Development and validation of a diagnostic model based on artificial intelligence for diagnosing pancreatic cancer

Introduction

The global incidence and mortality rates of pancreatic cancer (a malignant tumor originating from the pancreatic epithelium) have been increasing in recent decades. This trend can be attributed primarily to the increasing prevalence of lifestyle-related risk factors, including advancing age, smoking, increased alcohol consumption, and obesity (1,2). Most patients with pancreatic cancer are diagnosed at advanced stages of the disease, while much less frequently tumors are detected as incidental findings, during instrumental diagnosis (imaging) for other reasons, or in patients with high carbohydrate antigen 19-9 (CA 19-9) levels.

The most common form of malignant tumor is pancreatic ductal adenocarcinoma (PDAC—a malignancy that is formed in the exocrine portion of the pancreas), which accounts for approximately 85–95% of cases. This type of cancer develops in the exocrine part of the organ and has an extremely low 5-year survival rate of about 9% (3). Pancreatic adenocarcinoma can emerge from various types of precursor lesions, including pancreatic intraepithelial neoplasia (PanINs—precancerous neoplasm that develops in the small ducts of the pancreas), intraductal papillary mucinous neoplasia (IPMNs—precancerous neoplasm developing in the ductal system of the pancreas, consisting of cells that form papillary structures and produce mucin) and mucinous cystic neoplasms (MCNs) of the pancreas (precancerous neoplasms characterized by the presence of mucinous epithelium forming cystic structures surrounded by an epithelial membrane with the potential for malignant degeneration). This malignant transformation is instigated by a series of sequential genetic changes that exhibit substantial diversity. The timely detection of cancer at its early stages stands as one of the foremost priorities in oncology. Successfully addressing this challenge holds pivotal significance in modern medicine.

In this context, medical imaging plays an important role in detecting malignancies at all stages, from initial screening to monitoring the effectiveness of treatment. Currently, one of the most common methods of non-invasive detection of cancer is computed tomography (CT) with contrast. The data obtained with the help of these methods of radiation diagnostics allows assessment and interpretation of a range of tumor characteristics, i.e., size, shape, contour, degree of infiltration into adjacent tissues, arterial and venous invasion, regional lymph node involvement, as well as distant metastases. Standard imaging is often not informative enough for accurate diagnosing, and detecting cancer based on CT images “in manual mode” is a long process and requires a high level of expertise in radiology and oncology. High accuracy in diagnosing the cancer stage, resectability, and detection of metastases are necessary prerequisites for medical decision-making [neoadjuvant treatment, surgical treatment, chemoradiotherapy, palliative chemotherapy, targeted therapy and immunotherapy (4,5)]. In this regard, the task of introducing new diagnostic methods that fully utilise the achievements of modern information technologies is particularly acute.

Modern technologies of CT and magnetic resonance imaging (MRI) image processing based on mathematical algorithms of image analysis and deep learning (DL) are efficient tools for diagnostics. Mathematical algorithms for analysing medical images make it possible to identify subtle structural features associated with the pathology of disease-affected organs, and trained artificial intelligence (AI) associates these structural features with one or another disease phenotype. The efficacy of this method of diagnostics depends on both the mathematical algorithms used for image processing and the organisation of the AI training procedure. As a result, the integration of intelligent technologies significantly improves the accuracy of determining the stage of a tumor, the possibility of surgical intervention and the detection of metastases, which, in turn, contributes to more informed and timely clinical decisions.

Radiomics is an evolving field within medicine that assists in extracting quantitative data (density, shape, texture, etc.) from medical images and provides information for forecasting processes underlying tumor biology and intratumoral heterogeneity (6). Radiomics encompasses extracting manually designed attributes (shape, intensity, texture, and wavelets) from images, often focusing on segmented areas of interest (7). Quantitative imaging allows radiomics and dynamic imaging features to be considered individually or in combination, enabling clinical prediction models to be built based on radiomic signatures or imaging phenotypes, and allowing the assessment of clinical outcomes related to tumor biology (8).

