PubMed, Embase, and the Cochrane Library databases were systematically searched for studies published until 30 June 2025. The search was limited to English-language publications. A detailed search strategy was tailored for each database, using a combination of Medical Subject Headings (MeSH) terms and free-text keywords. The search strategy included the following subject headings and abstract words: artificial intelligence, radiomics, deep learning, machine learning, pancreas, pancreatic cancer, pancreatic adenocarcinoma, mass-forming pancreatitis, autoimmune pancreatitis, and focal chronic pancreatitis. The search strategy was: (“radiomics” OR “deep learning” OR “machine learning” OR “artificial intelligence”) AND (“pancreatic cancer” OR “pancreatic adenocarcinoma”) AND (“mass-forming pancreatitis” OR “autoimmune pancreatitis” OR “focal chronic pancreatitis”). Boolean operators (“AND”, “OR”) and proximity operators were used to broaden the search and capture variations in terminology. The complete electronic search strategies for all databases, including MeSH terms, keywords, Boolean operators, and date limits, are provided in Supplementary Table S1 for transparency and reproducibility. The strategy was adjusted to accommodate each database’s specific indexing systems and syntaxes. Following deduplication, two investigators independently screened titles and abstracts. Articles deemed potentially relevant underwent full-text review against predefined inclusion criteria. Studies satisfying these criteria were incorporated into qualitative synthesis and meta-analyses. Additionally, backward citation tracking of included articles and related systematic reviews was performed to identify supplementary evidence. Throughout the screening procedure, the corresponding author was consulted in order to settle any disputes among the authors on research selection.

The inclusion criteria were designed based on the following PICOS criteria to ensure comprehensive and unbiased searches as follows:

Exclusion criteria were as follows: (a) studies involving patients who had received neoadjuvant therapy (including immunotherapy, radiotherapy, or chemotherapy) prior to diagnostic imaging; (b) studies failing to provide extractable data for reconstructing 2×2 contingency tables or explicit diagnostic validation protocols; (c) non-original research publications (case reports, systematic reviews, editorials, and letters) and conference abstracts without full methodological details; and (d) investigations exclusively focusing on technical validation phases such as image segmentation algorithms or radiomic feature extraction pipelines.

Two reviewers (H.Y. Zhang and J. Lu) independently extracted the data from the included studies. The following data were extracted from the eligible studies: (a) study characteristics (the first author, year of publication, country, and study design); (b) lesion characteristics and numbers of PC or pancreatitis; (c) AI algorithms and subtype (DL or ML); (d) images modalities (CT, MRI, ECT, and PET/CT); (e) the reference criteria for pancreatitis and PDAC (histopathological or clinical diagnoses); and (f) the diagnostic accuracy data including true-positive (TP), false-positive (FP), true-negative (TN), and false-negative (FN) were extracted directly into contingency tables, which were used to calculate SE and SP. If a study provided multiple contingency tables for the same or different AI algorithms, we assumed that they were independent of each other.

The quality of the reviewed studies was evaluated using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) criteria. The tool assesses four key domains: “Patient Selection”, “Index Test”, “Reference Standard”, and “Flow and Timing”. For each domain, two reviewers independently evaluated the risk of bias (high/low/unclear) and applicability concerns (high/low). Discrepancies were resolved through discussion or consultation with a third researcher.

Statistical analyses were conducted by using STATA 17.0 (StataCorp LLC, College Station, Texas, USA) using the bivariate random-effects model. Coupled forest plots were generated to visually present pooled sensitivity and specificity. The pooled sensitivity and specificity of the differential diagnosis of pancreatitis and PC, and their 95% confidence intervals (CIs) were calculated using the relevant data extracted from each individual study. The sensitivity was calculated by dividing the number of patients who had a PC diagnosis by the total number of patients with PC, and the specificity was calculated by dividing the number of patients who had an MPF by the total number of MPF patients. I² values were computed to assess statistical heterogeneity, categorized as very low (0%–25%), low (25%–50%), medium (50%–75%), and high (>75%) [19,20]. Publication bias was determined by visual assessment of Deeks’ funnel plot, and statistical significance was evaluated by Deeks’ asymmetry test. p < 0.05 was considered statistically significant.

If significant heterogeneity presented, a subgroup analysis was conducted to investigate the cause of heterogeneity. The following covariates were used in the subgroup analysis: (a) subtype of AI algorithms (ML vs. DL); (b) imaging modalities (US, CT, MRI, or PET); (c) according to the pooled performance using the same dataset (AI algorithms or human clinicians); (d) according to the geographical distribution (Asia or non-Asia); (e) number of institutions from which data were obtained (single or multiple centers); (f) patient numbers (patients ≤100 vs. patients >100); (g) publication year (before 2020 vs. after 2020); (h) studies were categorized as low risk or high/unclear risk of bias, based on the overall judgment across the four domains (patient selection, index test, reference standard, and flow/timing); and (i) different AI algorithm (LASSO vs. SVM vs. ANNs).

