AI-based diagnosis of pancreatic cancer using computed tomography (CT) imaging has emerged as a highly active area of research [6]. Korfiatis et al. [7] developed an automated AI model trained on CT images from more than 3,000 patients and demonstrated that this model could detect pancreatic cancer a median of 475 days (range, 93–1,082 days) prior to clinical diagnosis. The proposed model utilizes an end-to-end convolutional neural network (CNN) architecture, which automatically extracts features directly from raw CT images. On diagnostic CT scans, the model achieved an accuracy of 92% and was able to effectively identify subtle lesions that are often missed by radiologists [7].

Cao et al. [8] introduced the Pancreatic Cancer Detection with AI model, a deep learning approach for detecting pancreatic lesions using non-contrast CT scans. The model was trained on data from 3,208 patients at a single center and achieved an area under the curve (AUC) of 0.986–0.996 in external validation across 10 centers with 6,239 patients [8]. Additionally, in a real-world clinical setting involving 20,530 consecutive patients, the model demonstrated a sensitivity of 92.9% and a specificity of 99.9% [8].

Chen et al. [9] developed a deep learning model using contrast-enhanced CT images from 546 patients with pancreatic cancer and 733 healthy controls. In the internal test set, the model achieved a sensitivity of 89.9% and a specificity of 95.9%, which was not significantly different from the sensitivity of radiologists (96.1%; p = 0.11) [9]. In a nationwide, real-world validation cohort, the model demonstrated a sensitivity of 89.7% and a specificity of 92.8%. Particularly, for small pancreatic cancers less than 2 cm in size, the model achieved a sensitivity of 74.7% [9].

Radiomics enables the automated extraction and analysis of quantitative features from medical images, such as CT and magnetic resonance imaging (MRI), which are often imperceptible to the human eye. These features include shape, texture, and intensity, and can be extracted using manual or automated methods. Based on these features, traditional machine learning algorithms can be applied to objectively analyze pathological information embedded within the images. Yao et al. [10] demonstrated that radiomics-based machine learning models are capable of detecting early signals of pancreatic cancer on CT scans obtained well before clinical diagnosis [11]. In a study by Mukherjee et al. [11], a model was developed to detect PDAC on CT scans acquired a median of 398 days (range, 93–1,092 days) prior to diagnosis. This model extracted 88 primary and grayscale radiomic features, selected 34 features, and evaluated four machine learning classifiers: k-nearest neighbors, support vector machine (SVM), random forest, and XGBoost [11]. In the test set, the model achieved high performance, with a sensitivity of 95.5%, specificity of 90.3%, and accuracy of 92.2% [11].

Comprehensive serum microRNA (miRNA) sequencing studies using automated machine learning have shown meaningful results [2]. In the study by Kawai et al. [2], serum samples from 212 patients with pancreatic cancer and 213 healthy controls, collected from 14 hospitals, were analyzed. A diagnostic model combining 100 highly expressed miRNAs with cancer antigen 19-9 demonstrated excellent performance, achieving an AUC of 0.99, sensitivity of 90%, and specificity of 98% [2]. Particularly, this model also exhibited high diagnostic accuracy in an asymptomatic early-stage (stage 0–I) pancreatic cancer cohort, with an AUC of 0.97, sensitivity of 67%, and specificity of 98% [2]. In a large multicenter study, Wang et al. [4] analyzed 798 miRNAs from five prospective cohort studies to identify circulating miRNAs associated with pancreatic cancer risk. This study utilized plasma samples collected within five years prior to cancer diagnosis and identified 120 risk-associated miRNAs in the discovery phase, of which three (hsa-miR-199a-3p/hsa-miR-199b-3p, hsa-miR-767-5p, and hsa-miR-191-5p) were validated in the replication phase [4].

