Role of artificial intelligence and novel visualization techniques in the early diagnosis of pancreatic cancer: a review

Full Text

INTRODUCTION

Pancreatic cancer is a gastrointestinal cancer characterized by rapid progression, invasiveness, asymptomatic course at early stages, and relapses after surgical treatment [1]. Pancreatic ductal adenocarcinoma (PDAC) is the most common type of pancreatic cancer. In 80%–85% of cases, it is characterized by a progressive course or distant metastases. Furthermore, the five-year survival rate in pancreatic cancer is 12%, which is the lowest of all cancers, highlighting the significance of early diagnosis and treatment [2].

Despite progress in understanding PDAC, the disease is still associated with a poor prognosis due to late diagnosis and limited treatment options in advanced stages. Difficulties in early detection, along with the growing incidence of PDAC due to risk factors such as obesity and diabetes mellitus in patients over 50 years, increase the prevalence of this condition [3, 4]. In 2023, a total of 62,210 new cases of PDAC were reported in the United States, with the mortality rate reaching 49,380 cases by 2022 [3]. Of the 57,600 cases reported in 2020, 55% of patients already had metastases at diagnosis [4]. In the United States, PDAC accounts for 2% and 5% of all cancers and cancer deaths, respectively, highlighting the need for earlier detection of this condition [5]. In 2019, pancreatic cancer accounted for 3.4% and 3.0% of all cancers in Russia in males and females, respectively. Over the last decade, the incidence of pancreatic cancer has increased by 11.4% in males and by 25% in females. In 2019, 19,930 cases of newly diagnosed pancreatic cancer were reported [6].

The United States Preventive Services Task Force (USPSTF) does not recommend screening for PDAC in adult asymptomatic patients, given the low prevalence of this condition and the lack of evidence that such screening can improve detection, survival, or mortality rates [7]. However, some guidelines recommend imaging examinations and case follow-up in high-risk patients [8]. Furthermore, a randomized controlled study is currently underway to assess the role of screening in patients with newly diagnosed hyperglycemia and diabetes mellitus for early detection of PDAC using computed tomography (CT) (NCT04662879) [9]. However, given the lack of specific biomarkers and limitations of CT and magnetic resonance imaging (MRI) in detecting PDAC measuring <2 cm, early diagnosis remains challenging even in high-risk patients [10].

In recent years, artificial intelligence (AI) has been used for image processing in the diagnosis of PDAC, with encouraging results [11]. AI refers to systems that replicate human intelligence and can learn to make decisions. The rapid advancement of AI technology, notably machine learning (ML) and deep learning (DL), has piqued clinicians' interest in developing novel integrated, reliable, and efficient diagnostic tools for medical care. DL is a significant advancement in addressing challenges associated with large-scale data collection, processing, and differentiation. For many years, the medical community has been unable to solve the existing issues. However, DL has proven highly effective in identifying complex patterns within high-dimensional data and can be applied across a wide range of scientific fields. Furthermore, DL systems are trainable and capable of functioning based on raw data, such as numbers, text, or even their combinations [12]. Trained AI-based models can process input medical images and provide outputs within seconds, significantly speeding up diagnosis. According to research, the efficacy of AI systems in detecting PDAC is comparable, if not superior to that of health experts [11, 13]. Significant improvements in data analysis speed further contribute to more effective diagnosis and treatment, reducing the workload in health care [11, 13].

Recent advancements in AI-based imaging and image analysis have improved the sensitivity and specificity of early PDAC diagnosis. These advancements include cancer-specific positron emission tomography (PET) radiotracers, ultrasound contrast agents, and AI-based image processing and analysis algorithms [13].