A sensitivity analysis was performed by excluding studies with high or unclear risk of bias to evaluate the influence of lower-quality studies on the overall pooled results. The pooled sensitivity, specificity, diagnostic odds ratio (DOR), and area under the summary receiver operating characteristic curve (AUC) were recalculated for the remaining studies The results were compared with the primary analysis to assess the stability and robustness of the findings.

A total of 408 studies were retrieved during the initial search. After removing 119 duplicates, 135 studies were excluded based on title and abstract screening, leaving 94 studies eligible for full-text review. Ultimately, 28 articles were included in the systematic review, all of which provided sufficient data for meta-analysis. The final analysis incorporated data from 3,279 patients. A total of 76 contingency tables, including TP, FP, TN, and FN, were extracted from the 28 eligible studies (3, 24, 25, 27–49). The complete details of the literature search and screening process are illustrated in Figure 1 and Supplementary Table S1.

The included studies were published between 2001 and 2025, with a notable concentration of 20 publications (71.4%) between 2019 and 2023. The majority of the studies (n = 17) were conducted in China, while the remaining studies were carried out in America, Japan, Romania, Germany, and India. Among the included studies, 25 employed retrospective designs, and 3 utilized prospective designs. Twenty-two studies were conducted at a single center, while six involved multiple centers. Eight studies compared the performance of AI models with clinicians using the same dataset. Twenty-four studies explicitly excluded low-quality images, whereas the remaining studies did not mention this process.

The imaging modalities used in the studies were categorized as follows: CT (n = 11), MRI (n = 4), US (n = 9), and PET/CT (n = 4). Twenty-three studies analyzed retrospectively collected data, while the remaining three studies used prospectively collected data. In terms of AI algorithms, the distribution was as follows: ML (25 studies) and DL (3 studies). Eight studies compared the performance of AI with that of expert clinicians. Detailed characteristics of the included studies are presented in Tables 1, 2 and Supplementary Tables S2, S3.

The quality of included studies was determined by the QUADAS-2 (Supplementary Figure S1). The detailed assessment results are presented in Supplementary Figure S2. Over half of the studies showed a high risk or an unclear risk of bias respectively for patient selections (n = 16) because these studies did not clearly describe patient characteristics, including prior testing, clinical presentation, study setting, intended use of the index test, or external evaluation. The risk of bias in the index test was low in 13 studies (46%) and high or unclear in 15 studies (54%). The risk of bias within the domains of reference standard tests was consistently low across all studies. Moreover, 15 studies remained unclear in the flow and timing domain, while 13 studies were low in this regard. The analysis revealed a potential risk of bias in the domain of patient selection, with the absence of randomization serving as a primary contributing factor. Overall, the included literature was deemed suitable for subsequent analyses in light of its quality.

The summary receiver operating characteristic (SROC) curves for the 28 included studies, comprising 76 contingency tables, are presented in Figures 2, 3. The pooled sensitivity (SEN) and specificity (SPE) were 89% (95% CI: 87%–90%) and 88% (95% CI: 86%–90%), respectively, with an AUC of 0.94 (95% CI: 0.92–0.96) for all AI algorithms. When the highest accuracy contingency table was selected from these 28 studies ( Table 3), the pooled SEN and SPE were 91% (95% CI: 88%–93%) and 90% (95% CI: 87%–93%), respectively, with an AUC of 0.96 (95% CI: 0.94–0.97), as shown in Figures 2, 4. However, substantial heterogeneity was observed among the included studies, with SEN having an I² = 77.14% and SPE having an I² = 75.61%.

Funnel plots indicated no significant publication bias among the 28 included studies with 76 contingency tables (p = 0.18) and when the highest accuracy contingency table was selected (p = 0.70), as illustrated in Figures 5a, b.

