Artificial intelligence in pancreatic cancer histopathology and diagnostics - implications for clinical decisions and biomarker discovery?

Abstract

Artificial intelligence (AI) and machine learning (ML) are rapidly advancing fields within computer science, driving significant progress in cancer diagnostics. Various ML models have been developed to assist diagnosis, guide therapy decisions, and facilitate early disease detection. In this review, we discuss diverse AI and ML approaches and critically evaluate their applications and limitations in pancreatic cancer histopathology, diagnostics, and biomarker discovery.

Graphical abstract

Introduction

Artificial intelligence (AI) and machine learning (ML) have been introduced into various clinical fields over the past decades to enhance data evaluation and interpretation, as well as to manage increasing volumes of data [1]. AI is a field of computer science focused on developing systems that can mimic cognitive functions associated with human intelligence, such as the ability to identify hidden patterns in data, extract them and provide decisions based on prediction modeling [2]. ML is a subset of AI that allows a computational system to learn, self-correct and improve accuracy based on training and experience. ML uses mathematical and statistical models to obtain decisions autonomously, without direct instructions. High-quality data is essential for effective training of ML models. The more accurate and relevant the training data are, the better the model performs [3].

Generally, ML models offer robust learning ability and unique advantages in processing complex patterns in diverse types of input data. Therefore, they are ideally suited for image analysis tasks, especially in the field of digital pathology [4]. They can also combine a large amount of inherently highly diverse data, such as demographic and clinical parameters, biochemical values, radiology and histopathology image information and molecular‒pathological data from sequence analyses. The AI and ML can support diagnosis and enhance the selection of the most appropriate treatment according to the patient’s needs, bringing the concept of personalized medicine closer to the clinical practice [1]. In addition, the AI and ML prediction capabilities can improve the modeling of disease-free survival (DFS) and overall survival (OS) by integrating various input data and identifying relevant parameters related to patient outcomes [5, 6].

Pancreatic cancer

Pancreatic cancer (PC) covers a broad range of malignancies that develop from acinar, ductal or neuroendocrine cells within the pancreatic parenchyma. For a detailed review of pancreatic cancer pathology, classification, and staging, see Haeberle and Esposito [7]. The most common malignant form, pancreatic ductal adenocarcinoma (PDAC), is a highly aggressive disease with a median survival of 11.7 months [8] and increasing frequency in population. The absence of early symptoms frequently leads to the diagnosis of PDAC at advanced stages, significantly limiting therapeutic response and patient survival. PDAC exhibits significant morphological diversity, molecular heterogeneity, and complexity, which directly impact prognosis and treatment efficacy. Morphological and molecular variations are observed not only among individual PDAC patients but also within a single tumor. Sántha et al. demonstrated that the morphological heterogeneity of pancreatic cancer correlates with its structural and functional diversity [9]. This substantial heterogeneity significantly contributes to treatment resistance and failure [10]. Currently, surgical resection is a potential curative option available for only 15–20% of patients [11].

Prognosis of PDAC patients can be estimated based on overall clinical condition, previous medical history, nutritional status and eligibility to neoadjuvant therapy and/or surgical resection. Other high-risk biological and clinical features include, but are not limited to, the size of primary pancreatic tumor lesions exceeding 2–3 cm, presence of invaded lymph nodes, elevated Carbohydrate Antigen 19−9 (CA 19−9) levels, significant unintentional weight loss, and relevant comorbidities. For patients with borderline resectable or locally advanced adenocarcinomas, where immediate radical resection is not recommended, neoadjuvant chemotherapy is typically administered as a primary intervention, with subsequent consideration for resection. In cases of metastatic PDAC, first-line systemic chemotherapy is initiated. Approximately half of these patients will subsequently receive second-line chemotherapy [12,13,14]. Despite recent progress in understanding the pathobiology of pancreatic cancer and radiology imaging-based diagnostics [15] there is a need for new computational approaches and biomarkers that would credibly predict disease progression or reveal early cases in population screenings.