The most effective method for training AI is DL (type of machine learning that incorporates algorithms based on the structure and function of the brain or neural networks). Neural networks possess the ability to learn and adapt from data on their own through training by data analysis (9). U-Net is a convolutional neural network that has a U-shaped architecture which allows fast and accurate image automatic segmentation. U-Net has been widely used for performing various medical imaging tasks, including the segmentation of organs, tumors, and other structures within images. Its success lies in its ability to process complex image structures and generate accurate segmentation masks, making it a powerful tool for medical image analysis. In recent years, there has been rapid progress in the use of AI in the field of medicine. AI-based models are used for screening, differential diagnosis, treatment, prediction of patients with pancreatic cancer. Together, these technologies can not only improve the accuracy of diagnosis, but also contribute to the development of a personalized approach in oncology.

Related work

To assess the current state and development of methods for automatic segmentation of medical images in the diagnosis of pancreatic cancer using AI, it is important to analyze existing research in this field. The reviewed works include the development of decision support systems based on DL models, as well as the use of optimization algorithms to improve the accuracy of segmentation and classification.

Vaiyapuri et al. in their work, the co-authors presented an intelligent medical decision-making system with DL support for the classification of pancreatic tumors intelligent deep-learning-enabled decision-making medical system for pancreatic tumor classification (IDLDMS-PTC) using CT images. The proposed IDLDMS-PTC method includes several subprocesses, namely preprocessing based on Gaussian filtering (GF), segmentation based on emperor penguin optimizer with multilevel thresholding (EPO-MLT), feature extraction based on MobileNet (a family of lightweight DL models), auto encoder (AE) classification, and parameter optimization based on multileader optimization (MLO). Using the EPO method for optimal threshold selection and the MLO algorithm for parameter tuning helps to achieve improved classification results. To evaluate the effectiveness of the IDLDMS-PTC method, a comprehensive experimental analysis was performed on a reference dataset. Extensive comparative results have revealed the promising performance of the IDLDMS-PTC model compared to existing methods (10). Gandikota et al. in their work developed the tunicate swarm algorithm (TSA) using the TSA with DL-based pancreatic cancer segmentation and classification (TSADL-PCSC) DL and classification method. To achieve this, the TSADL-PCSC method includes four processes, namely W-Net segmentation, GhostNet feature extractors, deep echo state network (DESN) classification, and TSA-based hyperparameter tuning. The experimental result of the TSADL-PCSC method was tested on a reference database of CT scans. The results obtained confirmed the higher accuracy and efficiency of TSADL-PCSC compared to other methods (11). Bagheri et al. used a deep convolutional neural network (DCNN) to segment the pancreas in a publicly available dataset, and also evaluated the impact of technical and clinical factors for pancreatic segmentation training. Using Dice similarity coefficient (DSC—is a statistical metric used to assess the similarity or degree of overlap two data sets), the segmentation accuracy was evaluated. The average rate of DSC segmentation of the pancreas was 78%±8%. Factors that significantly correlated with DSC were also identified: body mass index (BMI), visceral abdominal fat, pancreatic volume, and median and average CT attenuation in the vicinity of the pancreas (12). Khdhir et al. developed the Adaptive Learning Optimization-Convolutional Neural Network-Gated Recurrent Unit (ALO-CNN-GRU) mechanism for segmentation and classification of pancreatic cancer based on Antlion optimization and DL. Medical images are pre-processed to reduce noise. Segmentation to determine the affected area of the pancreas was processed by the Antlion optimization algorithm (ALO). Segmentation performed using the CNN model classifier and the Gated Recurrent Unit [GRU—a type of recurrent neural network (RNN)] (13). Nishio et al. presented and evaluated a combination of DL architectures and data augmentation methods for automated pancreatic segmentation in CT scans. Deep U-Net and basic U-Net were selected for pancreatic DL segmentation algorithms. Data augmentation methods included random cropping and patching (correction) of images [random image cropping and patching (RICAP)], mixup, and the traditional method. The best average DSC value was obtained using a deep U-Network with RICAP and mixup (14). Yang et al. presented the AX-Unet DL platform (DL architecture) for segmentation of pancreatic CT images. In AX-Unet, based on the U-Net encoder-decoder structure, the authors included a modified atrous spatial pyramid pooling module (ASPP) to obtain location information. The results of experiments on two publicly available datasets confirmed the superiority of the proposed AX-Unet model over modern methods (15). Zhao et al. studied the current application and prospects of AI in pancreatic cancer. The author, after analyzing the work on the use of AI in pancreatic cancer, came to the conclusion that despite the numerous positive prospects for integrating AI in the field of pancreatic cancer, certain limitations inevitably arise: insufficient transparency of DL models, difficulty interpreting results, and generalization problems when working with heterogeneous datasets. It is also noted that limited sampling negatively affects the stability and effectiveness of training (16).

