A clinician-centric intelligent method towards reliable pancreatic cancer vascular invasion assessment: a retrospective, multi-centre study.

A clinician-centric intelligent method towards reliable pancreatic cancer vascular invasion assessment: a retrospective, multi-centre study - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide 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 Lancet Reg Health West Pac . 2026 May 30;71:101890. doi: 10.1016/j.lanwpc.2026.101890 A clinician-centric intelligent method towards reliable pancreatic cancer vascular invasion assessment: a retrospective, multi-centre study Rui Guo a, j , Mengyao Zhang a, j , Hongzhang Zhu b, j , Xinlu Tang a , Wenli Fu a , Minhong Li c , Huijun Hu d , Xiang Xiao e , Weishen Wang f , Ying-Bin Liu g, h, i , Shi-Ting Feng b, ∗ , Tao Chen g, ∗∗ , Xiaohua Qian a, ∗∗∗ Author information Article notes Copyright and License information a School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China b Department of Radiology, The First Affiliated Hospital, Sun Yat-sen University, Guangzhou, China c Department of Radiology, The Second Affiliated Hospital of Guangzhou Medical University, Guangzhou, China d Department of Radiology, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou, China e Department of Medical Imaging, Nanfang Hospital, Southern Medical University, Guangzhou, China f Department of General Surgery, Pancreatic Disease Center, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China g Department of Biliary-Pancreatic Surgery, Renji Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China h Shanghai Key Laboratory of Systems Regulation and Clinical Translation for Cancer, Shanghai, China i State Key Laboratory of Systems Medicine for Cancer, Shanghai Cancer Institute, Shanghai, China ∗ Corresponding author. fengsht@mail.sysu.edu.cn ∗∗ Corresponding author. dr_chentao78@163.com ∗∗∗ Corresponding author. xiaohua.qian@sjtu.edu.cn j Joint first authors. Received 2026 Jan 7; Revised 2026 May 12; Accepted 2026 May 14; Collection date 2026 Jun. © 2026 The Author(s) This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). PMC Copyright notice PMCID: PMC13241961  PMID: 42256486 Summary Background Pancreatic cancer is highly aggressive, with high post-resection recurrence and poor survival. Accurate preoperative assessment of vascular invasion is essential but remains clinically unmet. This study aimed to enable reliable automated assessment of vascular invasion using routine computed tomography (CT). Methods We developed CRVIA, a clinician-centric artificial intelligence (AI) method that encodes clinical expertise into invasion-omics and emulates clinician’s decision-making through causality-enhanced modelling. The model was trained on 1251 cases and validated on internal (1104 cases) and external (3575 cases from five centres) cohorts. Reader studies involving eight radiologists were conducted to evaluate the efficacy of CRVIA’s assistance. Primary outcome measures included classification performance, feature distance, and inter-reader agreement. Statistical analyses were conducted using paired t -tests, Mann–Whitney U, Pearson’s chi-squared, and permutation tests. Findings The study cohort comprised 5930 cases from 2062 patients (median age 63.0 years [IQR 56.0–70.0]; 41.1% female). CRVIA achieved AUCs of 0.946 (95% CI 0.929–0.962) internally and 0.943 (95% CI 0.932–0.952) externally, outperforming nine comparison approaches spanning radiomics to foundation models. Performance remained stable across stages, vessel types, and segmentation variations. In reader studies, CRVIA exceeded senior radiologists’ accuracy, elevated junior radiologists to expert-level performance, and substantially improved inter-reader agreement. Interpretation This study presents a reliable, open-source AI tool (code: https://github.com/SJTUBME-QianLab/CRVIA ) for accurate, stable, and interpretable vascular invasion assessment in pancreatic cancer, which potentially supports precise treatment, encourages collaborative research, and ultimately benefits patients, especially for resource-limited regions. Funding National Natural Science Foundation of China and Shanghai Natural Science Foundation. Keywords: Pancreatic cancer, Vascular invasion assessment, Computed tomography, Invasion-omics, Artificial intelligence Research in context. Evidence before this study Preoperative assessment of vascular invasion in pancreatic cancer, crucial for determining resectability, has traditionally relied on subjective interpretation of routine CT scans using established morphological criteria (e.g., tumour–vessel contact, contour irregularity). We searched Google Scholar from inception to 31 December 2025 for studies published in English using the terms (“pancreatic cancer” OR “pancreatic ductal adenocarcinoma”) AND (“vascular invasion” OR “vessel invasion”) AND (“computed tomography” OR “CT”) AND (“artificial intelligence” OR “deep learning” OR “radiomics”). Regrettably, dedicated AI systems for this specific task are nearly absent. While radiomics has been applied to broader pancreatic cancer characterisation, its use for vascular invasion assessment remains limited, and such approaches are notably sensitive to imaging protocol variations. To date, no deep learning model—and certainly no clinician-centric AI framework—has been developed or validated for the fully automated, interpretable assessment of tumour-vessel invasion from preoperative CT. This clear unmet need highlights the urgency for a robust and clinically transparent tool to standardise this critical surgical decision. Added value of this study This study develops and validates an AI method for tumour–vessel invasion assessment in pancreas using routinely used CT. The proposed method, CRVIA, is a clinician-centric technologically transformative AI method to deliver clinically feasible and trustworthy assessments. Additionally, we curated a nationwide, large representative CT image dataset—comprising 5930 tumour–vessel cases from 2062 patients to validate it. CRVIA achieved superior performance (AUC up to 0.943 on external validation across centres) and maintained across stages, vessel units, and varied imaging preprocessing, establishing solid foundation for real-world clinical deployment. It outperformed multiple advanced approaches, even advanced large-scale foundation models. Crucially, in reader studies, CRVIA’s interpretable outputs elevated junior radiologists’ performances to expert-level and significantly improved inter-reader agreement, revealing clinical feasibility in real-world practice. Implications of all the available evidence Based on the evidence presented, CRVIA demonstrates that its clinician-centric AI paradigm—which explicitly encodes domain expertise into model design—can achieve expert-level, robust performance for the subtle and critical task of preoperative vascular invasion assessment in pancreatic cancer. Therefore, it can reduce reliance on subjective experience and provide guidance to clinicians across expertise levels, thereby holding promise for lowering barriers to precision pancreatic cancer care. Future work will focus on prospective validation of its impact on clinical workflows and patient outcomes. Introduction Pancreatic cancer is one of the most aggressive malignancies, characterised by particularly poor prognosis, with a post-resection recurrence rate above 80% 1 and a five-year survival rate of 13%. 