analysis represents a matched comparison in which both CAD and radiologists (reviewer 1: general and chest radiologist; reviewer 2: general and abdominal radiologist) were restricted to unenhanced CT images. Given the limited number of readers ( n = 2) and substantial inter-reader variability, these results should be considered exploratory. In the subsampled internal test set, the sensitivity of CAD (89.7%) was comparable with reviewer 2 (86.2%) but significantly higher compared with reviewer 1 (41.4%, P < .001), and the difference was mainly observed in tumors smaller than 2 cm (87.5% vs 0%, P = .016) ( Table 2 ). In the 3 PCs missed by CAD, reviewers 1 and 2 diagnosed none and one, respectively. Of the 17 PCs missed by reviewer 1, CAD correctly diagnosed 14; of the 4 PCs missed by reviewer 2, CAD correctly diagnosed 2. No significant difference was noted in specificity between CAD, reviewer 1, and reviewer 2 (90.3%, 100%, 87.1%, respectively). Combining reviewer 1 and CAD yielded 89.7% sensitivity and 90.3% specificity (ie, a case was diagnosed as PC if the diagnosis of either the reviewer or CAD was positive). Combining reviewer 2 and CAD yielded 93.1% sensitivity and 80.6% specificity. Table 2. Performance of CAD tool and radiologist reviewers for unenhanced CT. CAD Reviewer 1 P value Reviewer 1 and CAD a Reviewer 2 P value Reviewer 2 and CAD a Subsampled internal test set Sensitivity (%) 89.7 (72.6-97.8) [26/29] 41.4 (23.5-61.1) [12/29] <.001 89.7 (72.6-97.8) [26/29] 86.2 (68.3-96.1) [25/29] >.999 93.1 (77.2-99.2) [27/29] <2 cm 87.5 (47.3-99.7) [7/8] 0.0 (0.0-36.9) [0/8] .016 87.5 (47.3-99.7) [7/8] 75.0 (34.9-96.8) [6/8] >.999 87.5 (47.3-99.7) [7/8] 2-4 cm 84.6 (54.6-98.1) [11/13] 46.2 (19.2-74.9) [6/13] .062 84.6 (54.6-98.1) [11/13] 84.6 (54.6-98.1) [11/13] >.999 92.3 (64.0-99.8) [12/13] >4 cm 100.0 (63.1-100.0) [8/8] 75.0 (34.9-96.8) [6/8] .500 100.0 (63.1-100.0) [8/8] 100.0 (63.1-100.0) [8/8] >.999 100.0 (63.1-100.0) [8/8] Specificity (%) 90.3 (74.2-98.0) [28/31] 100.0 (88.8-100.0) [31/31] .250 90.3 (74.2-98.0) [28/31] 87.1 (70.2-96.4) [27/31] >.999 80.6 (62.5-92.5) [25/31] Normal 95.0 (75.1-99.9) [19/20] 100.0 (83.2-100.0) [20/20] >.999 95.0 (75.1-99.9) [19/20] 95.0 (75.1-99.9) [19/20] >.999 95.0 (75.1-99.9) [19/20] OPD 81.8 (48.2-97.7) [9/11] 100.0 (71.5-100.0) [11/11] .500 81.8 (48.2-97.7) [9/11] 72.7 (39.0-94.0) [8/11] >.999 54.5 (23.4-83.3) [6/11] External test set Sensitivity (%) 76.2 (60.5-87.9) [32/42] 64.3 (48.0-78.4) [27/42] .227 83.3 (68.6-93.0) [35/42] 73.8 (58.0-86.1) [31/42] >.999 85.7 (71.5-94.6) [36/42] <2 cm 77.8 (40.0-97.2) [7/9] 55.6 (21.2-86.3) [5/9] .625 88.9 (51.8-99.7) [8/9] 66.7 (29.9-92.5) [6/9] >.999 88.9 (51.8-99.7) [8/9] 2-4 cm 66.7 (44.7-84.4) [16/24] 58.3 (36.6-77.9) [14/24] .688 75.0 (53.3-90.2) [18/24] 70.8 (48.9-87.4) [17/24] >.999 79.2 (57.8-92.9) [19/24] >4 cm 100.0 (66.4-100.0) [9/9] 88.9 (51.8-99.7) [8/9] >.999 100.0 (66.4-100.0) [9/9] 88.9 (51.8-99.7) [8/9] >.999 100.0 (66.4-100.0) [9/9] Specificity (%) 86.4 (65.1-97.1) [19/22] 95.5 (77.2-99.9) [21/22] .625 81.8 (59.7-94.8) [18/22] 95.5 (77.2-99.9) [21/22] .625 81.8 (59.7-94.8) [18/22] Normal 88.9 (65.3-98.6) [16/18] 100.0 (81.5-100.0) [18/18] .500 88.9 (65.3-98.6) [16/18] 100.0 (81.5-100.0) [18/18] .500 88.9 (65.3-98.6) [16/18] OPD 75.0 (19.4-99.4) [3/4] 75.0 (19.4-99.4) [3/4] >.999 50.0 (6.8-93.2) [2/4] 75.0 (19.4-99.4) [3/4] >.999 50.0 (6.8-93.2) [2/4] Open in a new tab Data in parentheses are 95% CIs. Data in brackets are numerators and denominators used to calculate percentages. Subgroup analyses are exploratory. P values are unadjusted for multiple comparisons. Abbreviations: CAD, computer-aided diagnosis; OPD, other pancreatic diseases. a Treated as positive if either CAD or reviewer diagnosis is positive. In the external test set, differences between the model and 2 radiologists in sensitivity (64.3%, 73.8%, respectively) and specificity (both 95.5%) were not statistically significant, though the small sample size limited statistical power ( Table 2 ). Reviewers 1 and 2 diagnosed 3 and 4 of the 10 PCs missed by CAD, respectively ( Table S2 ). 