Personalized and tumor-informed circulating tumor DNA (ctDNA) assays have emerged as practical tools for identifying molecular residual disease (MRD) in pancreatic cancer [12,13]. In a study by Botta et al. [14], 1,329 plasma samples from 298 patients were analyzed, revealing that patients with ctDNA positivity during the MRD window had a significantly shorter median disease-free survival (DFS) (6.37 months) compared to ctDNA-negative patients (33.31 months; HR, 5.45; p < 0.0001) [13]. In a clinical laboratory validation study of digital droplet PCR-based KRAS ctDNA analysis, Groot et al. [12] examined 290 plasma samples from 59 patients with PDAC. Preoperative detection of ctDNA was significantly associated with shorter DFS and overall survival, and the persistence of ctDNA after surgery was correlated with a higher recurrence rate [12].

In a metabolomics study combining mass spectrometry and machine learning, a collective biomarker integrating databases of primary metabolites and phospholipids achieved an accuracy of 97.4% in classifying patients with PDAC [15]. Iwano et al. [15] developed a machine learning-based diagnostic algorithm using a SVM, which also demonstrated utility in evaluating the effects of neoadjuvant chemotherapy.

Selvan et al. [16] proposed an enhanced dynamic CNN architecture for CT-based imaging. This model introduces a DynamicConv2D layer, which generates the final output by computing a weighted sum of multiple parallel convolutional kernels, with the weights dynamically determined by the input. This approach enables effective feature extraction that adapts to variations in radiodensity and anatomical structure [16]. Additionally, various data augmentation techniques–such as rotation, scaling, and hue transformation–were applied to ensure robust learning performance under potential variations in image quality encountered in real-world clinical settings [16]. Through these architectural improvements, the proposed model demonstrated a marked performance gain in the detection of PDAC compared to conventional models [16].

Ensemble methods are machine learning techniques that combine the predictions of multiple individual models (base learners) to achieve better performance than any single model alone. This approach is analogous to aggregating the opinions of several experts to reach a final decision; even if individual models make different errors, appropriately combining their predictions can offset these errors and result in more accurate overall predictions. Hegde et al. [17] introduced the XGSVM ensemble model, which integrates XGBoost and SVM algorithms, achieving an accuracy of 95.8% and outperforming conventional single models. This model utilized a diverse dataset incorporating imaging data, genetic markers, clinical history, and demographic information, enabling the learning of distinct patterns from each data type [17].

Conventional CNNs process images in a grid-like fashion, treating all pixels equally. However, in pathological tissues, the spatial relationships and interactions between cells and structures are critical for diagnosis, and CNNs are limited in capturing these complex geometric properties and spatial dependencies [18,19]. Graph convolutional networks (GCNs) are deep learning models designed to process data structured as graphs rather than regular images. In this approach, a whole-slide image is divided into small patches, each represented as a node in a graph, while the relationships between patches are defined as edges, thereby converting the entire image into a graph structure [20,21]. Each node aggregates information from its neighboring nodes to update its own features, and through multiple layers, the model progressively learns both local and global characteristics. Wu et al. [21] developed a GCN-based model to distinguish aggressive adenocarcinoma and less aggressive pancreatic tumors from non-neoplastic cases, achieving an F1 score of 0.85, which indicates a well-balanced precision and recall.

Liu et al. [22] developed a CNN model to distinguish pancreatic cancer tissue from non-cancerous pancreatic tissue on CT images, using a retrospective cohort and cross-racial external validation. In patient-based analyses, the model achieved a sensitivity of 97.3%, specificity of 100%, and accuracy of 98.6% in local test set 1, and a sensitivity of 99.0%, specificity of 98.9%, and accuracy of 98.9% in test set 2 [22]. In an external validation cohort using data from a U.S. institution, the model demonstrated a sensitivity of 79.0%, specificity of 97.6%, and accuracy of 83.2%. Notably, for small tumors measuring 1.1–1.2 cm, the model missed only 1.7% of cases, which was substantially lower than the 7.0% missed by radiologists [22]. The deep learning-based platform developed in this study, PANCREASaver, automatically analyzes CT scans and highlights suspicious lesions. By integrating multiple AI techniques and clinical data processing modules, it serves as an integrated AI platform for clinical practice, supporting tasks ranging from diagnostic assistance to outcome reporting [22].