While AI and ML approaches have already been introduced in radiology image analysis, they offer the potential to integrate also other types of input data and enhance clinical decision-making or providing support for preventive programs. Park et al. evaluated the significance of early pancreatic cancer detection by developing a model that simulates incidence and mortality data in the U.S. The model included a hypothetical screening program that could reduce mortality. These findings suggest that screening for cancer 4–6 years before clinical diagnosis could lead to an average gain of 2.4 life-years [16].

The concepts and applications of AI in pancreatic cancer have attracted significant attention and have already been extensively reviewed [17,18,19,20,21]. In this work, we evaluated studies published from 2018 to April 2025. Inclusion criteria required the studies to report detailed information on ML model performance, involve at least 30 patients, and use PC/PDAC histological data as input for model training. Here, we aim to discuss recent advances in AI- and ML-based analyses of variable data inputs in PC, with a particular emphasis on histopathological diagnostics.

Learning schemes in computational histopathology of PDAC

The development of ML models for use in histopathology involves three principal approaches: supervised learning, unsupervised learning, and reinforcement learning. Currently, ML-assisted histopathological decision-making relies on supervised learning, which uses high-quality images that have been annotated by experienced pathologists to identify regions of normal and tumor tissue and typical dysmorphologies. In PDAC, morphological parameters such as normal tissue regions, invasive ductal structures, and the extent of desmoplastic fibroinflammatory stroma are defined within the training dataset. Additionally, the presence of PDAC histopathological variants, such as foamy cells, enlarged ducts, or vacuolated cells, can also be incorporated. Currently, various ML algorithms are designed to address specific problems in medical diagnosis. The concepts and categories of ML are illustrated in a Venn diagram (Fig. 1). AI methods for the analysis of histopathological images can be broadly categorized into two groups: traditional ML methods and deep learning methods. The biggest advantage of using deep learning methods over traditional methods is that they can process large and unstructured data such as images, text, and speech without the need to extract features manually. For a detailed description of the different ML models, their training and validation, see Greener et al. [22].

Fig. 1

Venn diagram demonstrating the principial hierarchy of AI models represented by machine learning algorithms, artificial neural networks and deep neural networks

Convolutional neural networks (CNNs) as a subset of Deep neural networks (DNN) are the often-used ML model in pancreatic histopathology to evaluate image data. CNN can recognize patterns directly from raw pixel images, requiring minimal preprocessing. CNNs can identify patterns with robustness to variability, distortion, and simple geometric transformations, such as rotations, translations, and scaling. This ability makes them highly effective for biological image-related tasks, e.g. object detection, image classification and segmentation, where the input data may vary in position, size, or orientation [23,24,25] or for analysis of whole slide images (WSI) [26]. This makes CNNs particularly valuable for analyzing complex, high-resolution tissue images, enabling more accurate and efficient diagnoses in pathology and the discovery of predictive biomarkers.

Applications of ML in diagnostics and prediction PDAC

Several studies have been published recently describing the use of ML for the early detection of pancreatic cancer. For example, Cao et al. used deep learning methods to detect and classify pancreatic lesions through non-contrast computer tomography. The model achieved the area under curve (AUC) value of 0.986−0.996 in detecting lesions and was not evaluated inferior to the average results of human radiologists [27]. Alternative approach was used by Placido et al. in their publication, who used longitudinal clinical records to predict the risk of pancreatic cancer. The AUC of the model for the training data was 0.88 for predicting the occurrence of cancer within 36 months and 0.83 for 3 months prior to diagnosis. The results for the validation set were lower, with AUC values of 0.78 and 0.76, respectively [28].