SEARCH METHODOLOGY

The search was conducted in PubMed, Google Scholar, and eLIBRARY.RU. The following keywords in Russian were used: ранняя диагностика рака поджелудочной железы (early diagnosis of pancreatic cancer); искусственный интеллект (artificial intelligence); протоковая аденокарцинома поджелудочной железы (pancreatic ductal adenocarcinoma); медицинская визуализация (medical imaging); наночастицы (nanoparticles). Furthermore, the following keywords in English were used: pancreatic cancer; artificial intelligence; early diagnosis pancreatic ductal adenocarcinoma; pancreatic cancer imaging. The search depth was from 2003 to 2025. The review primarily included articles published in the recent 5 years; however, earlier publications of high scientific merit were also considered. Where necessary, an additional search of other relevant publications addressing the clinical and prognostic significance of novel imaging methods in early diagnosis of pancreatic cancer was conducted.

Titles and abstracts were independently screened, followed by full-text review of relevant articles. The following publication selection algorithm was used:

• Duplicates were removed prior to the search;
• During the search, the titles and abstracts of the selected publications were screened for relevance and full-text access. Abstracts, articles, and works without full-text access were excluded;
• Full-text articles were assessed for eligibility.

Inclusion criteria were as follows: full-text reviews published in peer-reviewed journals, meta-analyses, systematic reviews, randomized controlled or experimental in vitro and in vivo studies, and case reports in Russian or English containing the keywords specified above. As a result, 60 publications were included in the present review.

MODERN IMAGING METHODS IN PANCREATIC DUCTAL ADENOCARCINOMA

Photon-counting CT, the most recent advancement in CT, improves the capabilities and diagnostic potential of medical imaging. In contrast to conventional CT, photon detectors enable photon counting and assessment of their interactions [14]. The key advantages of this technique are a greater contrast-to-noise ratio, improved spatial resolution, and increased visibility of lesions at a lower energy (50 keV), resulting from improved response of the contrast agent. This provides significant benefits for the visualization and differentiation of the pancreatic parenchyma both under normal conditions and in PDAC, which is typically relatively isodense. Previous research found that low energy (e.g., 40 keV) improves the visibility of PDAC measuring ≤3 cm [15]. Furthermore, up to 44% of tumors (especially those measuring ≤2 cm) that are not detected by conventional CT are isodense relative to the pancreatic parenchyma [16, 17].

Another promising area is the development of non-iodinated molecular contrast agents for CT. Nanoparticles can be used in various imaging techniques by incorporating materials or moieties that correspond to the physical principles underlying the specific technique. This allows using nanoparticles in CT, MRI, and ultrasound (US) examinations. Gold and iron oxide nanoparticles are the most studied. The former have optical properties used in photothermal therapy for selective cancer cell elimination by radiation exposure, whereas the latter are used in MRI and targeted therapy for both diagnostic and therapeutic purposes [18]. Photon-counting CT visualizes the K-edge¹ in the range of 25–150 keV, identifying nanoparticles (e.g., gold or gadolinium nanoparticles) even at extremely low concentrations, facilitating the early detection of primary neoplasms and metastases. Furthermore, the use of nanoparticles is based on their interactions with specific molecular targets in PDAC, improving the efficacy of theranostics² [19]. Nanoparticles are effective in the early detection of PDAC; however, their use is currently limited to identifying specific markers in biomaterials, and they cannot be applied as disease-specific imaging agents [20, 21].

PET molecular imaging remains an area of considerable interest in the diagnosis of PDAC. ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) is the most commonly used radiopharmaceutical in real-world practice. However, the National Comprehensive Cancer Network (NCCN) does not recommend it for the diagnosis of PDAC [22]. Similar to other cancers, detecting small lesions using positron emission tomography/computed tomography (PET/CT) with ¹⁸F-FDG is challenging, given the relatively low signal-to-noise ratio due to poor visualization of small tumors against the high activity of the pancreatic parenchyma, especially if inflammatory changes are present [13]. Therefore, the search for new drugs and methods to increase the efficacy of early diagnosis in PDAC is relevant.

