Cancer remains one of the world's most formidable public-health challenges. The Global Cancer Observatory reported almost 20 million new diagnoses and 9.7 million deaths in 2022 (Figure 1), translating to an economic burden that exceeds 1% of global gross domestic product each year (1). Recent market reports reflect the growing clinical interest and research activity in pancreatic cancer. Precedence Research (2024) forecasts strong expansion of the pancreatic cancer market through 2034 driven by rising investment in diagnostics (2), therapeutics and clinical trials. This commercial growth both supports and mirrors the increased funding for AI based detection and prognostic technologies that are the main focus of this review.

Despite incremental gains in screening and therapy, the absolute number of cancer-related fatalities continues to rise with population aging and growth (3). Pancreatic cancer (PC), of which approximately 90% are pancreatic ductal adenocarcinomas (PDAC), deserves special attention because its mortality far exceeds its incidence ranking. Globally it is the twelfth most commonly diagnosed malignancy, yet it already ranks seventh among cancer deaths (3, 4). Five-year relative survival in high-income countries has only recently crept into double digits—roughly 11%–13%—making PDAC the deadliest of the major solid tumors (5). Contributing factors include a stealthy symptom profile, an anatomically concealed primary site and a dense desmoplastic micro-environment that confers intrinsic resistance to cytotoxic therapies (3). Without intervention, PDAC is projected to become the second leading cause of cancer mortality in North America and parts of Europe by 2030 (4).

The global burden of pancreatic cancer has continued to rise over the past three decades. According to GLOBOCAN 2022, there were 510,992 new cases worldwide (1). Estimates from the Global Burden of Disease (GBD) 2019–2021 update indicate incident cases increased from about 489,862 in 2019 to 508,533 in 2021, while deaths rose from 486,869 to 505,752 over the same period (3). The age-standardized incidence rate (ASIR) decreased slightly from 6.04 to 5.96 per 100,000, whereas the age-standardized death rate (ASDR) declined from 6.03 to 5.95 per 100,000 (3). These aggregate figures mask striking geographic disparities. High-income regions such as North America, Western Europe and high-income Asia–Pacific have ASIRs approaching 10 per 100,000, whereas low-SDI countries have rates as low as 1.6 per 100,000 (3). Case fatality mirrors incidence because PDAC is frequently lethal once symptomatic. The incidence-to-mortality ratio was only ≈1.28 in 2024, and more than 85% of patients are diagnosed at an unresectable or metastatic stage (3).

Table 1 summarizes recent estimates from high-incidence regions. In the United States, the National Cancer Institute projects 67,440 new cases and 51,980 deaths for 2025. Although pancreatic cancer will account for only 3.3% of new malignancies, it will cause 8.4% of all cancer deaths (5). The United States has seen incidence climb by about 0.7% per year since 2001, and PDAC is forecast to become the second leading cause of cancer death by 2030 (4). European countries such as Hungary, the Czech Republic and Finland report age-standardized mortality rates exceeding 8 per 100,000 (4). In Asia, China already accounts for over 25% of worldwide PDAC deaths, and Japan has experienced one of the steepest rises in incidence among high-income countries (3). The median age at diagnosis is approximately 70 years, and more than 80% of patients present with unresectable or metastatic disease (3).

Socio-economic inequalities magnify these differences. A recent analysis using GBD 2021 data showed that countries with high SDI had ASIRs around 10 per 100,000 and ASDRs near 9.4 per 100,000, compared with rates around 1.6–1.7 per 100,000 in low-SDI countries (3). The number of disability-adjusted life-years (DALYs) attributable to pancreatic cancer rose from 10.9 million in 2019 to 11.3 million in 2021, highlighting the growing societal and economic toll (3). Projection models estimate that incident cases and deaths will both exceed 875,000 by 2044 (3).

