• A Multimodal AI Framework for the Early Identification of Pancreatic PaNET Cancer: Combining Biomarkers, Genetic Predisposition, and Radiological Data
  • Melika Fiuj,1,* Narges Etemadi ,2
    1. Department of Medicine, Faculty of Medicine, Tehran Medical Sciences, Islamic Azad University, Tehran, Iran
    2. Department of Microbiology, Faculty of Life Science, East Tehran Branch, Islamic Azad University, Tehran, Iran


  • Introduction: Pancreatic neuroendocrine tumors (PanNETs) are rare but clinically significant neoplasms, representing approximately 1–2% of all pancreatic malignancies. Despite their generally indolent nature, delayed diagnosis of PanNETs often results in advanced disease presentation, contributing to increased morbidity and mortality. Recent epidemiological data suggest a rising incidence of PanNETs, largely attributable to improvements in imaging modalities and heightened clinical awareness. However, due to their heterogeneous biological behavior—ranging from non-functional, asymptomatic lesions to hormonally active and aggressive tumors—early detection remains a formidable clinical challenge. The conventional diagnostic approach relies on a multimodal strategy incorporating cross-sectional imaging, functional imaging (such as ⁶⁸Ga-DOTATATE PET/CT), and histopathological confirmation. While these methods are indispensable, each carries inherent limitations, particularly in sensitivity, specificity, and accessibility, which hinder early-stage identification. Moreover, although serologic biomarkers like chromogranin A and neuron-specific enolase offer adjunctive diagnostic value, their specificity is often compromised by false positives and variable expression patterns. In this context, artificial intelligence (AI) has emerged as a transformative tool in medical diagnostics, offering the ability to analyze high-dimensional datasets from imaging, genomics, and proteomics with unprecedented precision. AI-driven approaches, including convolutional neural networks (CNNs), support vector machines (SVMs), and ensemble learning methods, have shown promise in detecting PanNETs, classifying tumor grade, and predicting prognosis based on integrated data sources. These capabilities not only enhance diagnostic accuracy but also pave the way for personalized medicine through the identification of novel biomarkers and clinically relevant subtypes. This study introduces PaNET-MultiDx, a novel multimodal AI framework designed to integrate radiological, genomic, and serologic inputs for the early and accurate detection of PanNETs. By leveraging the complementary strengths of deep learning and machine learning models, this approach aims to overcome the current limitations of conventional diagnostics and establish a foundation for precision oncology in neuroendocrine tumor management.
  • Methods: 1. Early Detection of Pancreatic Neuroendocrine Tumors (pNETs) Early detection of pancreatic neuroendocrine tumors (pNETs) is critical, as delayed diagnosis significantly contributes to increased morbidity and mortality, despite their typically indolent behavior. Although pNETs constitute only about 1–2% of all pancreatic neoplasms, their reported incidence has increased in recent decades, primarily due to advancements in diagnostic imaging and heightened clinical awareness [1]. While many pNETs are non-functional and remain asymptomatic until advanced stages, a subset are functional, secreting bioactive hormones that can lead to complex clinical syndromes. This variability underscores the necessity for prompt biochemical assessment and radiological evaluation [2]. The 5-year survival rate for patients diagnosed with localized pNETs exceeds 90%, but falls to below 40% in those presenting with distant metastases, emphasizing the prognostic value of early detection [3]. Currently diagnostic work-up typically requires a multimodal strategy involving cross-sectional imaging (CT, MRI), functional imaging—such as Ga-68 DOTATATE PET/CT—and histopathological confirmation via biopsy, frequently guided by endoscopic ultrasound [4]. In conclusion, considering the increasing incidence, heterogeneous clinical behavior, and stage-dependent prognosis of pNETs, there is an urgent need for heightened clinical vigilance and implementation of comprehensive diagnostic protocols to enable early and accurate diagnosis. 2. Biomarker-Based Diagnostics for Pancreatic Neuroendocrine Tumors Biomarkers for pNETs are extracted from various biological sources including serum, tumor tissue, and emerging liquid biopsy platforms such as circulating tumor DNA (ctDNA) and exosomes, enabling non-invasive and dynamic assessment of tumor behavior. Among the most studied biomarkers, chromogranin A (CgA) is elevated in approximately 60–80% of pNET patients, though its specificity is limited due to false elevations in non-neoplastic conditions [5]. Neuron-specific enolase (NSE) is present in 30–40% of cases and is more associated with poorly differentiated tumors. Pancreatic polypeptide (PP) and vasoactive intestinal peptide (VIP) are elevated in 20–30% and <10% of patients respectively, typically in functional tumors [6]. Recent advances in transcriptomic and proteomic profiling have identified novel markers such as NETest, a multigene blood-based assay with reported diagnostic accuracy exceeding 90%, offering superior sensitivity compared to single-analyte markers [7]. In conclusion, while traditional serum markers like CgA remain in clinical use, the integration of high-resolution molecular biomarkers significantly improves diagnostic yield and holds promise for personalized disease monitoring in pNETs. 