The detailed results of the subgroup analyses and exploration of potential sources of between-study heterogeneity are shown in Table 4 and Supplementary Figures S3–S20. Considering the developmental stage and inherent differences of the algorithms, we categorized them into ML and DL groups and performed subgroup analyses. In the DL group, the pooled SEN and SPE were 89% (95% CI: 82%–93%) and 85% (95% CI: 76%–91%), respectively. In the ML group, the pooled SEN and SPE were 89% (95% CI: 86%–90%) and 88% (95% CI: 86%–90%), respectively (Supplementary Figures S3a, b, S12). For different imaging modality, nine US studies had a pooled SEN of 90% (95% CI: 88%–92%), a pooled SPE of 88% (95% CI: 83%–91%), and an AUC of 0.94 (95% CI: 0.92–0.96). Eleven CT studies had a pooled SEN of 89% (95% CI: 86%–92%), a pooled SPE of 90% (95% CI: 86%–93%), and an AUC of 0.96 (95% CI: 0.93–0.97). Four MRI studies had a pooled SEN of 88% (95% CI: 83%–92%), a pooled SPE of 86% (95% CI: 82%–89%), and an AUC of 0.89 (0.86–0.91). Four PET/CT studies had a pooled SEN of 85% (95% CI: 81%–87%), a pooled SPE of 86% (95% CI: 74%–93%), and an AUC of 0.87 (95% CI: 0.84–0.90) (Supplementary Figures S4a–d, S13). In eight studies using the same databases, AI had a pooled SEN of 88% (95% CI: 85%–91%), an SPE of 88% (95% CI: 85%–90%), and an AUC of 0.93 (95% CI: 0.91–0.95); however, clinician had a pooled SEN of 77% (95% CI: 66%–85%), an SPE of 80% (95% CI: 71%–87%), and an AUC of 0.85 (95% CI: 0.82–0.88) (Supplementary Figures S5a, b, S14). Twenty-one studies were conducted in Asia, and seven were conducted outside Asia. The pooled SEN and SPE were 88% (95% CI: 86%–90%) and 88% (95% CI: 85%–90%), respectively, with an AUC of 0.94 (95% CI: 0.92–0.96) in the Asia group. The pooled SEN and SPE were 90% (95% CI: 88%–91%) and 89% (95% CI: 85%–92%), respectively, with an AUC of 0.92 (95% CI: 0.90–0.94) (Supplementary Figures S6a, b, S15). Twenty-two studies had single-center data and six studies had multiple-center data. The pooled SEN and SPE were 88% (95% CI: 86%–90%) and 88% (95% CI: 85%–90%), respectively, with an AUC of 0.94 (95% CI: 0.92–0.96) in the single-center group. The pooled SEN and SPE were 91% (95% CI: 86%–94%) and 88% (95% CI: 84%–92%), respectively, with an AUC of 0.95 (95% CI: 0.92–0.96) (Supplementary Figures S7a, bS16). Fourteen studies had sample sizes ≤100 and 14 studies had sample sizes >100. The pooled SEN and SPE were 87% (95% CI: 85%–90%) and 88% (95% CI: 84%–91%), respectively, and the AUC was 0.94 (95% CI: 0.91–0.95) for ≤100. For sample size >100, the pooled SEN and SPE were 90% (95% CI: 87%–92%) and 88% (95% CI: 85%–90%), respectively, for >100 with an AUC of 0.95 (95% CI: 0.92–0.96) (Supplementary Figures S8a, bS17). Seven studies were published before 2020 and 21 studies were published after 2020. The pooled SEN was 90% (95% CI: 86%–93%) for studies published before 2020, and 90% (95% CI: 86%–93%) for those published after 2020. The SPE was 88% (95% CI: 86%–90%) and 88% (95% CI: 85%–90%), respectively. The AUC was 0.95 (95% CI: 0.93–0.97) and 0.94 (95% CI: 0.92–0.96), respectively (Supplementary Figures S9a, bS18). Nine studies were low and 19 studies were high and unclear. The pooled SEN was 89% (95% CI: 85%–91%) for the high and unclear group, and 89% (95% CI: 84%–93%) for the low group. The SPE was 88% (95% CI: 86%–90%) and 88% (95% CI: 85%–90%), respectively. The AUC was 0.95 (95% CI: 0.92–0.96) and 0.94 (95% CI: 0.92–0.94), respectively (Supplementary Figures S10a, bS19). Six studies used LASSO, six used SVM, and five used ANNs. The pooled SEN and SPE were 86% (95% CI: 83–89%) and 87% (95% CI: 84–90%), respectively, for the LASSO group (AUC 0.92, 95% CI: 0.90–0.94). For the SVM group, the pooled SEN and SPE were 95% (95% CI: 90–98%) and 87% (95% CI: 82–94%), respectively (AUC 0.97, 95% CI: 0.95–0.98). For the ANN group, the pooled SEN and SPE were 88% (95% CI: 84–91%) and 84% (95% CI: 73–91%), respectively (AUC 0.90, 95% CI: 0.87–0.93) (Supplementary Figures S11a–cS20). Sensitive analysis results are shown in Supplementary Table S5.

