Pancreatic cancer is one of the deadliest cancers. Its prognosis is usually bad. Most patients get diagnosed at an advanced stage because of this.¹. Although relatively less common than other malignancies, pancreatic cancer accounts for a disproportionately high number of cancer-related deaths. In developed nations, lung cancer ranks seventh in cancer-driven deaths, according to the World Health Organization. According to predictions, the second leading cause of death from cancer will be lung cancer by 2030, mainly due to its frequency.^1–3 Gender differences are evident, as men exhibit a slightly higher incidence rate of 5.7 per 100,000 (34,530 cases) and a mortality rate of 4.5% (27,270 deaths), compared to women, who have an incidence rate of 4.9 per 100,000 (31,910 cases), and a mortality rate of 4.0%.^1,2 . This underscores the urgent need for improved diagnostic and treatment methodologies.⁴ Traditional imaging modalities, including CT scans, have been integral in diagnosing pancreatic cancer; however, they are often limited by their inability to detect small or early-stage tumors accurately.^5,6 Over the last few years, particularly with deep learning (DL), image analysis by means of artificial intelligence (AI) has shown very favorable performance as a routine radiological assessment aid. Systems powered by AI may be able to reliably and repeatedly perform important feature identification, fine texture assessment, volumetric quantification, and risk classification that is not immediately visible to the human eye. The new developments will assist in many important activities, such as early detection of tumors and localization of lesions. This review assesses the performance of these techniques, outlining their respective strengths and limitations. Distinct from earlier reviews, this paper systematically classifies AI models based on their functional approaches, such as segmentation, classification, or hybrid techniques, while also emphasizing the datasets employed, evaluation metrics applied, and their practical relevance to clinical settings. Furthermore, the study proposes future research avenues that will help to bridge the divergence between technological innovation and clinical applications.

The primary objective of this study is to assess the current state of AI models for pancreatic cancer diagnosis and to compare different AI frameworks and techniques. It also identifies key challenges and opens up new avenues for future research. It highlights a few essential gaps that need to be filled for the appropriate design and implementation of innovative AI-based solutions in the clinical setting.

The organization of the remaining paper is as follows. Specifically, the section “Pancreas Tumors in the Human Body” outlines pancreatic tumor in the human body. “Methods” section presents the selected studies and summarizes their characteristics. The “Results” section reports the results of the systematic analysis. Finally, there is a comparative analysis of AI models used in the section “Comparative Analysis of AI Techniques” . The section “Challenges” discusses the challenges and limitations associated with current diagnostic approaches. Sections “Future Directions and Scope” and “Conclusions” outline the future research directions and conclude the study.

Tumors are masses of abnormal pancreatic cells that proliferate uncontrollably and disrupt normal tissue function, leading to malignant transformation and pancreatic cancer. The most common type is ductal adenocarcinoma. In this type, the cells lining the ducts of the pancreas are involved.⁷ Patients with pancreatic cancer may experience symptoms such as abdominal pain, decrease in appetite, reduced energy levels, weight loss, and jaundice. Unfortunately, the disease is often detected after it has metastasized.^4,7

This paper surveys a total of 33 research studies from 2018 to 2025, focusing on PDAC diagnosis. In this section, we describe various AI-Driven Segmentation models for tumor localization, DL models for tumor classification, radiomics-based approaches using CT imaging, and early detection models. The performance of each category was evaluated using standard metrics such as DSC, Sensitivity, Specificity and AUC for clinical utility and feasibility of deployment.

The area of pancreatic cancer diagnosis using AI and advanced CT imaging has evolved rapidly in recent years. A multitude of research organizations have contributed significantly to this field. Tables 2 and 3 summarize studies published between 2018 and 2025, focusing on detection, segmentation, and classification of pancreatic cancer using DL and ML techniques.