8 of the 15 PCs missed by reviewer 1 and 5 of the 11 PCs missed by reviewer 2 were correctly diagnosed by CAD. When combined with CAD, sensitivity numerically increased to 83.3% for reviewer 1 and to 85.7% for reviewer 2, with 81.8% specificity. An exploratory decision curve analysis was performed to assess the potential clinical utility of the CAD tool ( Figure S1 ). Across a range of clinically relevant threshold probabilities, the CAD tool suggested a higher net benefit than the treat-all strategy and reviewer 1, and a net benefit comparable to that of reviewer 2. These findings suggest that the CAD tool may offer clinical utility when applied to unenhanced CT. However, the results should be interpreted with caution, as these reader- and subgroup-level analyses are exploratory. Distinguishing PC from normal pancreas and other pancreatic diseases Because radiologist reports were based on the interpretation of both enhanced and unenhanced CT, while the CAD tool analyzed only unenhanced images, this comparison represents a real-world benchmark rather than a matched assessment. In the internal test set comprising CT images immediately preceding PC diagnosis, the CNN-transformer hybrid model distinguished PCs from Normals and OPDs based on unenhanced CT with 90.8% (95% CI: 84.9-95.0) sensitivity, 93.0% (90.4-95.2) specificity [97.3% (94.9-98.8) for Normals, 83.0% (75.7-88.8) for OPDs], and 92.5% (90.2-94.5) accuracy in the internal hold-out test set [AUC 0.975 (0.957-0.993)]. The sensitivity of the model was not significantly different compared with that based on the formal radiologist reports of unenhanced and enhanced CT images (90.8% vs 95.4%, P = .210) ( Table 3 ). The specificity of radiologists in the internal test set could not be determined because cases identified or suspected by radiologists as having pancreatic tumors were excluded in selecting the control group, thereby eliminating any potential false-positive diagnoses. Table 3. Sensitivity of CAD tool for unenhanced CT and radiologist interpretation of unenhanced and enhanced CT. Internal test set External test set CAD vs Radiologist Sensitivity of CAD (%, unenhanced CT) Sensitivity of CAD (%, unenhanced CT) a Sensitivity of Radiologist (%, unenhanced and enhanced CT) P value Sensitivity of CAD (%, unenhanced CT) Overall 90.8 (84.9-95.0) [129/142] 90.8 (84.4-95.1) [118/130] 95.4 (90.2-98.3) [124/130] .210 76.2 (60.5-87.9) [32/42] Stage I 88.9 (51.8-99.7) [8/9] 100.0 (59.0-100.0) [7/7] 100.0 (59.0-100.0) [7/7] >.999 46.7 (21.3-73.4) [7/15] II 90.0 (78.2-96.7) [45/50] 89.4 (76.9-96.5) [42/47] 93.6 (82.5-98.7) [44/47] .688 93.3 (68.1-99.8) [14/15] III 100.0 (81.5-100.0) [18/18] 100.0 (79.4-100.0) [16/16] 81.2 (54.4-96.0) [13/16] .250 100.0 (29.2-100.0) [3/3] IV 89.2 (79.1-95.6) [58/65] 88.3 (77.4-95.2) [53/60] 100.0 (94.0-100.0) [60/60] .016 88.9 (51.8-99.7) [8/9] Tumor size (cm) <2 87.2 (72.6-95.7) [34/39] 86.8 (71.9-95.6) [33/38] 94.7 (82.3-99.4) [36/38] .375 77.8 (40.0-97.2) [7/9] 2-4 90.6 (80.7-96.5) [58/64] 89.3 (78.1-96.0) [50/56] 96.4 (87.7-99.6) [54/56] .289 66.7 (44.7-84.4) [16/24] >4 94.9 (82.7-99.4) [37/39] 97.2 (85.5-99.9) [35/36] 94.4 (81.3-99.3) [34/36] >.999 100.0 (66.4-100.0) [9/9] Open in a new tab Data in parentheses are 95% CIs. Data in brackets are numerators and denominators used to calculate percentages. Subgroup analyses are exploratory. P values are unadjusted for multiple comparisons. Abbreviation: CAD, computer-aided diagnosis. a Cases without formal radiologist reports were excluded. In the external test set, the model distinguished PCs from Normals and OPDs with 76.2% (60.5-87.9) sensitivity, 86.4% (65.1-97.1) specificity [88.9% (65.3-98.6) for Normals, 75.0% (19.4-99.4) for OPDs], and 79.7% (67.8-88.7) accuracy [AUC 0.892 (0.814-0.969)] ( Figure 