The high diagnostic accuracy of AI algorithms in this meta-analysis highlights their potential to complement or even surpass conventional imaging methods. The overlapping clinical and imaging features of PC and pancreatitis often lead to diagnostic uncertainty, which can result in delayed treatment for PC or unnecessary surgery for pancreatitis (50, 51). AI’s ability to extract and analyze high-throughput quantitative features from imaging data offers a promising solution to this problem (52). Notably, the pooled SEN and SPE of AI algorithms exceeded the performance thresholds typically observed in conventional imaging, suggesting that AI could serve as a valuable decision-support tool in clinical practice (53).

Subgroup analyses revealed that both ML and DL algorithms demonstrated comparable diagnostic efficacy, with ML achieving an SEN of 89% and an SPE of 88%, and DL achieving an SEN of 90% and an SPE of 88%. This suggests that the choice of algorithmic architecture may depend on specific clinical contexts and available computational resources. When stratified by imaging modality, AI models demonstrated generally high diagnostic performance across all four modalities. CT-based models achieved the highest pooled accuracy (AUC 0.95), indicating excellent discriminative ability. Ultrasound-based AI also showed strong performance (AUC 0.92), followed by MRI-based approaches (AUC 0.89). PET/CT-based AI yielded a somewhat lower accuracy (AUC 0.87) compared with the other modalities, although its performance remained clinically meaningful. These findings indicate that while AI can provide valuable diagnostic support across different imaging modalities, performance levels are not uniform. The inclusion of heterogeneous imaging modalities likely contributed to the overall inconsistency observed in the meta-analysis. Future research with larger, modality-specific datasets will be needed to determine the optimal imaging context for AI application in distinguishing PC from pancreatitis.

When analyzed separately, AI demonstrated generally high diagnostic performance across all four imaging modalities, although important differences were observed. CT-based AI models achieved the highest pooled accuracy with an AUC of 0.96, an SEN of 89% (95% CI: 86%–92%), and an SPE of 90% (95% CI: 86%–93%), and showed relatively higher heterogeneity (I² = 82.85% in SEN and I² = 73.43% in SPE). This strong performance may be attributed to the widespread availability of CT, larger training datasets, and more standardized acquisition protocols in the included studies.

Ultrasound-based AI also performed well, with an AUC of 0.94, an SEN of 90% (95% CI: 88%–92%), and an SPE of 88% (95% CI: 83%–91%), although heterogeneity was moderate (I² = 54.18% in SEN and I² = 69.71% in SPE). This variability likely reflects operator dependence and inconsistent image quality across centers. The relatively strong performance despite these limitations suggests that AI may help mitigate inter-operator variability by standardizing interpretation, though most studies were single-center with limited sample sizes.

MRI-based models reached a pooled AUC of 0.89, an SEN of 88% (95% CI: 83%–92%), and an SPE of 86% (95% CI: 82%–89%), but showed high heterogeneity (I² = 78.68% in SEN and I² = 15.83% in SPE). While MRI offers superior soft-tissue contrast, the diversity of acquisition sequences (T1WI, T2WI, and DWI) and lack of harmonization across institutions likely contributed to the variability. Standardized MRI protocols will be critical for improving reproducibility in future studies.

PET/CT-based AI demonstrated the lowest pooled accuracy with an AUC of 0.87, an SEN of 85% (95% CI: 81%–87%), and an SPE of 86% (95% CI: 74%–93%), accompanied by substantial heterogeneity (I² = 29.43% in SEN and I² = 91.07% in SPE). Small study numbers, differences in tracer selection, and heterogeneous reconstruction protocols may explain these results.

Collectively, these findings indicate that although AI shows diagnostic value across modalities, the magnitude of benefit is not uniform. The overall high heterogeneity (I² > 75%) in our meta-analysis can be explained, at least in part, by modality-specific differences. From a clinical perspective, CT- and ultrasound-based AI appear most promising for near-term application due to higher accuracy and accessibility, whereas MRI- and PET/CT-based approaches will require larger, multicenter datasets and standardized acquisition protocols to reduce heterogeneity and establish reliable clinical utility.

In our subgroup analysis by algorithm type, SVM-based models achieved the highest pooled diagnostic accuracy (AUC 0.97), followed by LASSO and ANN frameworks. Although these differences were not statistically significant, the trend suggests that kernel-based classifiers such as SVM may be particularly effective for radiomics-derived imaging features, whereas ANN performance may be limited by small sample sizes and non-standardized input data.