Pancreatic carcinogenesis is multifactorial, with both modifiable and non-modifiable determinants. Modifiable lifestyle factors include tobacco smoking, obesity, diabetes, diet and alcohol consumption. Smoking remains the strongest environmental risk factor: pooled analyses show that current smokers have nearly a two-fold increase in risk compared with never smokers, and heavy smokers (>25 cigarettes day⁻¹) can experience a 2.7-fold elevation (6). Risk declines after cessation but may not return to baseline until 15–20 years later (6). A meta-analysis reported that current smokers have a 75% increased risk of pancreatic cancer relative to never smokers and that elevated risk persists for at least a decade after quitting (6). Second-hand smoke exposure appears to play a minor role (6).

Obesity and metabolic dysfunction are increasingly recognized contributors. A pooled analysis of cohort studies found that obesity approximately doubles the risk of pancreatic cancer in both men and women, and each 5 kg m⁻² increase in body-mass index raises risk by about 12% (6). Adipose tissue produces pro-inflammatory cytokines and increases insulin resistance, creating a carcinogenic milieu. Type 2 diabetes mellitus (T2DM) is both a risk factor and a consequence of pancreatic cancer: long-standing T2DM increases PDAC risk by 1.5–2.0-fold, whereas new-onset diabetes confers a 5–8-fold increased risk within one to three years (6). Indeed, a large proportion of patients with pancreatic cancer have diabetes or impaired glucose tolerance at diagnosis (6). Other dietary and lifestyle factors—such as heavy alcohol consumption, diets rich in red and processed meats, low intake of fruits and vegetables, and chronic pancreatitis—have also been implicated, although effect sizes are generally smaller.

Non-modifiable factors include age, sex, ethnicity and genetic predisposition. Incidence increases dramatically with age and peaks between 70 and 74 years (3). Males tend to have higher age-standardized incidence and mortality rates than females across all age groups (3). Familial pancreatic cancer accounts for 5–10% of cases; germline mutations in BRCA1/2, CDKN2A, PALB2, STK11/LKB1, TP53 and mismatch-repair genes confer markedly elevated lifetime risks. Hereditary pancreatitis (PRSS1 mutations), Peutz–Jeghers syndrome and familial atypical multiple mole melanoma syndrome are notable syndromes requiring surveillance. Pancreatic cancer risk also varies by ethnicity; for example, African-American populations in the United States experience incidence and mortality rates about 30% higher than those of Caucasian populations, likely reflecting a combination of genetic, metabolic and socio-economic factors (3).

Early detection remains the cornerstone for improving pancreatic-cancer outcomes. However, current screening modalities lack sensitivity for precursor lesions, and the overall rarity of PDAC precludes population-wide screening. Standard imaging techniques such as abdominal ultrasound and computed tomography have limited ability to detect small pancreatic tumors. Pre-diagnostic CT(Computed Tomography) scans often fail to reveal abnormalities in more than half of patients, and subtle signs may precede the clinical diagnosis by 3–36 months (7). Consequently, only about 13.6% of PDAC cases are diagnosed while still localized, and over 85% present with locally advanced or metastatic disease (3). When patients are diagnosed at an early stage and can undergo complete resection followed by multi-modal therapy, median overall survival can exceed 60 months (7).

Given the low prevalence of PDAC in the general population, surveillance strategies are currently recommended only for high-risk individuals—those with strong family histories or pathogenic germline mutations. Ongoing prospective cohorts (for example, the CAPS and Dutch familial pancreatic cancer studies) monitor high-risk participants with annual magnetic resonance imaging (MRI) and endoscopic ultrasound (EUS). A recent update involving approximately 1,700 participants with familial or genetic risk factors reported that surveillance detected tumors at an earlier stage: 38.5% of screen-detected cancers were stage I compared with 10.3% in the general population; 5-year survival reached 50% among the surveillance cohort vs. 9% for non-screened patients (8). The study underscores the potential of targeted screening to extend survival but also highlights logistical challenges: surveillance requires specialized centers with multidisciplinary expertise, and false-positive results can lead to unnecessary interventions (8).