1 , 2 Improving patient outcomes hinges on precise treatment decisions, which require accurate vascular invasion assessment. 3 Regrettably, there is no effective preoperative clinical approach or assistive strategy. Currently, artificial intelligence (AI) has made significant strides in cancer diagnosis and treatment planning. 4 , 5 , 6 Here, we propose a clinician-centric AI method for automated tumour–vessel invasion assessment in pancreas using computed tomography (CT), a modality routinely used in clinical practice. This method provides reliable assessment with clinically interpretable evidence for treatment decision-making. According to the National Comprehensive Cancer Network (NCCN) guidelines, surgical resection is the only curative option for pancreatic cancer. 7 Vessel invasion status directly affects surgical intervention decision regarding tumour resectability and vascular reconstruction necessity. 7 Consequently, overestimating invasion delays necessary surgery, while underestimating it exposes patients to unnecessary, non-beneficial high-risk operations, both ultimately undermining the chances of survival. 8 Thus, accurate preoperative assessment of vascular invasion in pancreatic cancer is essential to patient-specific precise treatment plans, maximise benefits, and minimise harms. The NCCN surgical guidelines grade vascular invasion by the extent of the tumour–vessel contact, 3 and clinical studies have further established additional criteria with corresponding imaging cues, such as vascular deformation and contour irregularity. 9 , 10 , 11 However, current image-based methods rely only on visual, subjective experience-based cues, thus often fail to differentiate between visually similar but pathologically distinct subtle conditions, 12 such as vessel narrowing caused by invasion versus mere compression ( Fig. 1 a). This challenge highlights the critical need for an imaging-based AI system to objectively and automatedly assess tumour–vessel invasion in pancreas. Fig. 1. Open in a new tab Overview of this study. ( a ) Challenges of vascular invasion assessment in pancreatic cancer based on routinely used CT. Imaging-based preoperative assessment is essential for precise treatment yet faces major challenges in differentiating visually similar but pathologically distinct cases with subtle and complex patterns. ( b ) Conceptual diagram for our proposed Clinician-centric Reliable Vascular Invasion Assessment (CRVIA) method. This method represents a paradigm shift towards clinician-centric AI as it distills intelligence cues from clinical expertise, association logic, and diagnostic pathways of clinicians, and in turn offers stable high performance with clinician-centric interpretability. ( c ) Dataset. This study curated a nationwide, large representative dataset including 5930 tumour–vessel cases from 2062 patients of six medical centres. ( d ) CRVIA’s stability across diverse clinical conditions, covering six medical centres, four tumour–node–metastasis (TNM) stages, three clinically critical vessel units, and varied imaging preprocessing. ( e ) CRVIA’s clinician-centric interpretability in reader studies. CRVIA matches and even surpasses senior experts, helps junior radiologists approach senior-expert performance, and improves inter-reader agreement, underscoring its clinical potential for pancreatic cancer. CA = celiac axis. SMA = superior mesenteric artery. SMV = superior mesenteric vein. SV = splenic vein. PV = portal vein. Subtle invasion-related CT imaging differences—arising from microscopic vessel-wall changes—and limited sample sizes—owing to the rarity of pancreatic cancer and the difficulty of obtaining invasion labels (in contrast to other tasks with hundreds of thousands of samples)—impede current AI methods achieving clinically feasible, stable performance, despite their strength in implicit pattern learning for objective assessment. Moreover, the data-driven nature of such universal AI often misaligns with clinical expertise knowledge, so clinicians often cannot understand model outputs, do not trust them, and therefore cannot employ them in patient care. 13 , 14 Therefore, we developed a clinician-centric AI method that embeds prior clinical knowledge into a deep-learning framework, aiming at capturing subtle differences, improving assessment performance, and providing interpretable evidence. Clinician-centric intelligence encompasses three aspects: reliance on routine clinical imaging, stability across diverse clinical settings, and provision of interpretable and actionable evidence aligned with clinical expertise knowledge. This study proposes Clinician-centric Reliable Vascular Invasion Assessment (CRVIA), an AI method based on routine CT that assesses the invasion status of each tumour–vessel pair, which affects different treatments or surgery schemes for a patient. CRVIA’s clinician-centric paradigm—reflected in stability and interpretability—is achieved by integrating both clinical expertise knowledge (encoded as invasion-omics) and AI (implemented as a causality-enhanced graph neural network). CRVIA comprises three technical components ( Fig. 1 b): invasion-omics characterisation derived from clinical expertise, causal decision modelling driven by association logic, and hierarchical representation refinement driven by diagnostic pathways. We curated a nationwide, large representative CT image dataset ( Fig. 1 c)—comprising 5930 tumour–vessel cases from 2062 patients—for vascular invasion assessment in pancreatic cancer, covering six major medical centres, four tumour–node–metastasis (TNM) stages, and three clinically critical vessel units across five vessel types. Leveraging on this dataset, CRVIA was systematically evaluated—from heterogeneous imaging acquisition and preprocessing to expert-level verification. It demonstrated stable performance across centres, achieving an area under the receiver operating characteristic (ROC) curve (AUC) of up to 0.943 (95% confidence interval (CI) 0.932–0.952) externally, and maintained such stable high performance across all stages, vessel units, and even varied imaging preprocessing, establishing solid foundation for real-world clinical deployment ( Fig. 1 d). Moreover, CRVIA substantially outperformed nine technical approaches that are popular in medical field, including radiomics-based approaches, vision transformers, and advanced large-scale foundation models. Notably, in reader studies involving eight radiologists, CRVIA matched or even surpassed senior experts; and its clinician-centric interpretability fosters clinician understanding and trust, helping junior radiologists approach senior-expert performance and improving inter-reader agreement, thereby empowering clinicians to perform more accurate assessments ( Fig. 1 e). To sum up, this study proposes a novel clinician-centric AI method leveraging routinely used CT, exhibiting robust performance across diverse clinical conditions while providing interpretable clinical evidence. This design lays the foundation for integrating CRVIA into a future human-in-the-loop workflow, in which clinicians can review CRVIA’s evidence, accept reliable findings, override discrepant results, and contextualise assessments within the full clinical picture. Such a method has the potential to enable automated tumour–vessel invasion assessment, which is expected to reduce long-standing reliance on experience and