Ferber et al. [23] introduced a workflow that integrates a GPT-4-based autonomous clinical AI agent with a multimodal precision oncology platform to support personalized clinical decision-making. The system comprises three main components. First, a Vision Transformer is employed to sensitively detect microsatellite instability as well as KRAS and BRAF mutations from histopathology slides. Second, Medical Segment Anything Model is used to semi-automatically segment regions of interest across various imaging modalities, including CT, MRI, and ultrasound. Third, the platform integrates OncoKB, PubMed, and Google search interfaces to provide real-time access to the latest literature, genomic variant information, and clinical guidelines. In validation involving 20 real-world multimodal patient cases, the agent autonomously selected the optimal analytic module for each case in 87.5% of instances and achieved an overall clinical decision accuracy of 91.0%. These results suggest that the system can effectively assist clinicians in diagnostic and therapeutic decision-making by seamlessly integrating complex data streams [23].

Federated learning is a distributed training technique that enables the development of AI models by exchanging and aggregating model parameters (such as weights and biases), rather than transferring sensitive raw data to a central server [24]. This approach allows AI models to leverage large-scale and diverse datasets to improve accuracy and generalizability, while simultaneously ensuring data privacy, security, and regulatory compliance [24]. Jia et al. [13] developed a machine learning model using federated learning networks to analyze electronic health record data collected from across the United States [6]. Utilizing data from over five million patients, the researchers developed and validated two predictive models. Among these, the PRISM (pancreatic cancer risk prediction model) neural network identified approximately 35% of pancreatic cancer cases early using the same criteria, compared to only about 10% identified by the conventional clinical risk criteria (standardised incidence ratio-based criteria)–demonstrating more than a threefold improvement in early detection of high-risk patients [6,13]. Wang et al. [25] evaluated the effectiveness of a federated learning-based AI model for pancreas segmentation using abdominal CT images collected from two different hospitals. One institution provided 420 portal venous phase abdominal CT scans acquired for preoperative gastric surgery planning, while the other contributed 486 contrast-enhanced abdominal CT scans. Without sharing any raw data externally, the federated learning approach enabled model training across both institutions, resulting in superior pancreatic tumor prediction compared to models trained on single-institution data alone. Furthermore, the study demonstrated that applying additional server-side model optimization techniques could further enhance the performance of federated learning models [25].

Patel et al. [24] conducted a multi-institutional collaborative study utilizing federated learning. This research aims to develop an AI model for the diagnosis of PDAC using multimodal data, including CT scans and laboratory test results [24]. The core privacy-preserving mechanism of federated learning is that sensitive patient data never leave the local servers of participating medical institutions, thereby enabling inter-institutional collaboration without compromising data privacy [26]. The AI model is trained locally at each institution, and only the model parameters (such as weights) are transmitted to a central server for aggregation, while the raw patient data remain securely stored on local servers. Through this approach, major medical institutions across the United States can participate in the development of large-scale, multi-institutional AI models while fully preserving patient privacy.

The performance of AI models for the early detection of pancreatic cancer is highly dependent on the quality and diversity of training data. However, the lack of standardization in medical records, limited availability of curated datasets, and stringent legal and ethical requirements for patient privacy protection continue to pose major barriers [26]. Another major challenge is generalizability; AI models developed at a single institution often exhibit variable performance when applied to other institutions or diverse populations, due to factors such as ethnicity, geographic location, and differences in imaging equipment [27]. Therefore, multi-institutional validation studies with large cohorts are essential to ensure robust and generalizable performance [9]. In addition, the establishment of clear regulatory frameworks for Food and Drug Administration approval and clinical implementation of AI-based medical devices is crucial. For example, the European Union AI Act classifies medical AI systems as high-risk and imposes strict requirements throughout their entire lifecycle, necessitating compliance from providers of medical AI.