The advent of artificial intelligence and machine-learning techniques offers hope for earlier detection and more accurate risk stratification (9). Radiomics extracts high-dimensional quantitative features from imaging data, capturing subtle textural and morphological patterns that are imperceptible to the human eye. Deep-learning architectures—particularly convolutional neural networks (CNNs) and U-Net–based segmentation models—have demonstrated promising performance in identifying pancreatic lesions and classifying intraductal papillary mucinous neoplasms (IPMNs). A recent systematic review of AI-based IPMN imaging reported classification accuracies ranging from 60% to 99.6%, although heterogeneity in study populations, imaging protocols and analytic pipelines limits direct comparison (7, 10). Most studies relied on CT data despite guidelines favoring MRI, and many used small, single-center cohorts without external validation, leading to high risk of bias (7). Standardized frameworks, large multi-institutional datasets and rigorous external validation are urgently needed.

AI-augmented imaging may also facilitate detection of subtle pre-diagnostic changes. For example, radiomics and deep-learning models can segment the pancreas automatically and identify textural or shape alterations months to years before clinical presentation. In a recent review of AI-augmented imaging, radiomic signatures from routine CT were able to identify early changes that preceded diagnosis by 3–36 months (7). Such models, once validated and integrated into clinical workflows, could trigger further evaluation or enrolment into high-risk surveillance programmes. Nevertheless, adoption of AI in clinical practice raises issues of data privacy, algorithmic fairness, interpretability and regulatory oversight (7, 10).

Table 2 summarizes seven recent AI-focused reviews on pancreatic cancer. While these reviews document progress in CNNs, transformers, and radiomics, they consistently reveal gaps in comprehensive attention mechanism surveys, unified dataset visualizations and external validation strategies—limitations that our paper addresses.

While recent reviews organize AI studies by imaging modality [Podina et al.(11), Yao et al.(12)], clinical task (detection, segmentation, classification), data type [EHR-only: Mishra et al.(13)], or clinical subtype [IPMN: qadir et al.(7); PDAC vs. pancreatitis: Zhang et al.(14)]—approaches that effectively catalog method–modality pairings for specific protocols—our generational taxonomy provides a distinct meta-level view by organizing studies into three methodological waves: conventional machine-learning/radiomics pipelines (Generation 1, 2015–2020), deep-learning CNNs (Generation 2, 2017–2023), and attention/transformer-based multi-modal fusion (Generation 3, 2020–present). This temporal framework enables us to quantify performance evolution [AUC 0.84–0.98 in Generation 1 → 0.92–0.99 in Generation 2 → 0.996 in Generation 3 (15); segmentation Dice 0.19–0.70 → 0.57–0.87], synthesize findings across all modalities (CT/MRI, histopathology, genomics, biomarkers) within a single coherent framework, reveal validation trends (external validation 50% → 57% → 67% across generations; multi-center validation 0% → 29% → 22%), and identify the research frontier (attention-augmented, domain-adaptive architectures) by extrapolating from the architectural trajectory—advantages not accessible through modality-specific or task-specific organization. As detailed in Table 2, this generational perspective complements existing reviews by revealing temporal, cross-modality, and performance-evolution patterns that become visible only when viewing AI methods as successive generations building on architectural innovations.

Figure 2 illustrates global search interest in “Artificial Intelligence” from 2015 to 2025, highlighting sharp public attention peaks that coincide with major AI advances (16). Such public and media surges often track research investment and adoption cycles that accelerate translation of AI techniques into healthcare applications.

This review makes several unique contributions to the growing literature on AI-enabled pancreatic cancer detection and prognosis. First, it conducts a rigorous and reproducible search across PubMed and Google Scholar covering 2015–2025 to capture more than sixty peer-reviewed studies that apply machine learning, deep learning or attention-based methods to imaging, histopathology and molecular data. The search strategy reduces the risk of missed early-stage or conference works by including theses and preprints in addition to indexed journals.

This set of contributions positions our work as a comprehensive, methodologically rigorous and forward-looking synthesis that highlights both the promise of attention-based, multi-modal AI and the steps required for safe and equitable clinical translation.