3 ). Figure 3. Open in a new tab Receiver operating characteristic curves for differentiation between pancreatic cancer and non-pancreatic cancer. (A) Internal test set. (B) External test set. The AUC with 95% CI is shown for each cohort. Risk stratification and likelihood ratio of pancreatic cancer The CAD tool categorized CT images diagnosed as PC into 4 risk levels, with LRs ranging between 173.46 (24.19-1243.53) for the highest risk level and 1.53 (0.77-3.04) for the lowest level. CT images diagnosed as without PC were stratified into 3 risk levels, with LRs ranging between <0.01 (0.00-0.08) for the lowest risk level and 0.53 (0.23-1.22) for the highest level ( Table 4 ; Figure 4 ). Specific diagnostic confidence provided by the CAD tool. Level-specific LRs could not be analyzed in the external test set because 3 of the 4 risk levels of PC included no patients without PC. Table 4. Risk levels and corresponding likelihood ratios of pancreatic cancer and diagnostic confidence. Validation set Internal test set External test set Diagnosed as without pancreatic cancer by CAD tool Risk level Likelihood ratio based on CAD a Diagnostic confidence PC/Normal/OPD ( n ) Negative predictive value (%) Likelihood ratio PC/Normal/OPD ( n ) Negative predictive value (%) 1 <0.01 (0.00-0.08) Very strong 4/285/82 98.9 (97.3-99.7) 0.04 (0.01-0.10) 4/12/2 77.8 (52.4-93.6) 2 0.24 (0.08-0.78) Moderate 0/16/19 100.0 (90.0-100.0) 0.05 (0.00-0.77) 2/2/0 50.0 (6.8-93.2) 3 0.53 (0.23-1.22) Moderate 9/23/16 81.2 (67.4-91.1) 0.77 (0.38-1.55) 4/2/1 42.9 (9.9-81.6) Diagnosed as pancreatic cancer by CAD tool Risk level Likelihood ratio Diagnostic confidence PC/Normal/OPD ( n ) Positive predictive value (%) Likelihood ratio PC/Normal/OPD ( n ) Positive predictive value (%) 4 1.53 (0.77-3.04) Moderate 8/4/12 33.3 (15.6-55.3) 1.66 (0.73-3.81) 2/0/0 100.0 (15.8-100.0) 5 5.00 (2.61-9.58) Moderate 17/2/5 70.8 (48.9-87.4) 8.09 (3.42-19.11) 8/2/1 72.7 (39.0-94.0) 6 40.03 (14.69-109.09) Strong 53/3/7 84.1 (72.7-92.1) 17.65 (9.22-33.78) 9/0/0 100.0 (66.4-100.0) 7 173.46 (24.19-1243.53) Very strong 51/0/0 100.0 (93.0-100.0) 339.64 (21.09-5469.57) 13/0/0 100.0 (75.3-100.0) Open in a new tab Data in parentheses are 95% CIs. Abbreviations: CAD, computer-aided diagnosis; OPD, other pancreatic diseases; PC, pancreatic cancer. a Likelihood ratios were derived based on the validation set. Figure 4. Open in a new tab Risk Stratification according to the computer-aided diagnosis tool. Patient composition and likelihood ratios of pancreatic cancer across risk levels: (A) Validation set. (B) Internal test set. Patients diagnosed as negative and positive for pancreatic cancer were categorized into 3 and 4 risk levels, respectively. Sensitivity according to tumor size and stage and comparison with formal radiologist reports interpreting unenhanced and enhanced CT images There was generally no significant difference between the sensitivity of CAD for unenhanced CT images and that of formal radiologist reports for enhanced and unenhanced CT images across categories of tumor size and stage, except that radiologists had higher sensitivity than CAD for stage 4 PCs (100% vs 88.3%, P = .016). CAD achieved 87.2% (72.6-95.7) and 88.9% (51.8-99.7) sensitivity for < 2 cm and stage I PCs, respectively. Analyses stratified by tumor size and stage were exploratory, and P -values were not adjusted for multiple comparisons. Of the 6 PCs missed by radiologists, CAD diagnosed 5 (83.3%) and indicated the images containing the tumor as key images ( Figure 5 ). Of the 13 PCs missed by the CAD tool, review of the unenhanced CT images by the 2 radiologists blinded to the diagnosis identified 3 and 8 cases, respectively, as having pancreatic tumor or cancer, whereas neither radiologist detected focal pancreatic lesions in 5 cases ( Table S3 ). In the external test set, CAD achieved 77.8% (40.0-97.2) sensitivity for tumors less than 2 cm ( Table 3 ). Figure 5. Open in