To date, only a limited number of publications have investigated the detection of PDAC using ML coupled with standard histopathology techniques (Table 1). The ML models achieved high reliability above 90%. In all studies, the metrics that are appropriate for unbalanced cohorts, F1 score and balanced accuracy, have been used to evaluate reliability. However, direct comparison of the models is limited, as the studies included different numbers of groups that the classification models were trained for, variability in how the groups were defined or a lack of clear group definitions, differences in the number of data sources and/or scanners used for training, and variations in the diagnostic evaluation of images by pathologists.

In the pilot studies, models are typically trained using distinct groups, such as cancer and normal cohorts, due to the clear separation between these groups. Similarly, in publications by Fu et al. and Naito et al. two groups, PDAC patients, and healthy donors, were analyzed. However, in both publications, the characteristics of the control group were not provided in detail, and the extent of physiological heterogeneity was not considered as well. Due to the class imbalance in the study by Fu et al., F1 scores provide a more appropriate metric for evaluating model performance. Although Naito et al. utilized a dataset more than twice as large, the reported model reliability is comparable. This similarity may be explained by greater heterogeneity within the control group in the Naito et al. study, which is described as ‘non-adenocarcinoma’, in contrast to the Fu et al. study, where the control group consisted of normal pancreatic tissue [29, 30]. The greater heterogeneity within the dataset may explain why a substantially larger number of images was required to achieve an F1 score exceeding 0.95. This also emphasized the need for external dataset validation, which neither of these studies included.

Kriegsmann et al. successfully applied ML for classification of patients into three categories: PDAC, benign pancreatic conditions, and pancreatic intraepithelial neoplasia. Patients in the study were recruited from two different centers, each using distinct types of scanners. Importantly, patients from both centers were equally distributed between the training and test datasets [31]. Despite analyzing a smaller patients’ cohort, classifying into three groups, utilizing data from varied scanners across two centers, and including PDAC patients at various disease stages, this model achieved results comparable to two earlier publications.

Table 1 Application of ML in computational histopathology of pancreatic cancer

Janseen et al. [32] used ML algorithms to evaluate presence or absence of residual disease in PDAC patients after neoadjuvant therapy (Table 1). This multicentric study used data from 14 centers using four different scanners. The aim of the model was to distinguish residual pancreatic cancer from nonneoplastic pancreatic ducts in routine Hematoxylin and Eosin-stained WSI. The model achieved the overall F1 score of 0.78, and the best reliability calculated for each type of scanner separately was 0.81 [32]. Reliability is lower than in the above publications, but in this case the dataset was composed from multiple sources and scan types, color augmentation and normalization was needed, and the goal of classification was also different.

In addition to PDAC diagnosis, ML models were also used to identify grades (Table 1). In the publication, Sehmi et al. classified images into four groups - Normal, Grade I, Grade II, and Grade IIl. Histological images were stained with May-Grünwald-Giemsa and Haematoxylin and Eosin. For learning, 14 pre-trained CNNs were used and then the model with the best reliability was selected, having an F1 score of 0.96 [33]. Ghoshal et al. used the same dataset as Sehmi et al. but aimed to determine the uncertainty in the prediction of PC stage by the model using Bayesian CNNs [34]. The different objective in developing the model is probably the reason why the results of reliability for Ghoshal et al.‘s model differ from those obtained by Sehmi et al.

AI was also used to determine the molecular subtype of PDAC, specifically the classical or basal subtype. The model achieved a balanced accuracy of 96.19% for classical subtype od PDAC and 83.03% for basal molecular subtype [35]. The major advantage of this study is that reliability has been validated using an external dataset. The reliability for classifying the basal subtype is lower, but this is most likely due to the fact that this group was less represented in the data (19.4% cases and 25.1% images).

These examples demonstrate how successful and reliable ML models can be. When the network is properly trained on a sufficiently large heterogeneous dataset, its reliability can be very high for both the validation and test sets. Such algorithms can play a significant role in clinical practice, assisting with the efficient evaluation of large numbers of samples. The significant heterogeneity among tumors can considerably impact model performance. This is evidenced by cases where even experienced pathologists cannot reach a consensus on the precise diagnosis [36].