Histopathologically, PDAC is a hypovascular tumor composed of small tubules (ducts) in a dense fibrovascular stroma, with an infiltrative growth pattern [13, 23]. Many cancers, including PDAC, have elevated levels of cancer-associated fibroblasts that produce fibroblast activation protein (FAP) [23]. Several radiopharmaceuticals targeting FAP have been developed, notably the ⁶⁸Ga-labeled fibroblast activation protein inhibitor (⁶⁸Ga-FAPI), which inhibits this protein [24]. Several studies have supported the use of PET/CT with ⁶⁸Ga-FAPI in PDAC and its advantages over ¹⁸F-FDG [25–27].

However, like ¹⁸F-FDG, ⁶⁸Ga-FAPI has potential limitations, namely difficulties distinguishing between inflammation and tumor tissue [28]. However, studies on PET/CT with ⁶⁸Ga-FAPI for the diagnosis of PDAC in mouse models have demonstrated its advantages in early tumor detection compared to ¹⁸F-FDG [29]. Preliminary clinical data also indicate that this technique can be used in the early diagnosis of cancer and malignant transformation of mucinous pancreatic lesions [30, 31]. Further research is needed to assess the efficacy and cost-effectiveness of PET/CT with ⁶⁸Ga-FAPI. However, this radiopharmaceutical may become a viable option for the early detection of pancreatic cancer. Furthermore, radiopharmaceuticals that target integrins, epidermal growth factor receptor, and fibronectin extra domain B are being studied. Integrins are cell surface molecules that serve as fibronectin receptors, mediating cell interactions and promoting proliferation and angiogenesis. Many cancers are associated with elevated integrin levels [32, 33]. The epidermal growth factor receptor is a transmembrane glycoprotein essential for neoangiogenesis and tumor cell proliferation [34]. The fibronectin extra domain B is a protein that is found in the extracellular matrix in many cancers, including PDAC [35]. However, these radiopharmaceuticals require further research to assess their efficacy in the early diagnosis of PDAC.

Molecular ultrasound imaging is a promising technique for early detection of PDAC. Some researchers are developing microbubble contrast agents that target specific vascular endothelial markers in PDAC [36]. In a mouse model of PDAC, contrast-enhanced ultrasound with a microbubble contrast agent targeting the vascular endothelial growth factor receptor 2 detected small lesions measuring <3 mm [37]. Thymocyte differentiation antigen 1 is a molecular marker that has elevated levels in PDAC. In in vivo experiments, contrast-enhanced ultrasound with a thymocyte differentiation antigen 1-specific single-chain antibody detected PDAC in both orthotopic and transgenic mouse models [38].

Emerging evidence suggests that metabolic transformation plays a crucial role in the development of PDAC. Mutations in the KRAS oncogene, which are found in 90% of PDAC cases, promote tumor glycolysis. This is associated with increased activity of numerous glycolytic enzymes, including lactate dehydrogenase A, enhanced aerobic glycolysis, and lactate buildup, which promotes tumor growth [39]. PDAC is associated with decreased expression of genes encoding alanine aminotransferase, which regulates the conversion of pyruvate to alanine [40]. Noninvasive detection of these metabolic changes is possible, for example, using hyperpolarized ¹³C MRI, a novel molecular imaging technique with a high sensitivity and chemical specificity for metabolic processes that were previously difficult to diagnose [41–43].

Hyperpolarization enabled by dynamic nuclear polarization increases the sensitivity of visualization of ¹³C-labeled biomolecules, which are non-toxic to the body, by 10,000 times [41]. Hyperpolarized ¹³C-enriched pyruvate MRI allows detecting PDAC and assessing its progression in transgenic mouse models [40, 42]. For example, as pancreatic intraepithelial neoplasia progresses to PDAC, the ¹³C-alanine/¹³C-lactate signal ratio decreases gradually in the pancreas. A recent clinical study showed that hyperpolarized ¹³C-enriched pyruvate MRI can be used for a quantitative assessment of metabolic activity in healthy pancreas and in PDAC before and after systemic therapy [43]. These studies demonstrate that this technique can improve the diagnosis and treatment of PDAC in high-risk patients. Notably, hyperpolarized ¹³C-enriched pyruvate MRI has successfully passed clinical trials; it was found to be safer than PET/CT with ⁶⁸Ga-FAPI46, which was used as the initial molecular imaging technique [43].