a new tab Pancreatic cancer cases not identified by radiologist interpretation but diagnosed by the computer-aided diagnosis tool based on unenhanced images. (A, B, C) A case not identified in the formal radiology report despite interpretation of both unenhanced and contrast-enhanced CT images. (D, E, F) a separate case not identified by radiologists in the reader study when interpretation was restricted to unenhanced CT images. (A) and (D) show key unenhanced CT images highlighted by the CAD tool; (B) and (E) show the corresponding contrast-enhanced CT images demonstrating an indistinct pancreatic mass (arrowheads); (C) and (F) show Grad-CAM (gradient-weighted class activation mapping) visualizations of the model predictions, with heatmaps overlaid on key unenhanced CT images to indicate regions contributing to the model decision. Locating the PC or PC-associated secondary signs by key images Among 129 patients correctly diagnosed as PC by CAD in the internal test set, the key images indicated by the CAD tool included the tumor in 95 (73.6%) and secondary signs of PC (dilated pancreatic or bile duct) in 21 (16.3%), facilitating detection in those 116 (89.9%) patients. Similarly, among the 32 true-positives in the external test set, the key images could facilitate detection in 30 (93.8%) patients, including the tumor in 19 (59.4%) or secondary signs in 11 (34.4%) ( Table S4 ). Among the 33 false-positives in the internal test set ( Figure 6 ), the key images showed an unremarkable pancreas in 17, acute pancreatitis in 7, chronic pancreatitis with calcification in 2, and pancreatic cystic lesion in 2, whereas the key images did not include the pancreas in the remaining 5 cases. Among the 4 false-positives in the external test set, the key images showed an unremarkable pancreas in 2, chronic pancreatitis with calcification in 1, and a focal pancreatic lesion (neuroendocrine tumor) in 1. Figure 6. Open in a new tab Representative false-positive case identified by the CAD tool on unenhanced images. (A) Key unenhanced CT images identified by the tool. (B) Corresponding contrast-enhanced CT images showing a pancreatic cystic lesion (arrowheads). (C) Grad-CAM visualizations of the model predictions. Discussion We developed and externally validated a tool to diagnose PC on unenhanced CT, provide the LR of PC, and flag key images containing tumor or tumor-induced secondary signs to facilitate review. In a hold-out internal test set, the tool demonstrated 90.8% sensitivity, without a significant difference compared to formal radiologist reports interpreting both enhanced and unenhanced CT. Comparison with formal radiologist reports should be interpreted in light of the asymmetry in imaging inputs, as radiologist reports were informed by both enhanced and unenhanced CT, whereas the CAD tool was evaluated exclusively on unenhanced CT. Thus, this likely represents a conservative benchmark. When radiologists were similarly restricted to unenhanced CT in the reader study, the CAD tool demonstrated comparable or higher sensitivity; however, this finding should be interpreted with caution, given the limited number of readers and the substantial inter-reader variability observed. Collectively, these findings support a potential role for the CAD tool in opportunistic detection of PC on unenhanced CT. This tool is not intended to replace contrast-enhanced CT or MRI for definitive diagnosis of PC. Rather, it is designed for opportunistic detection on unenhanced CT examinations performed for unrelated clinical indications. By flagging examinations with imaging patterns suspicious for PC, the CAD tool may prompt timely confirmatory evaluation with dedicated pancreatic imaging. As diagnostic accuracy inevitably decreases with unenhanced CT, this tool additionally provides risk stratification and the LR of PC to convey its diagnostic confidence. LRs can be integrated with clinical context to inform post-test