3. Inherited Genomic Alterations in Pancreatic Neuroendocrine Tumors: Mechanisms, Markers, and Outcomes Patients with hereditary pancreatic neuroendocrine tumors (PanNETs) exhibit a 20–30% higher long-term mortality rate compared to those with sporadic tumors, primarily due to earlier onset, multifocality, and increased recurrence risk [8]. Approximately 10–15% of PanNETs are associated with inherited genetic syndromes such as multiple endocrine neoplasia type 1 (MEN1), von Hippel-Lindau disease (VHL), neurofibromatosis type 1 (NF1), and tuberous sclerosis complex (TSC) [9]. These germline mutations frequently affect key tumor suppressor genes, playing a critical role in early tumorigenesis and multifocal neoplasia. The MEN1 gene, located on chromosome 11q13.1, is the most commonly mutated, with germline mutations identified in up to 44% of hereditary PanNET cases; it encodes the nuclear scaffold protein menin, which regulates transcription, DNA repair, and cell proliferation [10]. The VHL gene, situated on chromosome 3p25.3, is mutated in approximately 13% of hereditary cases and leads to dysregulated hypoxia-inducible signaling and excessive angiogenesis. NF1, positioned at chromosome 17q11.2, is involved in about 5% of inherited PanNETs, where loss of neurofibromin results in RAS/MAPK pathway hyperactivation. Additionally, TSC1 on chromosome 9q34.13 and TSC2 on chromosome 16p13.3 are mutated in around 1–2% of cases, both affecting mTOR signaling and contributing to abnormal cell growth [11]. While some hereditary PanNETs exhibit indolent behavior, particularly non-functioning tumors, those associated with MEN1 and NF1 syndromes demonstrate worse prognoses, reinforcing the need for vigilant surveillance and individualized treatment strategies in genetically predisposed populations"Table 1".
  • Results: Limitations of Conventional Diagnostic Methods for Pancreatic Neuroendocrine Tumors (PanNETs) and the Need for AI-Enhanced Solutions Conventional diagnostic methods for pancreatic neuroendocrine tumors (PanNETs)—including radiographic imaging, histopathological analysis, and genetic testing—face notable limitations in accuracy, sensitivity, and early-stage detectability. Cross-sectional imaging techniques such as computed tomography (CT) and magnetic resonance imaging (MRI) often fail to differentiate PanNETs from other pancreatic masses due to overlapping radiologic features, with reported diagnostic sensitivity ranging from 63% to 82% depending on tumor size and radiologist experience. Histopathological assessment, though definitive, is invasive and subject to sampling error, particularly in heterogeneous or cystic lesions. Furthermore, while genetic screening for mutations in MEN1, DAXX, and ATRX has enhanced our understanding of PanNET pathogenesis, the penetrance of these mutations varies, and their predictive value for tumor behavior remains suboptimal, with sensitivity rates as low as 50% in sporadic cases [24]. Each of these modalities relies on distinct data sources—morphological patterns in imaging, tissue architecture in histology, and nucleotide sequences in genomics—yet none singularly provides a reliable, comprehensive diagnostic framework. In summary, the inherent limitations in sensitivity, specificity, and accessibility of these conventional approaches underscore the urgent need for integrative, AI-enhanced models capable of synthesizing multimodal data to improve early and accurate detection of PanNETs.
  • Conclusion: In summary, the proposed PaNET-MultiDx framework leverages a synergistic integration of high-resolution imaging, comprehensive genomic profiling, and established serologic biomarkers to address the current diagnostic challenges in pancreatic neuroendocrine tumors. By combining a convolutional neural network trained on contrast-enhanced CT and MRI datasets with a multilayer perceptron for mutation data (MEN1, DAXX, ATRX) and a random forest classifier for serum analytes (CgA, NSE, serotonin), the model harnesses modality-specific strengths while mitigating individual limitations. The attention-based fusion mechanism not only captures complex interdependencies across data types but also enhances interpretability through Grad-CAM visualization, thereby facilitating clinician confidence and translational applicability. Operational feasibility is supported by the availability of large, annotated imaging repositories (e.g., TCGA, institutional archives) and open-source AI toolkits (MONAI, PyTorch, scikit-learn), ensuring that PaNET-MultiDx can be developed, validated, and iteratively refined within real-world clinical workflows. Preliminary evidence from CNN-based and ensemble-learning studies—demonstrating accuracies exceeding 90% in imaging classification and genomic biomarker prioritization—underscores the strong foundation upon which this multimodal approach is built. Furthermore, the non-invasive nature of liquid biopsy and serologic testing positions PaNET-MultiDx as a highly accessible screening adjunct, with potential to detect early-stage lesions prior to overt clinical manifestation. Ultimately, by transcending the siloed constraints of single-modality diagnostics, PaNET-MultiDx represents a paradigm shift toward precision oncology in pancreatic neuroendocrine tumors. Its capacity to integrate radiographic phenotypes, molecular genotypes, and biochemical signatures offers a comprehensive lens for early detection, risk stratification, and treatment planning. Future prospective studies and multicenter collaborations will be pivotal in validating clinical performance, assessing cost-effectiveness, and establishing standardized guidelines for implementation, with the overarching goal of improving patient outcomes through timely and tailored interventions.
  • Keywords: Pancreatic neuroendocrine tumors adiogenomics CNN PaNET-MultiDx