A notable finding from this meta-analysis is the consistently higher pooled diagnostic performance of AI compared with clinicians in studies that directly evaluated both approaches using the same patient datasets. Across eight studies, AI achieved a pooled sensitivity of 88% and a pooled specificity of 88% (AUC 0.93), whereas clinicians achieved a pooled sensitivity of 77% and a pooled specificity of 80% (AUC 0.85). These results suggest that AI systems may provide comparable or numerically higher diagnostic accuracy than clinicians in differentiating PC from pancreatitis. However, this difference should be interpreted cautiously, as the included studies were heterogeneous in imaging modality, algorithm type, and clinician experience, and lacked formal paired statistical testing. Rather than replacing clinicians, AI may help reduce inter-reader variability and enhance diagnostic consistency, especially in settings with limited subspecialty expertise. However, it is important to note that AI systems are not intended to replace clinicians but rather to augment their diagnostic capabilities, particularly in complex cases where imaging features are ambiguous. However, this advantage was not consistent across all modalities and study settings, and several reports indicated comparable performance between AI and radiologists. These findings suggest that AI may complement clinicians by providing reproducible, quantitative assessments, but it cannot yet be regarded as superior or a replacement. For clinical practice, AI could be particularly useful in cases where imaging features are subtle or inter-observer variability is high, offering an additional layer of decision support. Still, the variability in algorithm design, training datasets, and external validation highlights the need for caution. Future multicenter prospective trials are required to determine whether AI can achieve robust and generalizable improvements in real-world workflows, and to clarify the conditions under which AI provides incremental benefit over expert interpretation (54).

The sensitivity analysis demonstrated that the pooled diagnostic accuracy of AI was largely unaffected by the exclusion of studies with high risk of bias. The recalculated estimates for sensitivity, specificity, DOR, and AUROC were nearly identical to those of the primary analysis, reinforcing the stability and reliability of the overall findings as shown in Supplementary Table S5. Interestingly, although slight numerical differences were observed—particularly a marginal increase in DOR and AUROC after excluding high-bias studies—these changes were not substantial. This trend may indicate that higher-quality studies tend to yield more consistent and slightly stronger diagnostic performance, possibly due to standardized patient selection and imaging protocols.

Nevertheless, caution is warranted, as such differences could also arise from variations in sample size, imaging modality, or algorithm type rather than methodological quality alone. Overall, the consistency of the results across sensitive analyses supports the robustness of the pooled conclusions and underscores the importance of maintaining methodological rigor in future AI-based diagnostic research.

Despite the promising results, significant heterogeneity was observed among the included studies (I² > 75%). Subgroup analyses identified several factors contributing to this heterogeneity, including algorithmic architecture, imaging modality, geographic origin, and publication year. For instance, studies conducted in Asia demonstrated slightly higher diagnostic accuracy (AUC = 0.95) compared to those outside Asia (AUC = 0.92). Similarly, single-center studies showed marginally better performance than multi-center studies, possibly due to standardized imaging protocols and reduced variability in data collection (54). These findings suggest that modality-specific factors, such as image resolution, acquisition protocols, and dataset availability, may partly explain the inconsistency observed in the overall pooled results. Nevertheless, the number of studies per modality remains limited, which restricts the ability to draw firm conclusions. Future studies focusing on single-modality AI applications, ideally with larger and more standardized datasets, will be essential to clarify the role of imaging modality in diagnostic performance.

There are several limitations in our meta-analysis. Firstly, almost all of the included studies were retrospective, which introduced potential bias in patient selection. Secondly, we manually calculated the data necessary to construct the confusion matrix in some studies due to the absence of full test in performance metrics. Additionally, some studies were abandoned as we failed to extract or calculate numbers of TN, TP, FP, and FN. We should be cautious in interpreting pooled estimates from quantitative analyses. Thirdly, despite the implementation of subgroup analyses and sensitivity tests, the sources of heterogeneity still waited for further detection. Similar high heterogeneity was also observed in recent systematic reviews evaluating the quality of radiomics models in other fields.

In summary, AI algorithms demonstrate robust diagnostic performance in differentiating PC from pancreatitis with CT- and ultrasound-based approaches currently appearing most promising for clinical translation. Importantly, AI should be viewed as an adjunct to, rather than a replacement for, radiologists, providing quantitative, reproducible insights that can support clinical decision-making. Future research should focus on prospective multicenter studies with standardized imaging protocols, external validation, and explainable AI frameworks to facilitate integration into radiology workflows. Combining AI with other data modalities, such as genomic or histopathologic information, may further enhance diagnostic accuracy and pave the way for precision medicine in pancreatic diseases.