probability and decision-making. This tool further flags key diagnostic images to aid physician interpretation. Together, the risk stratification and key image outputs provide context that may support clinical adoption by improving transparency of the AI analysis. This study had notable strengths. While prior research has demonstrated the potential of DL in detecting PCs on enhanced CT, 9 , 12 its application to unenhanced CT remains limited 6 , 18 , 19 due to the difficulties delineating the pancreas (ie, segmentation) and differentiating normal from abnormal pancreatic tissues without contrast enhancement, which are compounded by coexisting non-pancreatic diseases/conditions. 5 Rather than attempting to segment the pancreas for analysis, the approach located the upper abdomen, including the pancreas and surrounding structures, which may manifest tumor-induced secondary changes (eg, metastasis, dilated bile duct, lymphadenopathies) for analysis. The upper abdomen was analyzed by combining CNNs to capture local/granular features and transformers to utilize distant/global contextual clues for disease classification. This approach offered potential analytic advantages over using either CNN or transformer alone and further avoided the challenges of pancreas segmentation on unenhanced CT. A previous study employed DL to segment the pancreas and diagnose PC on unenhanced CT with promising results, 6 but it excluded patients with conditions that often alter the imaging manifestations of the pancreas and surrounding structures and hence could impact segmentation and disease classification, including acute pancreatitis, abdominal cancer treatment, severe ascites, and low imaging quality. Excluding acute pancreatitis could also affect generalizability. 20 Last, providing both qualitative and quantitative diagnostic information along with key images may help clinicians better understand and integrate AI analytics into decision-making. This study also had limitations. Radiologists’ sensitivity for PC might have been overestimated. Whereas CAD provided a definitive binary diagnosis, radiologists did not always confirm whether the diagnoses were PC when reporting confirmed/suspected pancreatic tumors. To avoid underestimating radiologists’ sensitivity, we considered radiologists as having diagnosed the PC if the reports described or suspected pancreatic tumors, regardless of whether PC was stated. In addition, the use of radiologist reports as part of the reference standard for defining non-cancer controls may introduce incorporation bias and potentially overestimate specificity. Although non-cancer cases were further verified through clinical records and cancer registry follow-up to exclude subsequent PC diagnoses, specificity estimates should nevertheless be interpreted with caution. Second, given the wide variability in radiologists’ diagnostic performance on unenhanced CT observed in our study, further Consequently, the LRs presented in this study should be regarded as exploratory and internally derived, and the wide confidence intervals observed in the external test set should be interpreted with caution. Although an exploratory decision-curve analysis was performed, validation in larger and more diverse cohorts is required to confirm the robustness of the LRs and the clinical utility of the model. In addition, neither the external validation cohort nor the reader study was based on a formal sample size or power calculation and thus the findings should be interpreted cautiously. Lastly, this study did not evaluate the clinical impact of the CAD tool. Although the model showed diagnostic performance comparable to that of radiologists, its value is most likely as an assistive system rather than a stand-alone tool. Future studies should assess its impact on radiologist performance and clinical workflow. In conclusion, this AI-based CAD tool detected PCs on unenhanced CT, provided risk stratification with post-test probability estimates, and flagged key images containing diagnostic clues. This tool may address the unmet clinical need of opportunistic PC detection on unenhanced CT. Supplementary Material umag017_Supplementary_Data umag017_supplementary_data.zip (1.4MB, zip) Acknowledgments The authors thank Nico Wu for assistance with data management and preprocessing, and Y.