This review is organized to walk the reader from clinical motivation to actionable research priorities (Figure 3): Section 1 provides the clinical and epidemiological background (incidence, mortality, key risk factors) and motivates the need for improved surveillance and AI-augmented detection, with focused subparts on risk factors, surveillance, and the role of emerging AI methods; Section 2 details the literature search and selection strategy (databases, search queries, deduplication, title/abstract triage, inclusion/exclusion criteria, full-text review and synthesis protocol) so readers can reproduce the corpus assembly; Section 3 synthesizes the surveyed AI work by methodological generation—Section 3.1 conventional machine-learning and radiomics pipelines, Section 3.2 deep-learning models for detection/segmentation/classification, and Section 3.3 attention- and transformer-based architectures and multi-modal fusion—highlighting representative studies and performance patterns; Section 4 catalogs the data sources that underpin the field (Section 4.1 two-dimensional imaging, Section 4.2 three-dimensional CT/MRI volumes, Section 4.3 radiomics, Section 4.4 clinical/registry data, Section 4.5 genomic/molecular assays, and Section 4.6 biofluid/biomarker panels) and links data regimes to suitable model classes; Section 5 discusses cross-cutting issues, methodological gaps, and interpretation (e.g., patient-level splitting, external validation, and domain shift); Section 6 outlines emerging directions and concrete recommendations (prospective trials, reporting standards, domain generalization, fairness, and federated learning); and Section 7 concludes with priorities for clinical translation.

We performed a structured, reproducible literature search and selection procedure to identify primary studies that applied machine-learning (ML), deep-learning (DL), or attention-based methods to pancreatic cancer imaging, pathology or molecular data. The goal was to capture methods that directly address detection, segmentation, classification or prognostication in pancreatic disease using computational approaches. The overall workflow is illustrated in Figure 4.

This systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement. The protocol was prospectively registered with [Open Science Framework (OSF), OID: https://doi.org/10.17605/OSF.IO/2DVHJ ].

This section synthesizes more than sixty peer-reviewed studies on artificial-intelligence methods for pancreatic-cancer detection, prognosis and treatment monitoring published between 2015 and 2025. We group the papers into three methodological generations—conventional machine-learning (ML) pipelines, deep-learning (DL) models, and attention- or transformer-enhanced frameworks—while emphasizing representative contributions and situating related work in context (Figure 5).

Figure 6 shows the global trend of AI publications by country and region between 2000 and 2025 (17). China's rapid ascent in research output, combined with continued contributions from the United States and the EU, helps explain why deep-learning and transformer approaches have become dominant in biomedical imaging and diagnostics.

A panoramic view of the data landscape clarifies why pancreatic-cancer AI has shifted from classical machine-learning to deep-learning and, most recently, to attention or transformer architectures. Instead of enumerating studies, the material below pools the datasets actually used by the short-listed papers (Table 8) by the primary data modality that governs network design and evaluation. Three trends emerge. First, modern imaging pipelines lean on ever larger, mostly private three-dimensional CT archives. Second, prognosis-oriented radiomics and -omics projects still depend on mid-sized, often single-center cohorts. Third, a handful of public benchmarks act as “connective tissue” for cross-paper comparison and pre-training.

The past decade has witnessed a rapid transition from conventional machine-learning pipelines to end-to-end deep networks and, most recently, to transformer-enhanced architectures for pancreatic cancer. Early studies relied on hand-crafted radiomic and clinical variables coupled with support-vector machines, random forests or shallow neural networks. These models provided useful proof-of-concepts but were limited by the size and heterogeneity of available datasets and required expert feature engineering (22, 26). Deep-learning approaches alleviated some of these constraints by learning hierarchical features directly from images. Modified VGG and ResNet models achieved internal accuracies around 98%–99% for binary detection when trained on balanced CT or histology datasets (46, 48). Hybrid stacks that combined denoising, segmentation and deep belief networks further pushed sensitivity to nearly 100% on curated cohorts (48). More recent work incorporates multi-dimensional attention and transformer components to better capture long-range context and focus the network on the small, heterogeneous pancreas and tumor region. For instance, the Multi-Dimensional Attention Gate network (MDAG-Net) introduces spatial, channel and multi-input attention gates; compared with a U-Net baseline on the Task07 Pancreas dataset it improved Dice, precision, recall and mean IoU by 5.3, 1.5, 12.7, and 7.6 percentage points, respectively. This shift illustrates how attention helps filter redundant background features and recalibrates convolution kernels to emphasize tumor regions.