Table 2 highlights that encoder–decoder frameworks based on U-Net architectures are used in the majority of segmentation studies. U-Net-like architectures preserve spatial hierarchies through skip connections, which are particularly beneficial for anatomical structures with irregular boundaries. Boers et al.³⁷ developed an interactive 3D U-Net with the capability of reducing the slice-wise inconsistencies and user re-annotating. Using volumetric convolutions and the Adam optimizer to study a private cohort of CT scans, DSC = 78%. The subsequent studies mainly focused on enhancing the representation of features through salience awareness and multi-scale refinement. Hu et al.²² proposed a framework based on DenseASPP to iteratively refine dissimilarity between the region of interest and the background. The DSC value on the NIH dataset ranges from 67.19% to 91.64%. Z. Chen et al.³⁶ and W. Li et al.²⁴ employed the methods of multi-scale feature fusion strategies. More recently, studies were conducted with emphasis on cross-dataset generalization and architectural sophistication. In this regard, Mahmoudi et al.³⁰ combined the outputs of Attention U-Net and Texture Attention U-Net (TAU-Net) for the pancreas and PDAC mass regions, respectively. They initially identified the performance gap related to organ and tumor segmentation. The reported mean DSC was approximately 0.64 for PDAC mass segmentation, highlighting the persistent performance gap between organ and tumor segmentation. The study by Amiri et al.²⁹ was proposed to extend the previously mentioned segmentation to pancreatic subregions and pancreatic ducts. Reinforcement learning-based anatomical navigation proposed by Amiri et al.²⁹, further extended the segmentation to pancreatic subregions and ducts, achieving a mean DSC of 0.70. Large-scale evaluations across Medical Segmentation Decathlon (MSD), The Cancer Imaging Archive (TCIA), and National Institute of Health (NIH) datasets by Mukherjee et al.²⁷ and Yang et al.³⁹ demonstrated that segmentation performance remains highly dataset-dependent, underscoring challenges related to domain shift and annotation variability.

Table 3 shows a considerably broader range of model architectures. They include classical ML/DL, transformers and hybrid pipelines. Several CNN-based classifiers were trained on publicly available CT-image datasets hosted on Kaggle. Nadeem et al.⁶ proposed a multi-stage pipeline for multi-class pancreatic lesion classification, which involved anisotropic diffusion filtering-based preprocessing, U-Net-based watershed segmentation, and AlexNet classification, and reported very high accuracy and AUC. There has been an increase in hybrid strategies involving transfer learning combined with traditional classifiers. For example, Alaca and Akmeşe⁵ utilized feature extractors DenseNet121 and InceptionV3, followed by nearest neighbors, support vector machines, and random forests. The authors found balanced accuracies of up to 92.5%. In the same way, Alaca⁷ used DARTS-optimized MobileViT models. Institutional and clinically curated datasets remain relatively limited in the current literature. However, such studies offer a more realistic estimation of the performance. Javed, Qureshi, Deng, et al.¹⁷ and Mitrea et al.⁴³ internally externally validated data source cohort study. Although the reported accuracy values were slightly lower in externally validated cohorts, these findings provide more clinically meaningful evidence of generalizability. Prior works in these domains, like Choi et al.⁴⁷ and Chu et al.⁴⁸, feature-engineering driven, included texture descriptors, CA19-9 biomarkers, volumetric and radiomics features in their studies. Therefore, conventional ML approaches may still yield competitive results in highly clinically constrained settings with limited data availability. The classification literature shows extensive variability across datasets and validation designs.

Overall, segmentation studies yield relatively stable DSC standardized datasets, such as NIH and MSD. Segmenting a tumor is harder than segmenting a whole organ. Classification studies indicate high accuracy values leveraging publicly available datasets, with high variability in institutional cohorts of patients. This demonstrates that dataset heterogeneity and external validation are, to a greater extent, an obstacle to clinical translation.

Despite notable advancements, there remains a significant challenge in detecting very small (<1 cm) or isodense tumors, which is crucial for improving patient outcomes. Although reinforcement learning-based anatomical maps utilize attention mechanisms and probability maps to segment pancreas regions, a notable gap remains in research specifically targeting fine-grained segmentation of the pancreatic duct, which is essential for diagnosing conditions such as PDAC.¹¹ The 3DGAUnet approach enhances volumetric feature representation and provides more detailed tumor segmentation, addressing the gap in accurate and effective segmentation of PDAC and its subregions. The large-scale pancreatic cancer detection model addresses the challenge of detecting very small or isodense tumors, a significant gap in early stage detection accuracy. The model showed good potential applicability to other non-contrast CT types of early pancreatic cancer.⁴⁸