provide guidance to clinicians across different levels of expertise. Furthermore, this advancement holds promise for lowering barriers to precision pancreatic cancer care, ultimately contributing to the broader development of extensive and intelligent precision medicine in pancreatic cancer. Methods Ethics approval This study was approved by the institutional ethics committees of the participating centres, with a waiver of informed consent for anonymised data: Clinical Research and Laboratory Animal Ethics Committee, the First Affiliated Hospital of Sun Yat-sen University ([2024]467); Renji Hospital Ethics Committee, Shanghai Jiao Tong University School of Medicine (RA-2021-094); Ruijin Hospital Ethics Committee, Shanghai Jiao Tong University School of Medicine (2020-234); Medical Ethics Committee of Sun Yat-sen Memorial Hospital of Sun Yat-sen University (SYSKY-2025-031-01); Medical Ethics Committee of Nanfang Hospital of Southern Medical University (NFEC-2024-608); and Clinical Research and Application Ethics Committee of the Second Affiliated Hospital of Guangzhou Medical University (2024-hg-ks-27). Retrospective analyses were conducted exclusively on anonymised samples, with no requirement for informed consent. Study design The objective of this study was to develop and evaluate CRVIA, a CT-based clinician-centric AI method for preoperative vascular invasion assessment in pancreatic cancer. The study design and model details were decided beforehand with no post-hoc changes, although an explicit protocol was not prepared. This study was not registered. There was no patient or public involvement during the design, conduct, reporting, interpretation, or dissemination of the study. This study adhered to the TRIPOD + AI reporting guidelines. Patients and the public were not involved in the design, conduct, reporting, or dissemination of this study. Participants and datasets This study retrospectively collected the largest CT image dataset to date for vascular invasion assessment in pancreatic cancer from six participating medical centres ( Fig. 2 ), comprising 5930 tumour–vessel cases from 2062 patients. Invasion labels were derived from intraoperative pathology or independent radiologist consensus (see Appendix p 3 for details). This dataset includes four stages and three vessel units (covering five key peripancreatic vessels) that inform different surgical decisions 15 , 16 : (1) celiac axis (CA)—determines resectability for pancreatic body/tail tumours; (2) superior mesenteric artery (SMA)—severe invasion contraindicates resection; (3) venous complex, including anatomically continuous superior mesenteric vein (SMV), splenic vein (SV), and portal vein (PV)—invasion often permits resection with reconstruction. Given these distinct clinical implications, each vessel is assessed separately in practice; and accordingly, we treat each tumour–vessel pair as an individual case. Fig. 2. Open in a new tab Data flow for this study of vascular invasion in pancreatic cancer. The curated dataset comprises three cohorts: a development cohort, an internal validation cohort, and an external validation cohort. Specifically, the development cohort (Centre-1-train) contains 1251 tumour–vessel cases from 417 patients at the First Affiliated Hospital of Sun Yat-sen University (FASY, Centre-1) collected between December 2016 and September 2021. The internal validation cohort (Centre-1-test) consists of 1104 cases from 369 patients collected at the same centre between October 2021 and December 2024. The external validation cohort comprises 3575 cases from 1276 patients, contributed by five participating centres: (1) Renji Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (Renji, Centre-2), (2) Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (Ruijin, Centre-3), (3) Sun Yat-sen Memorial Hospital of Sun Yat-sen University (SYM, Centre-4), (4) Nanfang Hospital of Southern Medical University (NHS, Centre-5), and (5) the Second Affiliated Hospital of Guangzhou Medical University (SAHG, Centre-6). Details of each centre are summarised in Table 1 , while the balance between the development and validation cohorts is reported in Supplementary Table S1 ( Appendix p 21). Table 1. Data characteristics and CT acquisition parameters. Centre name Centre-1-train Centre-1-test Centre-2 Centre-3 Centre-4 Centre-5 Centre-6 FASY FASY Renji Ruijin SYM NHS SAHG Vessel (invasion/non-invasion) 1251 (297/954) 1104 (231/873) 2617 (294/2323) 496 (203/293) 240 (48/192) 135 (22/113) 87 (31/56) Indistinct vessel (invasion/non-invasion) 579 (261/318) 479 (180/299) 960 (242/718) 301 (190/111) 101 (44/57) 51 (19/32) 36 (24/12) Patient 417 369 874 248 80 45 29 Female, Num. (%) 168 (40.3%) 152 (41.2%) 364 (41.6%) 100 (40.3%) 35 (43.8%) 18 (40.0%) 10 (34.5%) Age (IQR) 61.0 (53.0–68.0) 63.0 (54.0–69.0) 65.0 (59.0–72.0) 62.5 (56.0–69.0) 61.5 (53.5–67.0) 58.0 (50.0–67.0) 67.0 (57.0–73.0) Surgery, Num. (%) 241 (57.8%) 204 (55.3%) 598 (68.4%) 173 (69.8%) 42 (52.5%) 29 (64.4%) 18 (62.1%) TNM stage, Num. I 81 82 424 1 8 20 6 II 146 109 133 159 31 9 12 III 108 100 225 59 30 8 8 IV 82 78 92 29 11 8 3 Contrast phase Venous Venous Venous Venous Venous Venous Venous Slice spacing (mm) 0.299–1.25 0.699–1.25 0.599–1.6 0.5–5.0 0.699–1.0 1.0–1.5 0.8–1.6 Average pixel spacing (mm) 0.521–0.851 0.561–0.9375 0.566–0.958 0.576–0.976 0.554–0.976 0.578–0.888 0.578–0.832 Manufacturer Toshiba, Philips, GE, Siemens Siemens, Philips, Toshiba, GE, Canon Siemens, Toshiba, GE, UIH, Philips Philips, GE, Siemens GE, Siemens, UIH UIH, Philips, GE, NMS, Siemens GE, Toshiba, Philips Reconstruction kernel Br40d 3, B, FC03, FC08, Bv40d 3, FC18, FC13 Br40d 3, B, FC03, FC08, STANDARD, FC18, BODY B31f, FC08, STANDARD, B, Bf40d 2, Bf40d 1, B_SOFT_A IMR1 Soft Tissue, B, STANDARD, IMR1 Routine, B30f STANDARD, SOFT, Br40d 3, B_SOFT_B Br40d 3, B, STANDARD, F20, B_SOFT_B, IMR1 Routine STANDARD, FC12, FC08,B Tube voltage (kVp) 80/90/100/110/120/140 80/90/100/110/120/140 100–120 120/140 80/90/100/120 120 100/120 Tube current (mAs) 98–2116 95–1296 102–640 66–640 63–1052 107–590 60–431 Spiral pitch factor 0.638–3.0 0.6–3.0 0.6–1.375 0.6–1.375 0.9–1.375 0.6–0.9937 0.8–0.992 Open in a new tab Female or male denotes sex assigned at birth. Num. = number. IQR = median value with interquartile range. FASY = The First Affiliated Hospital of Sun Yat-sen University. Renji = Renji Hospital Affiliated to Shanghai Jiao Tong University School of Medicine. Ruijin = Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine. SYM = Sun Yat-sen Memorial Hospital of Sun Yat-sen University. NHS = Nanfang Hospital of Southern Medical University. SAHG = The Second Affiliated Hospital of Guangzhou Medical University. This study included patients who met the following criteria: (1) histopathologically confirmed pancreatic ductal adenocarcinoma located in the pancreatic head (including the uncinate process), neck, body, or tail; (2) availability of preoperative contrast-enhanced CT imaging acquired using a standardised pancreatic CT protocol and of sufficient quality for tumour–vessel invasion assessment; and (3) availability of definitive vascular invasion status determined by standardised clinical protocols (intraoperative exploration and pathological margin assessment for surgical patients; expert consensus-based imaging evaluation for non-surgical patients). Patients were excluded if they met any of the following conditions: (1) a history of pancreatic