In our work, we propose improving the U-Net architecture to development and validation of a new DL system for segmentation of pancreatic images on CT, which should contribute to more accurate detection of tumors pancreatic and increase the effectiveness of screening. We present this article in accordance with the TRIPOD reporting checklist (available at https://cco.amegroups.com/article/view/10.21037/cco-25-17/rc).

Methods

General information about the data used

The initial stage in designing the diagnostic model involved selecting data for training the neural network. For the segmentation of medical images and algorithm training, segmentation data from 280 studies of pancreatic cancer (multi-slice CT, axial sections, portal-venous phase of intravenous contrast) from an open source database—the Memorial Sloan Kettering Cancer Centre (17) were used. The source does not include clinical patient data. Segmentation was performed on CT images of 30 patients with pancreatic neoplasms (16.66%—pseudotumoral pancreatitis; 16.66%—benign pancreatic neoplasm; 66.66%—malignant pancreatic neoplasm: 15% of cases—stage II, 75% of cases—stage III–IV, respectively, 90%—pancreatic adenocarcinoma, 10%—neuroendocrine tumors of the pancreas). All these cases of pancreatic neoplasms were detected during examination at Regional Clinical Hospital of the Kaliningrad. The diagnoses within the patient cohort were confirmed through morphological diagnostics involving the examination of biopsy samples and/or surgical material. The image data were used for the algorithm development and testing. The amount of data for neural network training was determined based on 280 studies, which provided sufficient representativeness and variability for training a complex segmentation model. Additionally, images of 30 patients with confirmed pancreatic tumors were used to test and evaluate the accuracy of the model, which made it possible to test the quality of automatic segmentation and predicted parameters. Using the specialized three-dimensional (3D) Slicer software (version 5.0.2, www.slicer.org) the segmentation of medical images was performed manually. The study was conducted from 2023 to 2025.

The diagnostic model being developed, based on the U-Net trained neural network, is designed to automatically diagnose the presence of a malignant pancreatic tumor based on the analysis of CT images of the abdominal cavity. The assessment is carried out by analyzing CT images taken in the portal-venous phase of contrast using a trained neural network (U-Net) that segments areas with possible neoplasms.

Training and testing of the diagnostic model

To form the training and test samples of the algorithm, the conversion of medical image formats was carried out: from the source data presented in neuroimaging informatics technology initiative (NIfTI) and digital imaging and communications in medicine (DICOM) formats (in grayscale), a transition was made to the markup of slices stored in portable network graphics (PNG) format. This made it possible to record two classes (the pathology itself and the organ under study), using a separate channel for each.

Tumor segmentation can be performed manually, semi-automatically, or automatically. Manual segmentation is performed by an experienced radiologist and is based on expert knowledge. It is considered reliable but time-consuming and subject to variability between experts. In our study, CT images were annotated by several oncologists and radiologists to eliminate the risk of an error and to increase annotation reliability. After verification, the images were added to the database and used to test the accuracy and efficacy of the neural network algorithm. During the image markup process, multiple scenarios were taken into account, including combined markup of both the pancreas and the tumor, as well as separate markup pancreas and separate markup tumor.