Progress has been driven not only by algorithmic innovation but also by access to larger and more diverse datasets. Early efforts were restricted to two-dimensional slices or small, single-center cohorts; for example, the PCCD dataset comprised 1,800 portal-venous slices, and several urine-biomarker studies contained fewer than 600 samples. Contemporary studies leverage full three-dimensional volumes and multi-center registries. Large-scale cohorts such as PANDA aggregated more than 3,000 non-contrast CTs with external validation across nine hospitals, enabling robust training and external assessment (15). Public benchmarks like the NIH, MSD Task07, and Synapse datasets now provide standardized segmentation tasks, fostering fair comparison and transfer learning. Meanwhile, new modalities—fundus photography, plasma-omics, and multi-phase CT fused with clinical biomarkers—have diversified the data landscape (18, 79). This variety has underscored the importance of modality-specific architectures: two-dimensional patch CNNs remain effective for dense histology or fundus imagery, whereas three-dimensional U-Nets and transformers excel on volumetric CT, and attention-based fusion layers integrate imaging with lab tests or clinical variables.

The model-summary (Tables 5–7) and the data-source (Table 8) together highlight a consistent pattern: (i) radiomics and conventional ML are strongest for prognosis and longitudinal monitoring, where repeated measures (delta-radiomics on serial CT or pre/post-RT PET/CT) capture treatment-induced change and heterogeneity (26–28); (ii) CNNs dominate lesion detection/segmentation on volumetric contrast CT—especially when anatomical context (pancreas/duct masks, secondary signs) is injected—yielding high AUROC with clinically interpretable by-products (masks, heat-maps) (56, 57); and (iii) transformer/attention components are most compelling when fusing modalities (dual-phase CT with laboratory values, or imaging with clinical covariates) and when global context matters (e.g., multi-center non-contrast CT) (15, 75, 77). In short, the “what each method is good at” in Tables 5–7 maps directly to the data regimes cataloged in Table 8: hand-engineered features plus survival models for serial/prognostic settings; CNNs for voxel-level localization; and attention/transformers for multi-modal fusion and long-range dependencies.

Recent transformer-enhanced models demonstrate state-of-the-art performance on both detection and segmentation tasks. Hybrid DenseNet models incorporating channel and spatial attention and eleven clinical variables achieved 92.44% accuracy, AUC 0.971 and F1-score 0.936 for differentiating serous vs. mucinous cystic neoplasms on MRI. The dual-attention TransUNet (DA-TransUNet) integrates positional and channel attention blocks into a U-shaped transformer; on the multi-organ Synapse dataset it improved mean Dice by 2.32 percentage points and reduced the Hausdorff distance by 8.21 mm relative to TransUNet, with a pancreas-specific Dice gain of 5.73 percentage points. Such gains come at minimal computational cost—only a 2.54% increase in parameters. Similarly, the PancreasNet progressive residual transformer network combines Swin blocks, enhanced feature reweighting and regulated fusion; it reported 92.4% accuracy, 93.1% recall, 90.7% specificity, and Dice 0.87 on a 290-volume cohort, outperforming earlier convolutional CAD systems by 5–7 percentage points. Multi-modal attention further enhances classification: a dual-phase CT plus clinical biomarker model achieved 80% accuracy, 0.82 precision, 0.86 specificity, 0.74 recall, and AUC 0.83 for pre-operative lymph-node metastasis prediction, representing improvements of 0.10–0.16 points over radiomics baselines. At the extreme, the DeepOptimalNet cascade achieved a reported 99.3% accuracy, 99.1% sensitivity, and 99.5% specificity for pancreatic tumor classification in CT imaging (83), though such perfect scores on limited datasets warrant cautious interpretation and external validation. Collectively, these results indicate that attention-augmented architectures provide tangible benefits in focusing on subtle pancreatic features, balancing precision and recall, and reducing false positives.