inflammation or neoplasms, or abdominal surgery resulting in altered or disrupted pancreatic anatomy; (2) CT imaging of low quality, including but not limited to severe motion artefacts or metallic implant-related artefacts, such as biliary stents; (3) contraindications to iodinated contrast agents, such as a history of severe allergic reactions or renal insufficiency (serum creatinine > 2.0 mg/dL); (4) an interval exceeding one month between the initial CT examination and curative resection; or (5) lack of a definitive tumour–vessel invasion label, including but not limited to: unavailable intraoperative exploration records or pathological reports for surgical patients; incomplete MDT documentation for non-surgical patients. Evaluation strategy To systematically evaluate CRVIA’s assessment performance and clinical utility, we devised a four-dimensional evaluation strategy: (1) Performance on the internal validation cohort: Binary classification metrics were calculated on the internal validation cohort from the development centre (i.e., Centre-1-test), both overall and stratified by TNM stage. (2) Stability across centres on the external validation cohort: Binary classification metrics were used to investigate whether CRVIA maintains consistently high performance across diverse conditions. An independent multi-centre external validation cohort, spanning five leading hospitals in different regions, was employed to evaluate performance across centres and vessel units (CA, SMA, and Venous). (3) Stability across varied imaging preprocessing: We introduced small variations in tumour–vessel segmentation arising from manual refinements applied after automated segmentations, reflecting differences in annotator experience and knowledge. Stability was evaluated at two levels: decision-level, indicated by minimal variability in classification metrics, and feature-level, reflected by small distances between features derived from each segmentation and those from the raw segmentation. (4) Clinician-centric interpretability in reader studies: Eight radiologists, including both senior and junior, performed assessments initially and then reperformed with CRVIA assistance; we examined changes in assessment performance and inter-reader agreement. Methodology of CRVIA We proposed CRVIA, a preoperative assessment method based on routinely-used CT for vascular invasion in pancreatic cancer. CRVIA comprises two primary stages: (1) invasion-omics characterisation that encodes clinical expertise knowledge and experience, and (2) representation-based decision that emulates the decision-making framework of human experts. Methodological details are provided in Appendix pp 3–9, with involved variables and workflow in Supplementary Figs. S1 and S2 ( Appendix p 14). In the invasion-omics characterisation stage, clinical expertise knowledge and experience are encoded into a tailored invasion-omics representation, delineating four key imaging cues: (1) vessel boundary clarity, (2) tumour–vessel contact relationships, (3) morphological alterations of tumours and vessels, and (4) tumour thrombus presence, as illustrated in Supplementary Fig. S3 ( Appendix p 15). This design ensures the stability of invasion-omics features across heterogeneous imaging acquisition protocols and varied preprocessing operations (tumour and vessel segmentations), and those features provide interpretable evidence aligned with clinical expertise knowledge. In the representation-based decision stage, unified reasoning is achieved through causal decision modelling driven by association logic and hierarchical representation refinement driven by diagnostic pathways. Specifically, a causality-enhanced graph neural network (GNN) is developed to extract stable causal patterns by simulating clinicians’ multi-evidence association logic, thus improving assessment stability across heterogeneous data. Additionally, a hierarchical decision-tree-guided strategy is designed to distinguish between tumour–vessel interaction subclasses by simulating layered diagnostic pathways, thus recognizing subtle and complex invasion patterns and further enhancing stability, as presented in Supplementary Fig. S4 ( Appendix p 15). Ultimately, by integrating clinical expertise knowledge and diagnostic reasoning with GNN, CRVIA enhances its ability to capture subtle differences and generate clinically interpretable evidence. Statistical analysis Binary classification performance for tumour–vessel invasion assessment was evaluated using AUC, area under the precision–recall curve (AUPR), accuracy (ACC), balanced accuracy (BACC), 17 sensitivity (SEN), and specificity (SPE), where AUC was calculated using a deterministic non-parametric approach based on ranking and the trapezoidal rule, implemented via the metrics.roc_auc_score function in the Scikit-learn library. 18 To evaluate CRVIA’s stability across varied imaging preprocessing, Euclidean distances were calculated between features derived from each segmentation variant and those from the raw segmentation. Cohen’s kappa coefficient κ, reflecting inter-reader agreement, was computed using the metrics.cohen_kappa_score function in the Scikit-learn library. 18 Confidence intervals were estimated via 1000 bootstrap replications for most analyses. For experiments with small sample sizes—specifically the internal evaluation across the four TNM stages—confidence intervals were instead calculated from binomial probabilities via Wilson’s method. Paired t -tests were used to compare CRVIA with comparison approaches and to evaluate performance gains in radiologists with versus without CRVIA assistance. For assessing clinical characteristic balance, continuous variables were analysed using the Mann–Whitney U tests, and categorical variables using Pearson’s chi-squared tests. Permutation tests were performed to evaluate performance differences across subgroups stratified by potential confounders. Role of the funding source The funders of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. Results Performance on the internal validation cohort The assessment performance of CRVIA for tumour–vessel invasion was first evaluated on the internal validation cohort (Centre-1-test) of 1104 tumour–vessel cases. As shown in Fig. 3 a and b, CRVIA demonstrated excellent assessment performance, achieving an AUC of 0.946 (95% CI 0.929–0.962) and an accuracy of 0.912 (95% CI 0.894–0.928). Further subgroup analyses across the four TNM stages showed CRVIA’s performance stability, with AUCs ranging from 0.937 (95% CI 0.903–0.959) to 0.999 (95% CI 0.982–1.000) and accuracies from 0.903 (95% CI 0.865–0.932) to 0.923 (95% CI 0.883–0.950) ( Fig. 3 a and c). Additionally, analyses in Supplementary Fig. S5 ( Appendix p 16) showed no statistically significant differences in AUC across subgroups stratified by age, sex, TNM stage, and CT manufacturer. These findings verify that CRVIA achieves stable high assessment performance across all stages, thereby establishing a dependable foundation for clinical deployment. Fig. 3. Open in a new tab Evaluation of internal and external validation cohorts. ( a–c ) Internal validation of CRVIA (Centre-1: n = 1104). It contains four TNM stages (Stage I: n = 246, Stage II: n = 325, Stage III: n = 300, Stage IV: n = 233). ( a ) ROC curve. ( b ) ACC, BACC, SEN, and SPE of Centre-1. ( c ) ACC, BACC, SEN, and SPE across four TNM stages. ( d and e ) External validation of CRVIA across