-C.C. and P.-J.L. for imaging interpretation. Contributor Information Po-Ting Chen, Department of Medical Imaging, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan. Dawei Chang, Data Science Degree Program, National Taiwan University and Academia Sinica, Taipei, 106319, Taiwan. Yenjia Chen, Data Science Degree Program, National Taiwan University and Academia Sinica, Taipei, 106319, Taiwan. Pochuan Wang, Department of Computer Science and Information Engineering, National Taiwan University, Taipei, 106319, Taiwan. Andre Yanchen Yeh, Department of Medicine, National Taiwan University College of Medicine, Taipei, 100233, Taiwan. Kao-Lang Liu, Department of Medical Imaging, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan. Ming-Shiang Wu, Division of Gastroenterology and Hepatology, Department of Internal Medicine, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan; Internal Medicine, College of Medicine, National Taiwan University, Taipei, 100233, Taiwan. Wei-Chih Liao, Division of Gastroenterology and Hepatology, Department of Internal Medicine, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, 100233, Taiwan; Internal Medicine, College of Medicine, National Taiwan University, Taipei, 100233, Taiwan. Weichung Wang, Institute of Applied Mathematical Sciences, National Taiwan University, Taipei, 106319, Taiwan. Author contributions Po-Ting Chen (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing—original draft), Dawei Chang (Data curation, Formal analysis, Methodology, Software, Validation, Writing—original draft), Yenjia Chen (Data curation, Formal analysis, Methodology, Software, Validation), Pochuan Wang (Data curation, Methodology, Software, Validation), Andre Yanchen Yeh (Data curation, Methodology, Software, Validation), Kao-Lang Liu (Investigation, Supervision, Validation), Ming-Shiang Wu (Investigation, Resources, Supervision, Validation), Wei-Chih Liao (Funding acquisition, Methodology, Project administration, Supervision, Visualization, Writing—review & editing), and Weichung Wang (Funding acquisition, Project administration, Resources, Supervision, Validation, Visualization, Writing—review & editing) Supplementary material Supplementary material is available at Radiology Advances online. Funding This study was supported by the National Science and Technology Council (NSTC-113-2123-M-002-017- and NSTC-114-2123-M-002-010-). The funding source had no role in study design, data collection, analysis, interpretation, report writing, or the decision to submit this paper for publication. Conflicts of interest The authors declare that they have no competing financial interests or personal relationships that could have influenced the research in this study. W.C.L. holds issued and pending patents and has equity in PanCAD.ai. W.W. holds issued and pending patents and has equity in PanCAD.ai. Data availability Computer codes are available online ( https://github.com/davidzan/PCDoUCTwDL ). References 1. Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA Cancer J Clin. 2023;73(1):17-48. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Kang JD, Clarke SE, Costa AF. Factors associated with missed and misinterpreted cases of pancreatic ductal adenocarcinoma. Eur Radiol. 2020;31(4):2422-2432. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Nadeem M, Ather MH, Jamshaid A, Zaigham S, Mirza R, Salam B. Rationale use of unenhanced multi-detector CT (CT KUB) in evaluation of suspected renal colic. Int J Surg Ovid Technol (Wolters Kluwer Health). 2012;10(10):634-637. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Shaish H, Ream J, Huang C, et al. Diagnostic accuracy of unenhanced computed tomography for evaluation of acute abdominal pain in the emergency department. JAMA Surg Am Med Assoc. 2023;158(7):e231112. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Tummala P, Junaidi O, Agarwal B. Imaging of pancreatic cancer: An overview. J Gastrointest Oncol. 2011;2(3):168-174. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Cao K, Xia Y, Yao J, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med. 2023;29(12):3033-3043. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Zins M. Invited commentary: Failures in imaging diagnosis of pancreatic cancer: can we avoid them and how? Radiographics. 2023;43(11):e230217. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC. Early detection of pancreatic cancer on computed tomography: advancements with deep learning. Radiol Adv. 2025;2(5):umaf028. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Liu K-L, Wu T, Chen P-T, et al. Deep learning to distinguish pancreatic cancer tissue from non-cancerous pancreatic tissue: a retrospective study with cross-racial external validation. Lancet Digit. Health. 2020;2(6):e303-e313. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Chen P-T, Chang D, Yen H, et al. Radiomic features at CT can distinguish pancreatic cancer from noncancerous pancreas. Radiol Imaging Cancer. 2021;3(4):e210010. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Mukherjee S, Patra A, Khasawneh H, et al. Radiomics-based machine-learning models can detect pancreatic cancer on prediagnostic computed tomography scans at a substantial lead time before clinical diagnosis. Gastroenterology. 2022;163(5):1435-1446.e3. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Chen P-T, Wu T, Wang P, et al. Pancreatic cancer detection on CT scans with deep learning: a nationwide population-based study. Radiology. 2023;306(1):172-182. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Chang D, Chen P-T, Wang P, et al. Detection of pancreatic cancer with two- and three-dimensional radiomic analysis in a nationwide population-based real-world dataset. BMC Cancer. 2023;23(1):58. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. The Clinical Proteomic Tumor Analysis Consortium Pancreatic Ductal Adenocarcinoma Collection (CPTAC-PDA). The Cancer Imaging Archive. Accessed July 25, 2022. 10.7937/k9/tcia.2018.sc20fo18 [ DOI ] 15. Chu LC, Goggins MG, Fishman EK. Diagnosis and detection of pancreatic cancer. Cancer J. 2017;23(6):333-342. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Deeks JJ, Altman DG. Diagnostic tests 4: likelihood ratios. BMJ. 2004;329(7458):168-169. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Zhou X-H, McClish DK, Obuchowski NA. Statistical Methods in Diagnostic Medicine. John Wiley & Sons; 2009. [ Google Scholar ] 18. Ren S, Zhao R, Zhang J, et al. Diagnostic accuracy of unenhanced CT texture analysis to differentiate mass-forming pancreatitis from pancreatic ductal adenocarcinoma. Abdom Radiol (NY). 2020;45(5):1524-1533. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Ren S, Tang H-J, Zhao R, Duan S-F, Chen R, Wang Z-Q. Application of unenhanced computed tomography texture analysis to differentiate pancreatic adenosquamous carcinoma from pancreatic ductal adenocarcinoma. Curr Med Sci. 2022;42(1):217-225. [ DOI ] [ PubMed ] [ Google Scholar ] 20. Miller FH, Lopes Vendrami C, Hammond NA, Mittal PK, Nikolaidis P, Jawahar A. Pancreatic cancer and its mimics. Radiographics. 2023;43(11):e230054. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials umag017_Supplementary_Data umag017_supplementary_data.zip (1.4MB, zip) Data Availability Statement Computer codes are available online ( https://github.com/davidzan/PCDoUCTwDL ). Articles from Radiology Advances are provided here courtesy of Oxford University Press ACTIONS View on publisher site PDF (1.8 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top