In several numerical test experiments on an unchanged U-Net, it was shown that using a standard number of convolution cores and layers leads to overfitting of the network and, as a result, a decrease in its efficiency. In this regard, a truncated version of the network was created with a reduced number of layers and cores, this reduced the risk of overfitting and increased stability. To increase the accuracy, a post-processing unit was added to the algorithm, which makes a decision about the presence of a tumor in a given area. A normalization step was implemented according to the following formula:

fnorm(SM,tr,h,w)={1,ifSMh,w−min(SM)max(SM)−min(SM)>tr[1]

where SM is the segmentation map, min(SM) is the minimum estimate of the probability of a tumor, max(SM) is the maximum estimate of the probability, tr is the threshold, h,w are the height and width, respectively.

To enhance the quality of the dataset (enhance the contrast of the original image, remove artefacts and markers, and normalize and centre the image) required for training the neural network based on U-Net, an advanced preliminary assessment software module was developed. This algorithm, tailored for CT images of varying complexity, comprises three core elements: segmentation, computation, and background removal. The primary function of segmentation is a pixel-by-pixel assessment of a diagnostic image to detect the presence of the patient’s body. Upon detection of the patient’s body image at a given pixel location, the algorithm assigns a value of 1; conversely, if the patient’s body is not detected, the value assigned is 0.

The image background computation and removal module functions by computing the average brightness of the image and subsequently comparing each pixel with possible artefacts. This module enables the elimination of background disturbances, such as imaging artefacts and inadvertent inclusion of tomograph-related structural elements, along with other sources of interference. This results in a clean image background. The process is composed of the following stages: minimax normalization (this process ensures uniformity in pixel intensity values across the image, preparing it for subsequent enhancements), non-linear contrast enhancement (it accentuates significant features within the image while minimizing less critical details) and histogram equalization (redistribution of pixel intensity, enhancing details and structures that might have been obscured by variations in lighting or inherent imaging imperfections). Cumulatively, these pre-assessment stages refine the image, optimizing it for subsequent analysis by the neural network algorithm. This comprehensive process ensures quality preparation, enhancing the accuracy of diagnostic interpretations and the efficacy of decision-making in medical image analysis.

In the process of refining the neural network algorithm, targeted enhancements were made to optimize its performance. One major enhancement involved introducing augmentation, which generates artificial training examples from natural images. By utilizing preprocessing, augmentation, and fresh annotated data, the neural network underwent further training and refinement. During the refinement phase, a self-learning algorithm was developed. Reinforcement learning technology was selected as the primary approach for enhancing the algorithm’s capabilities autonomously. As part of adapting the final algorithm for real-world medical applications, a technical module was integrated. This module recalculates the size of tumors in centimetres, utilizing information from CT scans. This ensures accurate alignment with clinical dimensions. As a result, tumors measuring less than 1 cm were detected.

To ensure the objectivity of the predictor quality assessment, blind testing of predictors for the final result and other auxiliary predictors was conducted. Radiologists and oncologists evaluated the images and the results of automatic segmentation, as well as predictor parameters (tumor sizes). The entire assessment took place under conditions hidden from the automatic system and the results of the model in order to eliminate any possible biases. The assessment was carried out at the stage of final verification of the model, after completion of training and testing within the research team.

Statistical analysis

To evaluate the effectiveness of the diagnostic model, a statistical analysis of the results of the algorithm was performed in the study of CT images of abdominal organs with pancreatic cancer and CT images without pancreatic cancer. False positives were primarily observed in regions adjacent to large blood vessels during the portal-venous phase. The analysis showed that the risk of false positives is associated with similar textural characteristics of healthy vascular tissues and surrounding structures, which complicates their differentiation. Predictors (segmentation results, probabilistic estimates of the presence of a tumor, textural and radiological parameters) were used in the analysis to assess the accuracy and effectiveness of the model. They were processed using automatic methods and independently verified by experts, and their quality and impact on the study results were analyzed using statistical metrics (DSC, sensitivity, specificity, accuracy).