To provide a comprehensive quantitative overview, Table 9 aggregates reported performance metrics across the representative studies in Tables 5–7. Across conventional machine-learning approaches, AUC values ranged from 0.84 to 0.98, with accuracy spanning 81%–99.97%; deep-learning models improved these ranges to AUC 0.92–0.99 and accuracy 82.5%–99.8%, while also introducing segmentation capabilities (Dice 0.19–0.70). Attention and transformer-based architectures achieved the highest overall discrimination [AUC up to 0.996 for PANDA (15)] and best segmentation performance [Dice 0.87 for PancreasNet (74)], with F1 scores consistently exceeding 0.92 across all generations and peaking at 0.97 for the kernel attention network (65). This progression demonstrates measurable performance gains from conventional ML through deep learning to attention/transformer architectures, particularly for complex imaging tasks requiring spatial reasoning and multimodal data fusion. However, it is important to note that these metrics reflect heterogeneous datasets, tasks (detection vs. segmentation vs. prognosis) and validation strategies; therefore, direct cross-study comparisons must account for differences in sample size, prevalence, imaging protocols and patient-level vs. slice-level evaluation.

A few entries in Tables 5–7 (and related reports) claim near-perfect performance—e.g., 99.8% for a DCNN+DBN pipeline on the 1,800-slice PCCD set (48) and 99.6% for an NASNet → Elman hybrid on a 500-image two-class CT dataset (47). Based on the dataset properties summarized in Table 8 and the authors' own descriptions, several factors likely contribute.

First, both studies used relatively small, balanced research sets (hundreds to low thousands of images, not patients), often from one or few scanners/centers. Such homogeneity reduces nuisance variation and can inflate apparent performance. Second, PCCD and similar corpora are 2D slice collections where many slices from the same patient share textures, field-of-view and reconstruction kernels. If slices (or overlapping patches) from a given patient appear in both train and test folds, leakage occurs, and models learn patient/scanner signatures rather than disease biology, driving metrics toward 99%–100%. Third, several pipelines framed a binary “tumor vs. healthy” task with tumor-centric crops or post-hoc ROI selection, simplifying the decision surface and limiting challenging negatives (e.g., pancreatitis, cysts, post-operative change). Finally, neither study reports robust, multi-center external testing at the patient level; metrics were typically computed on balanced sets and may not reflect screening prevalence or case-mix.

These considerations help explain why 99%+ results are rare in larger, multi-institutional cohorts or prospective settings. In line with the trends in Table 8, when studies scale to diverse volumes and enforce patient-level splits with external validation, reported AUROC/accuracy typically settle into the 0.83–0.96 range for CAD, and Dice for pancreas/tumor segmentation into the 0.57–0.88 range—credible but not “perfect” (56, 69, 82).

To close the gap between Tables 5–8 and clinical reality, we recommend that future work:

Where feasible, anatomically informed attention (ducts, vessels) and multi-modal fusion should be prioritized for tasks that intrinsically require global context (e.g., lymph-node metastasis or subtype attribution) (15, 77).

Table 3 presents a comprehensive methodological quality assessment across 24 studies from three model generations (Machine Learning, Deep Learning, and Attention/Transformer architectures), systematically analyzing single-center vs. multi-center designs and internal vs. external validation strategies—key quality indicators that directly impact model generalizability and clinical translation.

The evolution of AI models has increasingly emphasized holistic patient context. Sex-specific RNA-Seq survival predictors illustrate how biological sex influences transcriptomic signatures and survival modeling (18). Multi-omics panels combining proteins and gene expression have outperformed single-omics classifiers, highlighting the synergy of heterogeneous data sources (20). The lymph-node metastasis study cited above fused dual-phase CT and eleven laboratory variables via an attention-based fusion module; this integration raised AUC from 0.72 (radiomics alone) to 0.83. Similar improvements are observed when fusing imaging with risk factors: logistic regression on urine biomarkers plus CA19-9 improved sensitivity and specificity to 96%, and multi-head fusion of dual-phase CT with clinical biomarkers increased lymph-node detection accuracy by over 10 percentage points. Beyond imaging, end-to-end frameworks now incorporate self-supervision and pseudo-lesion pretext tasks to reduce annotation burdens; models pre-trained on synthetic lesions achieved external accuracies up to 82.5% despite using only 10% of labeled data (60).