RISK STRATIFICATION USING ARTIFICIAL INTELLIGENCE WITHOUT CONSIDERING IMAGING FINDINGS

AI refers to a wide range of computational technologies that allow computers to solve problems that would normally require human intelligence. ML is a sub-field of AI, and DL is a specialized sub-approach within ML that uses convolutional neural networks. DL is widely used in medical imaging because it can automatically review existing data, analyze errors, and make more accurate predictions, in contrast to ML, which frequently requires manual feature extraction. ML models require manual conversion of medical images into numerical values (supervised by a human expert), whereas DL processes original images directly. Key modeling elements in DL include data collection and preprocessing, model selection, architecting, training, testing, and efficacy assessment [12].

Placido et al. [44, 45] used DL to analyze medical data of 6 million patients, including 24 ,000 patients with PDAC, from the Danish National Patient Registry, and 3 million patients, including 3900 patients with PDAC, from the United States Veterans Affairs (US-VA) database. The authors assessed whether the progression of pancreatic cancer during 36 months after the initial diagnosis can be predicted using data from electronic medical records, such as diagnosis codes and medical history. The most effective model had an area under the curve (AUC) of 0.879 (0.877–0.880); however, for the US-VA database, the AUC decreased to 0.710 (0.708–0.712). According to the authors, these differences in efficacy are likely associated with varying approaches to clinical recording and diagnosis coding, as well as different time periods of medical resource utilization. However, some symptoms and diagnosis codes reported 0–6 months before the PDAC diagnosis, including unspecified jaundice, bile duct and pancreatic diseases, abdominal and pelvic pain, weight loss, diabetes mellitus, and gastrointestinal neoplasms, were comparable in different populations. These promising findings demonstrate how clinicians can use electronic medical records to assess the risk of various disorders. These data can aid in detecting minor trends over time or risk factors that would be difficult to identify on examination, particularly if the patient does not have an attending physician. Future AI models, including laboratory and physical examination findings from medical records, such as glucose levels and body weight changes, may improve the early diagnosis of PDAC, enabling timely identification of patients at risk and targeted influence on modifiable risk factors [44, 45].

The main challenge in developing reliable AI models and systems is the heterogeneous quality of electronic medical records, as well as limited access to large-scale personal data and associated safety concerns. The majority of medical records are unstructured, with erroneous or excessive information, preventing an effective, accurate analysis necessary for efficient AI models [12, 13, 45].

Imaging is essential for diagnosis, staging, and treatment planning in PDAC. However, both CT and MRI have limited diagnostic accuracy in detecting tumors measuring <2 cm (69% and 82%, respectively). Therefore, radiomics and DL may improve early detection of PDAC [46]. For example, the U-Net segmentation model and its variations are used to detect PDAC by identifying the pancreas and distinguishing it from other structures on relevant CT slices [47]. The classification model detects the presence or absence of pancreatic cancer. Input data for these classifiers are focused regions of the pancreas (regions of interest) identified on CT images using a segmentation model. The output is a score that indicates the probability of cancer. The efficacy of this technique is determined by two key factors: the quality of pancreas segmentation and classifier performance. However, the former is critical for accurate detection of abnormalities. Recent advancements in DL, including medical image segmentation models such as the Medical Segment Anything Model (MedSAM), have enabled accurate pancreas segmentation on CT images [48]. Classifier performance is determined by the volume and diversity of training data [49]. A classifier must consider various characteristics of the tumor, including PDAC (e.g., size, shape, and stage), as well as patient characteristics (age, sex, and race). Research on ML and DL using imaging findings has confirmed their efficacy in early detection of PDAC [50–53].