In order to increase the objectivity and reliability of evaluating the effectiveness of the developed diagnostic algorithm, the results of segmentation and determination of the presence of a tumor in medical images were independently evaluated. To do this, the specialists who analyzed the images did not have access to the results of the trained neural network model, which eliminated the impact during the initial markup. Similarly, in the subsequent assessment of the final diagnostic results, the experts conducted an independent assessment without knowledge of the automatic conclusions of the neural network. The results of the expert evaluation were compared with automatic predictions to determine the accuracy, sensitivity, specificity, and other characteristics of the diagnostic model.

Four types of metrics were used to quantify segmentation results: DSC, accuracy, sensitivity, and specificity.

• The Dice metric is calculated using the following formula:

  $$D=\frac{2}{N}{\displaystyle \sum _{n=1}^N\frac{{\displaystyle {\sum }_{h=1}^H{\displaystyle {\sum }_{w=1}^W{L}_{h,w}^{\left(n\right)}\cdot f_{norm}\left(S{M}^{\left(n\right)},tr,h,w\right)}}}{{\displaystyle {\sum }_{h=1}^H{\displaystyle {\sum }_{w=1}^W\left[L_{h,m}^{\left(n\right)}+f_{norm}\left(S{M}^{\left(n\right)},tr,h,w\right)\right]}}}}$$[2] where (n) is the element number of the test sample, f_norm (SM,tr,h,w) is the normalization function that translates the probability matrix of the presence of a cancerous tumor in a given area (i,j) using the threshold tr, into label 1, if there is a tumor, and into label 0 in the opposite case.

  $$f{}_{norm}\left(SM,tr,h,w\right)=\{\begin{array}{cc}1,& \text{if}\frac{S{M}_{h,w}-\min \left(SM\right)}{\max \left(SM\right)}>tr\\ 0,& \text{else}\text{if}\frac{S{M}_{h,w}-\min \left(SM\right)}{\max \left(SM\right)}\le tr\end{array}$$[3]

• Sensitivity—the probability of correctly identifying the presence of disease cancer. It is calculated as the ratio of true positive results to the sum of true positives and false negatives.
• Specificity—the probability of correctly recognizing the absence of disease. It is calculated as the ratio of true negative results to the sum of true negatives and false positives.
• Accuracy—the overall proportion of correctly classified cases among all examined cases. It is calculated as the ratio of the sum of true positive results and true negative results to the total number of cases studied.

In further work to improve the diagnostic model, it is planned to expand the sample and cross-check using various performance indicators such as the receiver operating characteristic (ROC) curve and the area under the curve (AUC) in order to more accurately determine the boundary of the optimal classification threshold. In addition, it is planned to use statistical analysis to assess the impact of the introduction of additional functions, such as textural and radiological parameters, which can improve the accuracy of the system in the diagnosis of pancreatic cancer.

Ethics statement

The Memorial Sloan Kettering Cancer Center database is a publicly available dataset, and all information about the participants has been depersonalized. The current study used anonymous data extracted from a public database. Thus, the approval of the ethics committee was not required. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The permission of the Ethics Committee of Regional Clinical Hospital of the Kaliningrad was obtained for the use of data from Regional Clinical Hospital of the Kaliningrad (protocol No. 7-23 dated 05.10.2023). Informed consents (consent to participate and consent to publish) were obtained from all participants (all participants are over 18 years old).

Results

Structure and description of the diagnostic model

The block diagram of the diagnostic model is shown in Figure 1.

Figure 1 The block diagram of the diagnostic model.