ML-assisted biomarker discovery in pancreatic cancer

Timely diagnosis of PDAC allows earlier initiation of therapy, more accurate monitoring, and ultimately leads to improved patient survival. CA 19−9 is the most widely used biomarker associated with PC and PDAC [37, 38]; however, it lacks sufficient sensitivity and specificity for effective PDAC screening [39]. Therefore, there is a critical need to develop reliable, preferably non-invasive biomarkers for early detection of PDAC. Several studies reported biomarker development using different types of ML models, input data or sample source, reaching AUC values higher than 0.9 (Table 2).

Table 2 AI in the biomarker discovery for PC diagnosis

Mahawan et al. proposed a in silico ML pipeline for constituting composite biomarker signatures from public gene expression databases, and demonstrated its efficacy on 15 selected genes for prediction of risk of PDAC metastasis [40]. Iwano et al. introduced collective biomarkers, based on 36 peripheral blood serum metabolites, identified by support vector machine-driven analysis of liquid chromatography-mass spectrometry (MS) spectra [41]. Karar et al. proposed a combination model composed of one-dimensional CNN and long short-term memory that identifies patterns in urinary biomarkers to automatically diagnose PDAC [42] and Wolrab et al. used lipidomic profiling to discriminate between PDAC and healthy controls [48]. Multicentric proteomic analysis by Athanasiou et al. revealed serum protein signatures correlating with early stage PDAC and its subtypes [43]. Chen et al. created a model that integrates microbial signatures and CA 19−9 levels [47]. Firpo et al. [44] used several separate models for training in the first step and subsequently combined them into a single model using generalized boosted regression.

All ML models demonstrated promising efficacy; however, their direct evaluation is limited. Besides the type of input data and ML-model, the studies also differed in patient inclusion criteria. Some publications included patients in all stages of PDAC [40, 46,47,48] whereas others did not specify the stage [42, 45] or focused only on selected stages, namely T2 and T3 [41] or stages I and II [43, 44]. The definition of the control group also varied in studies, ranging from “healthy controls” [40,41,42, 45,46,47,48] which were not fully described in all cases, or including patients with chronic pancreatitis, family history of PDAC or genetic predispositions [43] and patients with chronic pancreatitis and intraductal papillary mucinous neoplasm [44]. Karar et al. stratified the dataset into three classes for ML model development– PDAC, healthy controls, and benign hepatobiliary disease [42]. A valid experimental design is therefore a critical prerequisite for reliable, reproducible and comparative outputs.

In summary, ML represents a powerful approach for identifying novel biomarkers to facilitate the detection of pancreatic cancer. As illustrated by the examples above, ML models can attain high accuracy despite considerable differences in model design or methodology.

Integration of input data from multiple modalities

A promising approach to achieve even more accurate predictions or prognoses in PDAC uses a combination of different data sources as inputs to an AI model. In cancer, histological data can be integrated with immunohistochemical, metabolic, genetic, and radiographic data. This integration can increase the effectiveness of clinical workflows, improve diagnostic accuracy, and ultimately lead to personalized diagnosis and treatment plans for patients. Chen et al. reported using used weakly supervised, multimodal deep-learning algorithm for successful prediction of prognosis of 14 cancer types including PDAC based on WSI (Table 3).

Table 3 ML models combining multiple types of data for PAAD

This study highlighted the importance of using multiple data sources, specifically a combination of histology images and molecular profile data, to predict outcomes and reveal prognostic features that correlate with poor and favorable outcomes in 14 cancer types. Data from 5720 patients was used to construct the model for all evaluated cancer types. Models for each single type of cancer have also been established. To create the model for PDAC, 166 patients were used. The C index of the whole dataset was 0.645, compared with 0.585 for the model using histological data alone and 0.607 for the model using molecular features alone. Similarly, the survival AUC improved from 0.616 for the model using only molecular features and 0.598 for histological images to 0.662 when both were combined. These results are similar to the model evaluations for PDAC, which are presented in Table 4 [49]. For both models, a combination of both datasets gave the best results.