five centres (Centre-2: n = 2,617, Centre-3: n = 496, Centre-4: n = 240, Centre-5: n = 135, Centre-6: n = 87). ( d ) ROC curve. ( e ) ACC, BACC, SEN, and SPE. ( f–h ) Performance comparison of CRVIA versus other approaches (n = 3575), where ROC curve and AUPR are not applicable to two approaches (“Rad + DT” and “Rad + SVM”) due to their binary outputs. ( f ) ROC curve, where statistical comparisons between CRVIA and other approaches across five centres were performed using one-sided paired t -tests (n = 5), with details in Supplementary Table S2 ( Appendix p 21). ( g ) AUPR. ( h ) ACC of each centre and the average across five centres. CRVIA achieves stable high performance across all stages and centres, outperforming all comparison approaches. ( b , c , e , g ) Error bars indicate 95% CI and bar centres represent the mean value of the metric. ( f , g ) See Appendix pp 10–12 for curve legend and x-axis labels. Stability across centres on the external validation cohort To evaluate CRVIA’s generalizability across different medical centres, we performed independent testing on the external multi-centre validation cohort (3575 cases from Centre-2 through Centre-6). CRVIA achieved an outstanding overall AUC of 0.943 (95% CI 0.932–0.952), despite considerable heterogeneity among centres in imaging conditions, such as acquisition protocols and scanner types ( Table 1 ). Performance across the five centres remained consistently high (AUCs: 0.939–0.991, accuracies: 0.865–0.943; Fig. 3 d and e), and no significant AUC differences were observed across age, sex, TNM stage, and CT manufacturer subgroups in Supplementary Fig. S6 ( Appendix p 17), even for hard-to-assess indistinct cases characterised by high tumour–vessel proximity ( Appendix p 9 and p 18). This result further confirmed CRVIA’s stability under challenging conditions, where tumour–vessel boundary is blurred and invasion or non-invasion appear visually similar. Ablation studies ( Appendix p 12 and p 18) showed each of CRVIA’s three key technical components contributed significantly to its high performance, supporting the reasonableness of CRVIA’s technical scheme. These results collectively highlight CRVIA’s strong generalizability across widespread medical centres, thereby showing potential for real-world clinical deployment. To show CRVIA’s comparative advantages, we further conducted wide comparison against representative methods. Due to the lack of readily available automated models for tumour–vessel invasion assessment, we implemented nine comparison approaches spanning three paradigms, from conventional to advanced ones: (1) radiomics-based 19 , 20 features with classifiers; (2) widely-used deep learning models 21 , 22 ; and (3) state-of-the-art large-scale foundation models 23 , 24 (see Appendix pp 10–12 for details). CRVIA achieved a significantly higher overall AUC (0.943, 95% CI 0.932–0.952), AUPR (0.732, 95% CI 0.689–0.773) and accuracy (0.903, 95% CI 0.894–0.912) than all comparison approaches ( Fig. 3 f and g and Appendix p 18) and maintained superior accuracy across every individual centre ( Fig. 3 h). Overall, in comparison with existing approaches, CRVIA exhibited high performance across diverse centres. We further evaluated CRVIA’s performance across key vessel units (CA, SMA, and venous complex; see Fig. 1 c for spatial arrangement) in the external cohort. CRVIA achieved high AUCs for arterial (CA: 0.963, 95% CI 0.902–0.994; SMA: 0.970, 95% CI 0.946–0.986) and venous (0.859, 95% CI 0.837–0.879) units, with balanced accuracy, sensitivity, and specificity ( Fig. 4 a and b). Notably, it maintained high sensitivity for arteries (CA: 0.810, 95% CI 0.625–0.962; SMA: 0.828, 95% CI 0.721–0.918) despite their lower invasion incidence, underscoring reliability in challenging low-prevalence scenarios. In contrast, other methods showed marked instability across vessel units ( Fig. 4 c). Overall, CRVIA consistently outperformed all comparison approaches across all vessel units, demonstrating its superior performance in anatomically complex, multi-vessel scenarios critical for guiding distinct treatment implications across vessel units in pancreatic cancer. Fig. 4. Open in a new tab Stability validation across different vessel units and varied preprocessing. ( a ) ROC curve of CRVIA across three vessel units. CA: n = 1,023, SMA: n = 1,276, Venous: n = 1276. ( b ) ACC, BACC, SEN, and SPE of CRVIA across three vessel units. ( c ) AUC comparison of CRVIA versus other approaches across three vessel units, where AUC is not applicable to two approaches (“Rad + DT” and “Rad + SVM”) due to their binary outputs. CRVIA achieves stable high performance across all vessel units and consistently outperforms all comparison approaches. ( b and c ) Error bars indicate 95% CI and bar centres represent the mean value of the metric. ( d ) ACC, BACC, SEN, and SPE of CRVIA for different tumour–vessel segmentation variants (n = 496). Error bars indicate 95% CI and bar centres represent the mean value of the metric. ( e ) Euclidean distances between features obtained from each segmentation and those from the raw segmentation. CRVIA’s invasion-omics features (“Ours”) were compared with the Radiomics features. Each box plot shows the median (centreline), upper and lower quartiles (box limits), 1.5 × interquartile range from the upper and lower quartiles (whiskers), and outliers (points). CRVIA maintains stable high performance and stable invasion-omics features across varied preprocessing. Stability across varied imaging preprocessing In real-world clinical settings, AI models remain sensitive to variability introduced by heterogeneous image preprocessing. Segmentation of tumours and vessels is a key preprocessing step for tumour–vessel invasion assessment in pancreas. After automated segmentation, manual refinement is typically required to improve segmentation precision, yet inevitably introduces small fluctuations that may undermine segmentation consistency. To assess the stability of CRVIA under such variabilities, we designed three types of preprocessing involving six sets of segmentation variants ( Appendix pp 9–10): (1) raw automated segmentations, (2) expert-refined versions of those raw segmentations (used as the reference for all other experiments), and (3) the expert-refined segmentations perturbed via four morphological transformations designed to emulate common inter-clinician annotation differences. All evaluations were conducted on the external validation cohort from Centre-3. CRVIA’s AUCs ranged from 0.929 (95% CI 0.904–0.953, on raw segmentations) to 0.949 (95% CI 0.929–0.968, on reference segmentations) with a span of 0.020 ( Fig. 4 d), demonstrating its stable high performance against varied preprocessing. At the feature level, invasion-omics features also exhibited minimal distortion across segmentation conditions ( Fig. 4 e): using features derived from raw segmentations as a baseline, the Euclidean distance to other segmentation variants remained low (median values 0.108–0.210); by contrast, conventional radiomic features exhibited substantially greater variability (median values 0.116–0.653), highlighting the superior stability of CRVIA’s invasion-omics characterisation derived from clinical expert. In summary, CRVIA exhibits strong stability for both feature extraction and diagnosis, effectively mitigating the effects of varied imaging preprocessing. Therefore, CRVIA has significant