The input data for the developed diagnostic model comprise 310 CT images (multi-slice CT, axial sections, portal-venous phase of intravenous contrast). The training and testing datasets were formed using a random selection approach, allocating 90% of the data for training and reserving 10% for testing purposes. Before inputting CT data into the neural network for analysis, a preliminary pre-processing stage takes place. It involves image centering and scaling, followed by the creation of an a priori probability mask that identifies the presence of a tumor.

During centering and scaling, the first step entails calculating and eliminating the background from the image. Background removal is achieved by computing the average brightness along rows and columns. Threshold values of 0.08 for rows and 0.06 for columns are applied, this ensures accurate localization of the area of interest in the image. Subsequently, images are resized to a 256×256 pixel resolution. The module tasked with generating the a priori probability mask, indicating the potential presence of a cancerous tumor, is designed to compute this mask for the centred image. A mask was created that represents a monochrome image and is calculated from centered images for each pixel:

Mi,j=1255⋅N∑n=1NLi,j[4]

M^i,j={1ifMi,j(M)>0.52elseifMi,j(M)≤0.5[5]

where, M_i,j is an estimate of the a priori probability of detecting a tumor in a given area of the image (the mask of the a priori probability), N is the number of examples in the dataset, and L is the segmentation maps from the dataset.

This mask adopts a monochrome image format and is computed for each pixel across the entire image. The mask serves a dual purpose. Firstly, it functions as an attention map through multiplication with the input image. This process effectively emphasizes areas of interest within the image. Secondly, the mask plays a role in incorporating the probability of detecting a cancerous tumor into the operation of the neural network algorithm. Essentially, the mask aids the neural network in considering the likelihood of tumor detection during its analytical processes.

An input tensor with dimensions 256×256×2 is generated by combining the input image and the outcome of multiplying the mask with the image. This input tensor is then directed to the neural network’s first input. At the same time, the logarithm of the a priori probability mask, denoting the potential presence of a cancerous tumor, is fed into the second input of the neural network. Combining these two data provides the model with information not only about visual signs, but also about the likelihood of a tumor.

The neural network used in the diagnostic model CT image analysis system is based on the U-Net architecture. This architecture comprises a sequence of subsampling and convolutional layers. In the modified version of the U-Net architecture used, the main differences are a reduction in the number of convolutional cores (to reduce overfitting), the use of an a priori probabilistic mask and a classifier to determine the presence of a cancerous tumor on a slice of a computer image. Furthermore, batch normalization layers were employed to counteract internal covariance shifts. The input layer of neural network information is processed by a specialized encoder, which is essential for extracting diverse complex variations of textural features. Subsequently, these features are input into both the cancer classifier and the segmentation algorithm. Each component within the neural network architecture is tailored to address a specific range of tasks, and the algorithm’s seamless operation as a whole result from the combined performance of all elements. The architecture of the neural network is shown in Figure 2.

Figure 2 The architecture of the neural network.

Training and testing of a diagnostic model for analysing CT images

The output of the neural network yields two primary results: an assessment of the likelihood of a cancerous tumor and a segmentation map. Once texture features are extracted within the encoder, the workflow branches into two distinct directions. To calculate the probability estimate, a flattening layer is introduced, transforming the third-order tensor into a vector. This layer plays a pivotal role in aligning the convolutional layer output with the fully connected layer, which utilizes a logistic activation function.

The chosen loss function is binary cross-entropy, selected due to the implementation of the sigmoid activation function. This strategic choice simplifies the achievement of a global minimum, significantly enhancing and facilitating the training process. Segmentation accuracy is calculated as the ratio of correctly classified pixels to total pixels, and the tumor detection probability is computed similarly. Through a sequential numerical experiment, values for binary cross-entropy were computed for both the training and validation datasets. It is noteworthy that there is an inverse relationship between the training epoch number and the magnitude of binary cross-entropy, as illustrated in Figure 3.

Figure 3 Dependency of errors on training epochs.