The previously reviewed literature has indicated the challenges involved in pancreatic cancer detection and segmentation. Numerous studies report similar challenges, including limited data availability, heterogeneous imaging data, generalizability, clinical translation issues, etc. In order for state-of-the-art technologies to become obsolete, new methodologies, multimodal investigations, and more clinically relevant diagnostic models must advance the field. A promising direction involves advanced DL architectures, particularly 3D convolutional networks integrated with attention mechanisms. Recent studies show that attention-based frameworks positively impact the simultaneous pancreas and its subregions and tumors segmentation. To illustrate this concept, PanSegNet was proposed by Zhang et al.⁵⁰, for integrating linear self-attention modules into the encoder–decoder of nnU-Net, which gave DSCs over 88% for all pancreatic organs on multi-center CT images of 140 cases from the MSD challenge.^49,50. Future research should explore hierarchical and multi-level attention, transformer-based encoders, and lightweight attention modules for clinical deployment. There is also an emerging research trend around building hybrid and efficiency-oriented models. Li et al.²⁴ proposed attention-augmented adversarial U-NetsClick or tap here to enter text. and Amiri et al.²⁹ stated a reinforcement learning-based pancreas anatomical mapping that enables promising accuracyClick or tap here to enter text.. Moreover, they found that they can maintain a high accuracy with fewer parameters and less computation, which means the optimized model can run in a real environment with limitations. Going forward, hybrid model development and application on larger datasets, with patient data from multiple institutions, should be explored. Combining the different data types should be an important direction for the future that would allow a holistic disease characterization for early disease prediction and intervention. Once again, a large-scale ensemble detection model with CT, and clinical features by Chen et al.²⁰ achieved an AUC of 0.95 with good sensitivities.⁸ Besides, a radiomics-based early prediction framework predicted pancreatic cancer with an advance of up to 36 months before clinical diagnosis, given by Chen W. et al.²¹. Further research concentrating on effective multimodal fusion methods and related longitudinal modeling can pave the way for personalized risk stratification and early intervention in the future. After this, we need architectural designs for model generalization and interpretability. Above all, developing such strategies depends on the availability of data from large cohorts and variation among institutions. Hence, a critical milestone in a translation pathway is leveraging the federated learning (FL) paradigms. FL provides a setting to carry out multi-institutional model training without private data sharing. Importantly, future research must ensure rigorous clinical validation, explainability, and seamless workflow integration. Though there is a reported high accuracy for some experimental settings, most of these studies are based on retrospective evaluation that lacks prospective validation and an assessment of interpretability. Aligning future research directions explicitly with identified limitations—such as poor external validation, limited subregion focus, and absence of decision-support integration—may facilitate the translation of AI models into meaningful improvements in pancreatic cancer screening, diagnosis, and patient outcomes.

The recent emergence of AI-assisted detection and characterization targeting pancreatic cancer has become commonplace in modern times. The growing interest in clinical translation represents more than a methodological advancement. According to an analysis of the research carried out, a significant improvement has been achieved in ML and DL architectures. CNN, attention-based methodologies, and hybrid frameworks have demonstrated significant progress. Pancreas segmentation, tumor detection, and early risk prediction represent major advancements in the field. In addition, these approaches have demonstrated performance characteristics in controlled study settings. There also exist performance indicators for the long-standing issues of pancreatic cancer early detection. Early tumor detection remains technically challenging due to class imbalance, subtle imaging features, and anatomical heterogeneity. Moreover, anatomical variability and tumor subtype heterogeneity further complicate model development. These challenges are commonly observed in pancreatic cancer and other oncological conditions. Significant challenges remain for large-scale clinical translation. In multi-institutional studies, prospective challenges include model interpretability, clinical evaluation, etc. Additionally, other challenges include integration into clinical workflows, domain adaptation, data harmonization, and fixed model generalization. We also face a challenge concerning across-protocol generalization and changes in scanners. This includes the use of complex DL architectures, attention and transformer-based mechanisms, compensation for lack of data by means of federated learning and generative modelling, and setting common evaluation protocols across heterogeneous populations and imaging platforms. Most importantly, the use of AI should not be restricted to detection and segmentation. It must also evaluate treatment response, characterize the tumor and predict disease progression. In summary, translating AI technologies successfully could radically change how pancreatic cancer is managed, leading to better outcomes and curbing mortality in patients. Such advancements may contribute to earlier detection, personalized treatment planning, and improved patient survival outcomes.