Mukherjee et al. [50] found that radiomics-based ML models can identify and quantify patterns typical for early PDAC on segmented images. For example, a support vector machine-based ML model could detect visually occult PDAC signs on pre-diagnostic portal venous phase CT scans taken for other indications 3 to 36 months before the final clinical diagnosis of PDAC was made [51]. In the test subgroup, the median time from pre-diagnostic CT to the diagnosis of PDAC was 386 (97–1092) days. Notably, during the preliminary clinical interpretation, the CT findings were considered negative for PDAC. Furthermore, the proposed support vector machine model showed high specificity for both an independent internal dataset and the public dataset National Institutes of Health-Pancreas CT (NIH-PCT) [52]. The model detected PDAC based on CT findings significantly more effectively than radiologists [50]. Moreover, the authors assessed the robustness of a radiomics-based support vector machine model for detecting visually occult PDAC signs on pre-diagnostic CT scans by simulating common variations in image acquisition and radiomics workflow using image perturbation methods. The model was robust within a broad range of variations, indicating its reliability [53].

Furthermore, a fully automated convolutional neural network-based AI model for early detection of PDAC was developed [54]. The authors used the previously proposed automated approach with AI-based volumetric segmentation [55, 56]. The model used one of the largest available datasets (~3000 CT scans). Notably, the authors excluded CT scans with biliary stents, because such devices are a source of bias, and AI identifies them as tumors [57]. An automated three-dimensional convolutional neural network accurately detected PDAC based on CT findings, regardless of isodensity. Despite being exclusively trained on CT scans with larger tumors, the model could detect PDAC on pre-diagnostic CT scans at a median 475 days (15–16 months) prior to diagnosis [54]. Moreover, the model was effective in high-risk populations, including patients with diabetes mellitus, potentially increasing the proportion of resectable PDACs by three times compared to currently reported values [58]. The model’s performance was consistent across variations in patient demographics, scanner specifications, and image acquisition protocols [54].

Cao et al. [59] proposed pancreatic cancer detection with artificial intelligence (PANDA model), an approach that can detect pancreatic lesions with high accuracy via CT. This model detects, segments, and classifies lesions by their subtype. The model was trained on a dataset of 3208 patients from a single center; moreover, a multicenter validation was performed involving 6239 patients across 10 sites. PANDA effectively detected pancreatic lesions, with an AUC of 0.986–0.996. The study included 33 radiologists with various levels of experience, from residents to pancreatic imaging specialists. They interpreted non-contrast CT scans for 291 patients. PANDA surpassed the average radiologist performance by 34.1% in sensitivity and 6.3% in specificity for PDAC detection. Furthermore, 15 pancreatic imaging specialists interpreted contrast-enhanced CT scans of the same patients. PANDA outperformed readers by 13.0% in sensitivity and 0.5% in specificity, even when only non-contrast CT scans were used. PANDA maintained high AUC values during external validation with datasets from China, Taiwan, and the Czech Republic. Moreover, PANDA demonstrated high sensitivity (92.2%) in a subanalysis of small PDAC lesions (<2 cm). The model was additionally tested on a dataset of 20,530 patients, which is particularly susceptible to reduced performance due to numerous subtle changes. In this case, PANDA also demonstrated consistently high sensitivity and specificity in the early detection of PDAC (>96% and 99.9%, respectively). According to the authors, PANDA could potentially serve as a new tool for large-scale PDAC screening using large databases, including non-contrast CT scans that are routinely taken for various indications [59].

Furthermore, DL can be used to detect other pancreatic neoplasms. Park et al. [60] proposed a three-dimensional DL model capable of detecting seven different types of solid and cystic pancreatic lesions, including PDAC, neuroendocrine neoplasms, and intraductal papillary mucinous neoplasms. The model was trained using CT findings of patients after pancreatectomy and patients without pancreatic abnormalities. The model's performance was compared with that of two experienced radiologists. In test set 1, the DL model achieved an AUC of 0.91, which is consistent with radiologists' performance (AUC 0.92–0.95). In test set 2, the model performed worse than radiologists (AUC 0.87 vs. 0.95–0.96, p < 0.001). However, this approach is a significant step forward in the automated detection of pancreatic tumors measuring up to 1 cm.