In other types of cancer, Hou et al. reported a similar conclusion, using a multimodal model to predict the survival of hepatocellular carcinoma patients. Dataset used in the study was constructed using WSI and mRNA expression levels from publicly available databases. The prediction of OS based on both datasets achieved a C index of 0.746 and surpassed that of models that were based only on either histopathological data (0.714) or genetic information (0.666) [50]. Mobadersany et al., used a combination of genetic biomarkers and histology images from 796 samples as inputs to a CNN to predict the survival of patients with gliomas. The best model performance was achieved by combining data from different sources leading to improved prognostic prediction probability [51]. In contrast, Höhn et al., developed a skin cancer classifier to distinguish between melanomas and benign nevi, and reported that its performance was comparable or even better when trained on a single dataset versus a combination of two datasets. Here, 430 WSI and basic patient data (age, sex, anatomical site of lesion) were used for construction of CNN. The AUC of the patient data model was 0.770, and the balanced accuracy (BA) was 0.720, while the WSI model achieved an AUC of 0.923 and a BA of 0.832. Using both datasets, the best approach achieved an AUC of 0.896 and a BA of 0.830 [52]. Although the differences in performance between using WSIs and both datasets are minimal, the study underscores that incorporating multiple datasets does not necessarily enhance classifier accuracy.

Therefore, including low- or non-informative variables from different data sources does not necessarily improve the accuracy of a classifier. In that case, a decrease in model reliability usually results from overfitting, particularly in models built with numerous uncorrelated variables. Initially, adding features can enhance accuracy, but beyond a certain point, it may lead to a decline. In fact, including too many features can cause the model to fit noise rather than meaningful patterns. This is especially relevant in supervised learning, where the number of samples should be proportional to the number of parameters [53, 54].

Challenges and limitations

A major challenge in training ML algorithms is the requirement for large volumes of precisely annotated (labeled) data [55]. Moreover, annotations are difficult to obtain, as they are time-consuming and must be performed and validated by expert pathologists. A possible solution to this problem is the use pretrained networks, e.g., ImageNet by Stanford Vision and Learning Lab [56] followed by additional training on a smaller set of annotated images [57]. This approach streamlines and accelerates the training process, although it does not always result in better performance than a network that was trained from the beginning [58]. Additionally, the features representations learned by pretrained ML models may not be well-suited for all histopathological images. To overcome this limitation, disease-specific digital pathology datasets have been proposed and made publicly available [59,60,61,62].

Most ML prediction models typically distinguish PDAC patients from either healthy individuals or those with other types of lesions. However, it’s crucial to consider various pancreatic changes that differ from PDAC [20, 63]. Therefore, a computational tool capable of discriminating not only PDAC but also other distinct pancreatic diseases is highly needed. Unbalanced cohorts within the datasets then pose another major challenge, potentially leading to biased accuracy estimates if an appropriate evaluation metrics are not properly applied. Furthermore, the histopathological heterogeneity of PDAC likely contributes to the variability in treatment responses [64]. Consequently, improving the performance of ML models will require a large, diverse dataset representing different PDAC subtypes and forms. Future development ML models should prioritize patient staging, rigorously correlating it with histopathology and OS. Additionally, investigating the response to neoadjuvant/adjuvant chemo(radio)therapy and exploring multiple treatment options is highly relevant [18].