Although remarkable progress has been made, several challenges remain. Many studies are retrospective and single-center; prospective, multi-institutional trials are needed to verify generalisability and clinical impact. Domain shift across scanners, patient populations and acquisition protocols can degrade performance; attention-based causal regularization and uncertainty estimation may mitigate such shifts. Interpretation and fairness are critical: anatomically guided attention and uncertainty-aware mechanisms provide more transparent saliency maps and allow models to defer to clinicians when confidence is low. In addition, counterfactual explanations should be incorporated to further enhance interpretability and strengthen clinician trust. Early detection remains an open frontier. Delta-radiomics and longitudinal EHR models have shown that subtle texture changes or clinical patterns can predict pancreatic cancer 24 months before diagnosis (22). Recent work suggests that radiomic signatures in apparently normal pancreas can precede diagnosis by over a year. Translating such findings into practice will require precise volumetric segmentation and integration with high-risk cohort selection. Combining imaging with genomics, proteomics, and metabolomics may uncover prodromal signatures, while non-invasive modalities such as fundus photography already hint at systemic biomarkers. Finally, data privacy, algorithmic bias and the integration of AI outputs into clinical workflows require careful consideration. The next decade will likely see the convergence of self-supervised learning, causal graph modeling, multi-modal fusion and federated training to deliver interpretable and equitable AI tools for the early detection, staging, and management of pancreatic cancer. To ensure clinical translation, future studies must also adopt prospective, population-based trial designs that directly evaluate real-world screening and triage performance.

Among the surveyed studies, PANDA (9-center, 3,208 training CTs, 5,337 external validation cases, AUC 0.984–0.987) (15) and Chen et al.'s nationwide study (1,473 real-world CTs, AUC 0.95, 89.7% AI sensitivity vs. 74.7% radiologist for < 2 cm tumors) (82) represent the closest candidates for clinical deployment, distinguished by multi-center validation and head-to-head clinician comparisons. Beyond imaging, Nasief's delta-radiomics (treatment response prediction, external AUC 0.98) (26) and biomarker panels [PancRISK urine test, AUC 0.94 (32); plasma amino-acids, AUC 0.86 (33)] show strong translational potential through interpretability and integration with existing workflows. However, regulatory approval faces critical barriers: 83.3% of studies lack multi-center validation required by FDA/EMA for generalizability evidence (Table 3); only four studies documented prospective reader comparisons mandated for Software as a Medical Device (SaMD) submissions; and domain shift effects [11.8% performance drop in cross-ethnic validation (60)] highlight inadequate robustness testing. The EU AI Act (2024) further mandates algorithmic transparency, bias audits across patient subgroups, and uncertainty quantification—requirements met by < 5% of reviewed studies.

This review has traced the methodological progression from classical machine-learning and radiomics to deep representation learning and, most recently, attention- and transformer-enhanced architectures for pancreatic cancer detection, segmentation and prognosis. Attention-augmented models and multi-modal fusion have delivered consistent gains in diagnostic accuracy, segmentation Dice and AUC across a range of tasks and datasets, demonstrating clear potential to augment human readers and to enable earlier, non-invasive detection.

Despite these advances, several barriers remain before routine clinical deployment is possible. Many published studies are retrospective, single-center or underpowered; evaluation metrics and reporting are heterogeneous; and model robustness to domain shift, scanner variability and population differences is often insufficiently tested. Moreover, challenges around interpretability, fairness, data governance and workflow integration persist and must be addressed alongside pure algorithmic improvements.

To accelerate safe and equitable translation we recommend the following priorities:

In summary, the combination of anatomically guided attention, principled multi-modal fusion, and rigorous external validation offers the most promising path toward clinically useful AI for pancreatic cancer. If future work couples these technical advances with prospective trials, clear reporting standards and careful attention to fairness and deployment, AI could materially improve early detection, staging and personalized management for patients with pancreatic cancer.