Chen et al. [11] proposed a DL model to detect PDAC using portal venous phase contrast-enhanced CT scans, with sensitivity, specificity, and AUC of 89.9%, 95.9%, and 0.96, respectively. There were no significant differences between the DL model sensitivity and radiologist conclusions (90.2% and 96.1%, respectively; p = 0.11). The model remained effective during external validation using 1473 CT scans (669 patients with PDAC and 804 controls) from institutions throughout Taiwan, with sensitivity, specificity, and AUC of 89.7%, 92.8%, and 0.95, respectively. The model demonstrated an adequate sensitivity of 74.7% for tumors smaller than 2 cm, which may go undetected in clinical practice.

ROLE AND POTENTIAL OF ARTIFICIAL INTELLIGENCE IN THE EARLY DIAGNOSIS OF PANCREATIC DUCTAL ADENOCARCINOMA

The studies discussed in this review demonstrate significant progress in the early detection of PDAC using AI technologies. However, introducing them into clinical practice requires a comprehensive prospective, multicenter validation. The assessment must take into account different patient populations and health facilities to ensure the reliability of AI-based approaches. The integration of AI into pancreatic imaging presents numerous challenges. The primary challenge is the lack of large open-source datasets (>10,000 patients), which are essential for effective training and testing of AI models. The majority of existing systems are not open source, resulting in a lack of transparency of training processes, datasets, and model parameters. This affects the reproducibility of results, forcing many institutions to develop in-house models using small datasets, which frequently fail to show comparable efficacy. Large open-source datasets will enable comparing existing models to objectively assess their efficacy and update initial settings. Another issue is the varying quality of available open-source datasets for pancreatic imaging, which limits their applicability for developing reliable, effective AI models. Moreover, there are limitations related to reimbursement, AI integration into routine practice, and comprehensive training of end users, including radiologists and other specialists, which prevent the widespread use of AI technologies in pancreatic imaging.

Some DL models have integrated distribution maps to identify regions of interest important for decision-making on CT scans. However, they merely provide a subjective interpretation, with no formal rationale for outputs. One promising area is interpretable AI systems that classify relevant visual signs using human-understandable explanations. Another option is AI models that can do quantitative assessments and estimate diagnostic uncertainties. Bayesian neural networks, Monte Carlo methods, and deep ensembles assess the uncertainty of AI outputs during forecasting. Accurate quantitative assessments may improve the interpretation of AI forecasts, with an emphasis on ambiguous cases where the model is uncertain due to atypical signs or low-quality images. The majority of modern approaches rely solely on CT findings and do not consider other relevant clinical factors, such as sex, age, medical history, and body mass index. Another promising area is integrating AI models with new biomarkers to improve the accuracy and efficacy of early detection of PDAC. Further research should focus on developing DL models capable of assessing these variables and detecting PDAC more accurately.

RISKS AND LIMITATIONS OF ARTIFICIAL INTELLIGENCE IN MEDICAL IMAGING

AI opens up new prospects in the diagnosis of various disorders, including pancreatic cancer, facilitating earlier detection of abnormalities and improved visualization. However, despite significant advancements in DL and computer vision, integrating AI into clinical practice presents numerous challenges. The main risks and limitations are associated with diagnostic accuracy, training data quality, interpretability of models, and ethical and legal aspects of technology use [11].

Diagnostic Accuracy and Reliability

One of the biggest issues of applying AI in medical imaging is the risk of false-positive and false-negative results [7]. AI models may erroneously classify benign neoplasms as malignant or fail to detect tumors at early stages, delaying treatment. False-positive results can lead to overdiagnosis and unnecessary invasive procedures, increasing the burden on patients and healthcare systems [44].

Another significant limitation is that the efficacy of AI systems depends on the quality of data. Algorithms are trained on medical images; however, differences in equipment specifications, image acquisition parameters, and visualization standards may impair diagnostic accuracy [50].