potential for real-world clinical deployment, since it accommodates variability introduced during preprocessing (such as segmentation tools and annotators) while maintaining reproducible, stable performance. Clinician-centric interpretability in reader studies To assess the clinical feasibility of CRVIA in real-world applications, we conducted two structured reader studies where radiologists evaluated cases first without and then with CRVIA assistance. A total of 65 patients were randomly selected from the internal validation cohort (Centre-1-test), yielding 77 indistinct tumour–vessel cases. The study had two main validation objectives: (1) to compare CRVIA’s performance with radiologists with different levels ( Fig. 5 b), and (2) to examine whether CRVIA with interpretability can truly assist them, including improving performance of less experienced junior radiologists ( Fig. 5 b) and enhancing inter-reader agreement ( Fig. 5 c). Fig. 5. Open in a new tab Results of reader studies. ( a ) An example tumour–vessel case (clinical label: invasion) in the second reader study. It presents the original CT image, 2D and 3D segmentation masks, along with the invasion-omics features and final invasion assessment generated by CRVIA. A structural disruption level of 1.57 and low vessel boundary clarity values (0.071 and 0.053) indicate invasion due to vessel narrowing and boundary blurring, respectively, consistent with CT evidence of vascular invasion. ( b ) ACC, SEN, and SPE of four junior radiologists, four senior radiologists, their AI-assisted versions (“+AI”), and CRVIA (n = 77), where statistical comparisons between independent (“Solo”) and AI-assisted (“+AI”) assessments were performed using one-sided paired t -tests (n = 8), with details in Supplementary Table S3 ( Appendix p 21). Due to the clinician-centric interpretability, CRVIA reaches (and even surpasses) senior-level performance and can improve both junior and senior radiologists. ( c ) Inter-reader agreement (Cohen’s kappa coefficient κ) between different radiologists without and with CRVIA assistance. All pairwise agreement increased, most substantially among junior radiologists. Eight certified radiologists from Centre-1—a leading medical centre in a major metropolitan area—participated in the reader studies: four juniors (abdominal imaging fellows in their first or second year of training) and four seniors (attending radiologists with > 10 years’ experience). In the first reader study without AI assistance, each radiologist independently assessed invasion status for case based on original CT images, blinded to the invasion labels and other readers’ assessments. In the second reader study, radiologists re-assessed each case after receiving a structured AI-assisted report (following a four-week washout period to mitigate recall bias). An AI-assisted report comprises the original CT image, 2D and 3D segmentation masks, and information generated by CRVIA, including invasion-omics features and the final invasion assessment ( Fig. 5 a). For example, in Supplementary Fig. S10a ( Appendix pp 19–20), a structural disruption level of 1.57 and low vessel boundary clarity values (0.071 and 0.053) indicate invasion due to vessel narrowing and boundary blurring, respectively, consistent with CT evidence of vascular invasion. By contrast, Supplementary Fig. S10e , despite compression-related narrowing similar to Supplementary Fig. S10a , shows non-invasion inferred by CRVIA: higher vessel boundary clarity and zero tumour–vessel contact, with a clear fat plane maintaining separation. Supplementary Fig. S10 display and explain additional representative examples spanning typical conditions, such as narrowed vessel, blurred boundary, tumour thrombus, and non-visualised vessel branch, demonstrating that CRVIA’s invasion-omics features align with established imaging manifestations. Together, these results show that invasion-omics features closely align with clinical expertise knowledge and accumulated experiences, thereby earning clinician’s trust to facilitate assessment revision. The resulting AI-induced changes are summarised in Fig. 5 b and c. CRVIA achieved an accuracy of 0.805, surpassing both junior (0.649–0.714) and senior (0.714–0.779) radiologists by 12.7–24.0% and 3.3–12.7%, respectively. These findings confirm that CRVIA can reach (and even surpass) senior-level performance with high-quality automated assessments. CRVIA’s assistance boosted the accuracies of four junior radiologists by 0.078, 0.091, 0.117, and 0.078, respectively—bringing them in line with senior-level performance. Senior radiologists’ accuracies also improved by 0.052, 0.104, 0.078, and 0.065, respectively. These results underscore that CRVIA improves performance for both inexperienced and highly experienced clinicians, which is attributable to its reliable conclusion with interpretable evidence that aligns with clinical expertise knowledge. Further analysis of the assessment agreement revealed that, inter-reader agreement among junior radiologists (Cohen’s kappa coefficient, denoted by κ) increased from 0.278–0.513 to 0.615–0.812, and agreement between junior and senior radiologists improved from 0.116–0.690 to 0.637–0.920 ( Fig. 5 c). These results indicate that CRVIA functions as a valuable cognitive aid to help less experienced clinicians approach expert-level performance and accelerate the development of their clinical skills. Overall, these reader studies confirmed the strong assessment performance and clinician-centric interpretability of CRVIA, and highlighted its effectiveness in assisting clinicians, thereby revealing clinical feasibility in real-world practice. Discussion Vascular invasion assessment is a crucial preoperative decision in pancreatic cancer, directly determining resectability and guiding treatment planning. 25 , 26 We propose CRVIA, a novel clinician-centric AI method that initially enables automated reliable assessment of tumour–vessel invasion in pancreas using only routine CT scans. CRVIA exhibits clinically desirable stability across a large-scale multi-centre cohort with diverse conditions, while offering clinician-centric interpretability, thereby holding promise for filling a pivotal clinical gap in pancreatic cancer. Despite its critical importance, the assessment of tumour–vessel invasion in pancreatic cancer has long posed a challenge for both clinical practice and medical AI. Tumour invasion often begins as subtle disruptions or infiltrations of the vessel wall that are difficult to discern on CT, as tumour and vessel wall exhibit similar attenuation, blurred boundaries and obscured minute signs. Methodologically, distinguishing such visually similar yet pathologically distinct states is inherently difficult, and further complicated by centre- and protocol-dependent imaging heterogeneity. Moreover, the rarity of pancreatic cancer, combined with the limited availability of valuable invasion annotations, yields substantially smaller labelled datasets than those available for common tasks (e.g., thoracic disease diagnosis via chest radiography 27 ), restricting the capacity of data-driven AI models to achieve stable and high performance. Together, these factors represent the fundamental obstacles that CRVIA is designed to overcome. This study addresses these limitations through two key characteristics. Methodologically, CRVIA moves beyond purely data-driven learning to emulate