Generalization remains a challenge for ML models, which often perform well on training data but poorly on data from different sources. This issue represents one of the greatest challenges to the widespread implementation of AI in medicine. Carrillo-Perez et al. used two datasets in their publication, where one contained histological images of PDAC patients and the other a control set, and showed that the model trained on the dataset created this way achieved accuracy of 99.04%, and when the model was evaluated on a different dataset, achieved an accuracy 67.20% [65]. To minimize this problem, it is crucial to create a training set that includes as much variability as possible, incorporating data from different centers, instruments, and varying staining protocols. In addition, Carrillo-Perez showed that the model trained on dataset integrating multiple sources does not preferentially discriminate between the classes, but rather between the datasets, indicating that stronger normalization would be needed when using datasets from various sources [65]. A similar problem was successfully addressed by Pečinka et al. for MS data recorded from whole cancer cells [66]. Since the problem of model generalization can often result from overfitting, it is necessary for publications to clearly state how this issue has been addressed. Authors should explicitly describe the methods and strategies implemented to prevent overfitting, such as cross-validation techniques, regularization methods, or the use of external validation datasets. Additionally, it is important to outline the specific measures implemented to minimize overfitting, and therefore ensuring the robustness and reliability of the reported performance of the ML model. Another technical issue significantly affecting AI analysis, but usually not the interpretation by a human pathologist is the color variability of histological samples. Some studies have shown that deep learning algorithms’ ability to segment or classify histological images decreases when there is high color variability or inconsistency within the training set [67, 68].

Predictive algorithms are inherently limited to recognizing only the patterns they have been trained on. While ideally an algorithm would be trained to identify all potential pathological conditions, this approach is often extremely difficult or even impossible. A practical alternative is to develop AI methods that, in addition to providing predictions, also generate a certainty score for each prediction [69]. When WSI are used to train algorithms, their immense size limits computational efficiency, as these gigapixel images are too large to fit into a Graphics Processing Unit’s (GPU) memory. Several approaches address this issue; a common method assumes that all patches—small tiles into which the image is divided—contain morphological information correlated with the label of the entire image [70].

The integration of ML models into clinical practice for PDAC remains in the early stages. Baxi et al. outline a phased framework for the clinical implementation of AI technologies. The initial phase involves validating the model’s reliability, a process overseen by a pathologist to ensure diagnostic accuracy. Contingent upon successful validation, ML applications may subsequently be introduced as adjunctive diagnostic tools. This would be followed by prospective clinical trials designed to evaluate model performance in real-world settings. In the final phase, the model would undergo continuous refinement and optimization through the incorporation of expanding datasets [71].

Regulatory and ethical consideration

The application of AI in medical practice raises important ethical considerations, particularly in relation to data confidentiality, patient privacy, data security, and informed consent. Since a large amount of sensitive information needs to be collected from patients and physicians to train and validate the model, it is necessary to consider the ethical aspects of data access from the perspective of public interest but also from the perspective of preserving the right to privacy of an individual [21, 72]. At the same time, data security should be considered to minimize the risk of data loss or hacking, as this could cause harm to patients [73, 74]. Another problem in the implementation of AI in medical practice is the absence of legal standards that address the issue of responsibility in case of misdiagnosis by ML-based tools, as they can have fatal consequences [19, 75]. Furthermore, bias in the model development process may also be an important obstacle. If the training set is not sufficiently diverse, biases in terms of demographic and socio-economic factors may occur. Over or under representation of some groups can cause poor model performance, incorrect predictions and even harm a patient [73, 76].

Conclusion

In recent years, various AI and ML based strategies have been developed to support diagnostic evaluations. With the growing availability of data, these models have begun to match—and in some instances exceed—the diagnostic performance of human experts. The integration of multimodal data sources has further enhanced the predictive accuracy and classification capabilities of ML models, offering improved clinical decision support. These advancements have the potential to reduce diagnostic turnaround times and assist in complex cases where definitive diagnoses are difficult to establish. While substantial barriers to routine clinical implementation remain, AI and ML represent a promising innovation with the potential to transform PDAC histopathology, streamline identification of novel biomarkers and improve prognosis of cancer patients.