Limitations in Training and Application of Artificial Intelligence Models

AI models must be trained on large datasets to function properly; however, clinicians have limited access to such data [59]. The majority of training datasets are poorly balanced, with most images obtained at late rather than early stages of PDAC [12, 59]. As a result, models are less accurate in identifying tumors that require early detection.

Furthermore, medical images used to train models vary in quality and format depending on the type and settings of equipment [54]. This impedes the integration of trained models into real-world practice, because the model's accuracy may vary between images from different clinics.

Another major concern is the black box problem. Most DL models make decisions without explaining how they reached a certain conclusion. This limits their use in clinical practice, because physicians cannot verify the reasoning behind the AI's conclusion. To use AI models in clinical practice, their decisions must be interpretable and human-understandable [12].

Ethical and Legal Issues

The question of legal liability for errors made by AI models remains unresolved [50]. If an incorrect diagnosis is made, the physician, model developers, or health facility may be held liable, limiting the technology's wider adoption.

Another major concern is personal data protection. The use of AI requires processing large volumes of medical data, including CT and MRI scans, electronic medical records, and genetic data. Medical data processing can lead to data leakage, necessitating strict compliance with data protection laws such as the General Data Protection Regulation (GDPR) in Europe and the Health Insurance Portability and Accountability Act (HIPAA) in the United States [33].

Furthermore, the risk of bias in AI systems must be considered. If a model is trained on data with gender, racial, or social biases, this may affect diagnostic accuracy in specific patient populations. For example, a model largely trained on data of Caucasian patients may be less effective in detecting PDAC in Asian or African patients [13, 44].

Limitations of Artificial Intelligence in Clinical Practice

Despite the great promise of AI systems, physicians may be reluctant to incorporate them into routine practice. Many clinicians are concerned about the accuracy of AI models and their ability to replace human expertise. To successful implement AI systems, healthcare professionals must be trained, and the actual clinical benefit of AI must be confirmed [12, 13].

Another limitation stems from incompatibility between many existing hospital information systems and new AI models, as well as differences in image acquisition protocols [12]. To implement AI models, hospital information systems must be properly updated, which is costly and labor-intensive. Furthermore, there are currently no unified protocols for AI integration into diagnostic workup, limiting its use in routine practice [13].

Model development and validation require additional expenditures. Collecting clinically relevant data, as well as testing and certifying models, requires significant investment from health facilities and model developers. This limits the availability of AI systems and prevents their use in less affluent clinics.

AI systems hold great promise in medical imaging; however, integrating them into clinical practice poses significant challenges. The risk of misdiagnosis, reliance on data quality, interpretability of decisions, and ethical and legal aspects must be considered. Successful AI integration necessitates unified AI model training and testing, as well as improved explainable AI methods, data confidentiality, and regulatory requirements [12, 13].

Only a comprehensive approach based on high-quality data, interpretable models, and secure infrastructure will enable AI to seamlessly integrate into the diagnostic workup, improving early detection of PDAC [13].

CONCLUSION

Significant progress has been made in early detection of PDAC using AI technologies. Current approaches include pre-imaging risk stratification and increased data volume through analysis of electronic medical records. They are frequently based on DL methods. Despite substantial achievements, clinical implementation of AI technologies remains challenging. New radiopharmaceuticals, imaging techniques, and biomarkers have been introduced; however, further research is needed to assess their role in the early diagnosis of pancreatic cancer. Furthermore, significant progress in early detection of pancreatic cancer, especially using AI, is impossible without research based on high-quality, representative data. Combined use of AI models and biomarkers is a promising area of future research to improve theranostics² in various cancers, including PDAC.