clinical diagnostic workflow within AI framework. It formalises clinical expertise, association logic, and diagnostic pathways, representing a paradigm shift from purely data-driven AI to clinician-centric intelligence. At the data level, our development of CRVIA is supported by the largest cohort for this task to date (5930 cases from six centres). Embedding clinical knowledge not only guides the extraction of subtle discriminative features, but also reduces reliance on large datasets alone, making the current cohort afford effective and reliable model development. Beyond these two key characteristics, the invasion-omics within the clinician-centric paradigm of CRVIA constitutes a distinctive innovation of this study. Traditionally, radiomics transforms medical images into geometric or statistical descriptors, offering a degree of objectivity. However, radiomic descriptors lack domain-specific medical semantics compared to expert-defined clinical imaging signs, 9 , 28 , 29 and rely heavily on precise image preprocessing vulnerable to inter-centre imaging heterogeneities, thereby constraining its clinical utility. By contrast, CRVIA quantifies and objectifies traditional qualitative knowledge of tumour–vessel relationships, redefines radiomics to invasion-omics through directly embedding clinical expertise into feature design to generate clinically meaningful and robust evidence. Leveraging its methodological characteristics and distinctive innovation, CRVIA delivers two critical advantages, as follows: Firstly, CRVIA demonstrated clinically desirable stability across medical centres, TNM stages, and vessel unit, which is an essential prerequisite for clinical deployment. In external validation across five major centres with heterogeneous image qualities, it achieved an overall AUC of 0.943 (95% CI 0.932–0.952), markedly outperforming other state-of-the-art approaches ( Fig. 3 f). Subgroup analyses further confirmed that its accuracy remained high across different TNM stages and vessel units ( Figs. 3 and 4 ). Notably, CRVIA maintained high performance even under varied image preprocessing within standardised workflows ( Fig. 4 ). Overall, these findings demonstrate CRVIA’s consistently high and stable performance, providing a solid foundation for scalable AI-assisted accurate assessment of vascular invasion in pancreatic cancer across diverse healthcare settings. Secondly, CRVIA provides clinician-centric interpretability, another indispensable clinical requirement for building trust and facilitate adoption only if their decision mechanisms are transparent and explainable. CRVIA directly encodes high-quality clinical expertise knowledge into AI modelling, transforming implicit AI reasoning into explicit and interpretable decision evidence. Each CRVIA-generated assessment is anchored to specific clinically meaningful cues, providing clinicians with a fully traceable and verifiable rationale. In reader studies, this interpretability allowed clinicians to understand and verify CRVIA’s predictions, achieving close alignment with expert judgment—for example, by highlighting subtle blurring of boundary or minor morphology alterations. This embedded clinical knowledge enables CRVIA to match or even surpass senior clinicians’ performance, while interpretable evidence guides junior clinicians to improve their assessment accuracy and inter-reader agreement. Consequently, CRVIA is expected to advance AI-assisted accurate preoperative assessment of vascular invasion within an understandable and trustworthy paradigm. In summary, CRVIA is a promising clinically deployable AI method to advance vascular invasion assessment in pancreatic cancer. By providing reliable assessments, it can facilitate more precise clinical decisions regarding resectability, vascular reconstruction, and adjuvant therapy, while still depends on other factors such as patient physical status and lymph node metastasis. 25 CRVIA’s improvements in clinical decision-making may ultimately benefit patient survival, even as outcomes remain influenced by surgical strategy, operative expertise, and the quality of perioperative care. 26 Hence, further investigation is warranted for decision-making and patient survival. This study utilised venous-phase CT, widely available in clinical practice, to maximise deployment potential, and future work will explore multi-modal CT–MRI fusion and multi-phase CT integration to enhance vascular invasion assessment, potentially addressing current failure cases ( Appendix p 20 and p 22). Additionally, although CRVIA is based on clinically motivated causal assumptions, incorporating formal causal inference—such as explicit causal graphs and deconfounding algorithms—could further strengthen theoretical robustness and interpretability. This work aims to offer a feasible pathway for collaborative advances in pancreatic cancer: we release open-source code to narrow expertise gaps, expand access to precise care (particularly for resource-limited regions), encourage further research, and ultimately benefit patients worldwide. Contributors R.G., X.T., and X.Q. contributed to the conception and design of the work. H.Z., M.L., H.H., X.X., W.W., Y.L. S.F., and T.C. contributed to the data acquisition and organisation. M.Z., H.Z., W.F., M.L., H.H., X.X., W.W., Y.L., S.F., and T.C. carried out the data preprocessing. R.G., M.Z., X.T., and W.F. contributed to the technical implementation and evaluation. M.Z. and X.Q. contributed to technical discussions. H.Z. and S.F. conducted the reader studies. R.G., M.Z., H.Z., X.T., W.F., W.W., Y.L., S.F., T.C., and X.Q. wrote and revised the manuscript. R.G., H.Z., X.T., T.C., and X.Q. contributed to the data access and verification. X.Q. supervised the study, and S.F., T.C., and X.Q. had final responsibility for the decision to submit for publication. All authors reviewed and approved the manuscript. Data sharing statement The proposed method was mainly implemented in Python using PyTorch, and the source code is publicly available at https://github.com/SJTUBME-QianLab/CRVIA . We utilised retrospective data collected from six major medical centres for model development and evaluation. Due to patient privacy regulations, institutional review board approvals, and data use agreement constraints, these data are not currently publicly available. A de-identified example dataset ( https://doi.org/10.6084/m9.figshare.32090647 ) sufficient to run the analysis scripts is publicly available in the code repository and via the Figshare link referenced therein. Declaration of interests The authors declare no competing interests. Acknowledgements This work was supported by the National Natural Science Foundation of China (Grant No. 62371286, 82471127, and 82573548), and was also sponsored by the Shanghai Natural Science Foundation (Grant No. 24ZR1444100). Footnotes Appendix A Supplementary data related to this article can be found at https://doi.org/10.1016/j.lanwpc.2026.101890 . Contributor Information Shi-Ting Feng, Email: fengsht@mail.sysu.edu.cn. Tao Chen, Email: dr_chentao78@163.com. Xiaohua Qian, Email: xiaohua.qian@sjtu.edu.cn. Appendix A. Supplementary data Appendix mmc1.pdf (3.5MB, pdf) References 1. Stoop T.F., Javed A.A., Oba A., et al. Pancreatic cancer. Lancet. 2025;405(10485):1182–1202. doi: 10.1016/S0140-6736(25)00261-2. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Siegel R.L., Kratzer T.B., Giaquinto A.N., Sung H., Jemal A. Cancer statistics, 2025. CA Cancer J Clin. 2025;75(1):10–45. doi: 10.3322/caac.21871. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Hong S.B., Lee S.S., Kim J.H., et al. Pancreatic cancer CT: prediction of resectability according to NCCN criteria. Radiology. 2018;289(3):710–718. doi: 10.1148/radiol.2018180628. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Pan X., AbdulJabbar K., Coelho-Lima J., et al. The artificial intelligence-based model ANORAK improves histopathological grading of lung adenocarcinoma. Nat Cancer. 2024;5(2):347–363. doi: 10.1038/s43018-023-00694-w. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Liang J., Zhang W., Yang J., et al. Deep learning supported discovery of biomarkers for clinical prognosis of liver cancer. Nat Mach Intell. 2023;5(4):408–420. [ Google Scholar ] 6. Gao Y., Ventura-Diaz S., Wang X., et al. An explainable longitudinal multi-modal fusion model for predicting neoadjuvant therapy response in women with breast cancer. Nat Commun. 2024;15(1):9613. doi: 10.1038/s41467-024-53450-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Tempero M.A., Malafa M.P., Al-Hawary M., et al. Pancreatic adenocarcinoma, version 2.2021, NCCN clinical practice guidelines in oncology. J Natl Compr Cancer Netw. 2021;19(4):439–457. doi: 10.6004/jnccn.2021.0017. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Park W., Chawla A., O’Reilly E.M. Pancreatic cancer: a review. JAMA. 2021;326(9):851–862. doi: 10.1001/jama.2021.13027. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Rigiroli F., Hoye J., Lerebours R., et al. CT radiomic features of superior mesenteric artery involvement in pancreatic ductal adenocarcinoma: a pilot study. Radiology. 2021;301(3):610–622. doi: 10.1148/radiol.2021210699. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Noda Y., Mizuno N., Kawai N., et al. Determination of arterial invasion in pancreatic ductal adenocarcinoma: what is the best diagnostic criterion on CT? Eur Radiol. 2023;33(5):3617–3626. doi: 10.1007/s00330-023-09521-3. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Zhou Y., Wang J., Zhang S.-L., et al. A CT radiomics-based risk score for preoperative estimation of intraoperative superior mesenteric-portal vein involvement in pancreatic ductal adenocarcinoma. Ann Surg Oncol. 2023;30(2):1206–1216. doi: 10.1245/s10434-022-12664-0. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Giannone F., Capretti G., Hilal M.A., et al. Resectability of pancreatic cancer is in the eye of the observer: a multicenter, blinded, prospective assessment of interobserver agreement on NCCN resectability status criteria. Ann Surg Open. 2021;2(3) doi: 10.1097/AS9.0000000000000087. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Reddy S. Explainability and artificial intelligence in medicine. Lancet Digit Health. 2022;4(4):e214–e215. doi: 10.1016/S2589-7500(22)00029-2. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Kundu S. AI in medicine must be explainable. Nat Med. 2021;27(8):1328. doi: 10.1038/s41591-021-01461-z. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Zaky A.M., Wolfgang C.L., Weiss M.J., Javed A.A., Fishman E.K., Zaheer A. Tumor-vessel relationships in pancreatic ductal adenocarcinoma at multidetector CT: different classification systems and their influence on treatment planning. Radiographics. 2017;37(1):93–112. doi: 10.1148/rg.2017160054. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Schneider M., Hackert T., Strobel O., Büchler M.W. Technical advances in surgery for pancreatic cancer. Br J Surg. 2021;108(7):777–785. doi: 10.1093/bjs/znab133. [ DOI ] [ PubMed ] [ Google Scholar ] 17. García V., Mollineda R.A., Sánchez J.S. Iberian conference on pattern recognition and image analysis; 2009. Springer; 2009. Index of balanced accuracy: a performance measure for skewed class distributions; pp. 441–448. [ Google Scholar ] 18. Pedregosa F., Varoquaux G., Gramfort A., et al. Scikit-learn: machine learning in python. J Mach Learn Res. 2011;12:2825–2830. [ Google Scholar ] 19. Zwanenburg A., Vallières M., Abdalah M.A., et al. The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology. 2020;295(2):328–338. doi: 10.1148/radiol.2020191145. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Mayerhoefer M.E., Materka A., Langs G., et al. Introduction to radiomics. J Nucl Med. 2020;61(4):488–495. doi: 10.2967/jnumed.118.222893. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. He K., Zhang X., Ren S., Sun J. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2016. 2016. Deep residual learning for image recognition; pp. 770–778. [ Google Scholar ] 22. Dosovitskiy A., Beyer L., Kolesnikov A., et al. International Conference on Learning Representations; 2021. 2021. An image is worth 16x16 words: transformers for image recognition at scale. [ Google Scholar ] 23. He Y., Guo P., Tang Y., et al. Proceedings of the IEEE/CVF international conference on computer vision and pattern recognition; 2025. 2025. VISTA3D: a unified segmentation foundation model for 3D medical imaging; pp. 20863–20873. [ Google Scholar ] 24. Pai S., Bontempi D., Hadzic I., et al. Foundation model for cancer imaging biomarkers. Nat Mach Intell. 2024;6(3):354–367. doi: 10.1038/s42256-024-00807-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Stoop T.F., Theijse R.T., Seelen L.W., et al. Preoperative chemotherapy, radiotherapy and surgical decision-making in patients with borderline resectable and locally advanced pancreatic cancer. Nat Rev Gastroenterol Hepatol. 2024;21(2):101–124. doi: 10.1038/s41575-023-00856-2. [ DOI ] [ PubMed ] [ Google Scholar ] 26. Strobel O., Neoptolemos J., Jaeger D., Buechler M.W. Optimizing the outcomes of pancreatic cancer surgery. Nat Rev Clin Oncol. 2019;16(1):11–26. doi: 10.1038/s41571-018-0112-1. [ DOI ] [ PubMed ] [ Google Scholar ] 27. Ma D., Pang J., Gotway M.B., Liang J. A fully open AI foundation model applied to chest radiography. Nature. 2025;643:488–498. doi: 10.1038/s41586-025-09079-8. [ DOI ] [ PubMed ] [ Google Scholar ] 28. Liu L., Wang Q., Adeli E., Zhang L., Zhang H., Shen D. Exploring diagnosis and imaging biomarkers of Parkinson’s disease via iterative canonical correlation analysis based feature selection. Comput Med Imag Graph. 2018;67:21–29. doi: 10.1016/j.compmedimag.2018.04.002. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Slavkova K.P., Kang R., Kazerouni A.S., et al. MRI-based radiomic features for risk stratification of ductal carcinoma in situ in a multicenter setting (ECOG-ACRIN E4112 Trial) Radiology. 2025;315(1) doi: 10.1148/radiol.241628. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Appendix mmc1.pdf (3.5MB, pdf) Articles from The Lancet Regional Health: Western Pacific are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (4.4 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases 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

Related documents

• Locally Advanced or Metastatic Pancreatic Adenocarcinoma#1086891
• Biomarkers in Predicting Response to Treatment in Patients Who Have Undergone Surgery for Pancreatic Cancer#1140131
• PanMETAI - a high performance tabular foundation model for accurate pancreatic cancer diagnosis via NMR metabolomics#1097449
• Randomized Trial for Optimal Number of Passes Required for Molecular Profiling During EUS-FNB of Pancreatic Cancer#907977
• Research of the Application of Pancreatic Cancer Screening Artificial Intelligence Model "PANDAPro"#876639