ADDITIONAL INFORMATION

Author contributions: F.T. Musaeva, Z.M.Kumykova, A.I. Ushaeva, A.S. Khasieva: published data search and analysis, writing—original draft, writing—review & editing; E.R. Sumenova, Ya.Yu. Kulinskaya, R.R. Yakupova, A.A. Mustafin: published data search and analysis, writing—original draft; A.Kh. Islamagulov: writing—review & editing; T.S. Elipkhanova: published data analysis, writing—original draft (certain manuscript sections); E.S. Ozerova: formal analysis, preparation of the final section of the article; D.A. Khusnutdinova: writing—review & editing; A.A. Nabiullina: conceptualization, supervision. All the authors approved the version of the manuscript to be published and agreed to be accountable for all aspects of the work, ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Ethics approval: Not applicable.

Funding sources: No funding.

Disclosure of interests: The authors have no relationships, activities, or interests for the last three years related to for-profit or not-for-profit third parties whose interests may be affected by the content of the article.

Statement of originality: No previously published material (text, images, or data) was used in this work.

Data availability statement: The editorial policy regarding data sharing does not apply to this work.

Generative AI: No generative artificial intelligence technologies were used to prepare this article.

Provenance and peer review: This paper was submitted unsolicited and reviewed following the fast-track procedure. The peer review process involved two external reviewers, a member of the editorial board, and the in-house science editor.

 

¹ The K-edge of absorption is a sharp increase in the absorption coefficient of X-ray or γ-radiation that occurs when the energy of the radiation exceeds the binding energy of an electron in the K-shell of an atom (the innermost shell).

² Theranostics is an interdisciplinary approach to medicine that combines diagnostics and treatment within a single technology or drug.

About the authors

Ferida T. Musaeva

North Ossetian State Medical Academy

Email: feridamusaeva@yandex.ru
ORCID iD: 0009-0000-1407-7189
Russian Federation, Vladikavkaz

Elizaveta R. Sumenova

North Ossetian State Medical Academy

Email: lsumenova@bk.ru
ORCID iD: 0009-0001-8159-0860
Russian Federation, Vladikavkaz

Almaz Kh. Islamgulov

Bashkir State Medical University

Author for correspondence.
Email: aslmaz2000@rambler.ru
ORCID iD: 0000-0003-0567-7515
SPIN-code: 8701-3486
Russian Federation, Ufa

Zalina M. Kumykova

North Ossetian State Medical Academy

Email: kumykova_2001@mail.ru
ORCID iD: 0009-0007-5243-6796
Russian Federation, Vladikavkaz

Tamila S. Elipkhanova

Maikop State Technological University

Email: eltamila01@mail.ru
ORCID iD: 0009-0006-2901-5443
Russian Federation, Maikop

Alina I. Ushaeva

Russian University of Medicine

Email: ushaeva21@list.ru
ORCID iD: 0009-0007-3888-5683
Russian Federation, Moscow

Amina S. Khasieva

Maikop State Technological University

Email: Khasievaamina999@gmail.com
ORCID iD: 0009-0002-8153-4647
Russian Federation, Maikop

Ekaterina S. Ozerova

The Russian National Research Medical University named after N.I. Pirogov

Email: ozerovaekaterina201@gmail.com
ORCID iD: 0009-0004-8740-1313
Russian Federation, Moscow

Dina A. Khusnutdinova

Kazan Federal University

Email: dinakhusnutdinova02848@gmail.com
ORCID iD: 0009-0002-0562-8414
Russian Federation, Kazan

Alina A. Nabiullina

Kazan Federal University

Email: a.ayratovnaa@gmail.com
ORCID iD: 0009-0004-4365-444X
Russian Federation, Kazan

Yana Yu. Kulinskaya

Russian University of Medicine

Email: Yana.Kulinskaya00@mail.ru
ORCID iD: 0009-0000-7187-0044
Russian Federation, Moscow

Roksana R. Yakupova

Bashkir State Medical University

Email: roksana.yakupova.01@mail.ru
ORCID iD: 0000-0001-5869-607X
Russian Federation, Ufa

Arthur A. Mustafin

Bashkir State Medical University

Email: zacartim@mail.com
ORCID iD: 0009-0006-4747-6972
Russian Federation, Ufa

References

Supplementary files