The term “training epoch” denotes a complete cycle of training a neural network using all the available training data. In our study, the accuracy of segmentation was initially computed as the ratio of segmented pixels to the total pixels in an image. Similarly, the probability of tumor detection for the classifier was also determined using the same approach. However, subsequently, the Dice coefficient was chosen as the preferred metric to assess the quality of segmentation. This metric has been recognized in various sources as a dependable indicator of the effectiveness of the segmentation algorithm.

To comprehensively evaluate the neural network’s performance, a comparison was conducted between the segmented tumor regions generated by the neural network and the oncologist-annotated segmentation from the test dataset, prior to any normalization adjustments (specifically, using the function f_norm). As demonstrated in Figure 4, the segmented areas exhibit a high degree of alignment.

Figure 4 Comparison of segmentation performed by a medical specialist (left) and the neural network algorithm (right).

It is important to note a change in segmentation methodology. The previously employed binary cross-entropy loss function was replaced by the Dice loss function. This change in approach yielded a classification accuracy of 88% for pancreatic cancer classification and a segmentation accuracy of 70% for pancreatic tumors. The sensitivity of the diagnostic model algorithm was 80%, and the specificity was 96%. To improve the statistical efficiency of the model, a neural network algorithm was further trained based on CT scans of patients with pancreatic cancer obtained at Regional Clinical Hospital of the Kaliningrad. The retraining process is implemented in a fine-tuning mode, in which the specialist manually annotated 20% of the training sample, explicitly indicating the location of the tumor and areas of its absence. These annotations were used for subsequent training of the model on the remaining 80% of the data, taking into account the previously mentioned 20% of the sample. As a result of the application of this approach, sensitivity and specificity indicators were achieved at the level of 98%, and the classification accuracy was 88%.

This system provides an iterative learning mechanism: after the initial image analysis, the doctor identifies questionable slices and, if necessary, clarifies annotations, additionally training the algorithm on the data obtained. In this way, the process of self-learning of the neural network model is implemented in practical use, which helps to increase its diagnostic value.

Discussion

Analysis of the results and possible causes of errors

In this study, a neural network model based on U-Net was developed and integrated to analyze CT images in the diagnosis of pancreatic cancer. In particular, the algorithm achieved a segmentation accuracy of 70% according to the Dice metric, which indicates a sufficient degree of overlap between automatic segmentation and expert opinions. The achieved indicators—classification accuracy of 88%, sensitivity and specificity of 98%—indicate the effectiveness of the proposed model. The results confirm the high potential of the developed model for the detection and segmentation of pancreatic tumors on CT images. Similar outcomes have been observed in various recent works (18,19).

It is important to understand which clinical and technical factors can influence the quality of pancreatic segmentation. In the course of the work, a few isolated occurrences of false positives were observed, primarily in areas adjacent to significant blood vessels during the porto-venous phase of contrast enhancement. This is probably due to the fact that healthy vascular tissue and surrounding structures may have textural features similar to those of a tumor, which makes it difficult to accurately differentiate. Other studies have also noted that the accuracy of segmentation can be influenced by other factors such as body mass index, visceral abdominal fat, pancreatic volume, standard deviation of CT attenuation within the pancreas, as well as the median and average CT attenuation in its immediate vicinity (12).

In the future, it is planned to conduct a more extensive evaluation of the effectiveness of the neural network and to explore the integration of supplementary texture and radiomic features. This approach is aimed at improving the accuracy of recognition by the diagnostic system of malignant neoplasms and healthy tissues.

Ways to improve the diagnostic model

Further studies are planned to improve the quality and effectiveness of the diagnostic model, including:

• Expanding the training dataset by adding medical images with different types and stages of tumors, as well as images with large variations in anatomical structures;
• The introduction of additional features (textural, radiomic, morphological) that can improve differentiation between tumor and healthy pancreatic tissues;
• The use of more complex neural network architectures [for example, TS-Total Segmentator, Abdomen Atlas U-Net (AAUNet), Abdomen Atlas Swin transformers U-Net (AASwin)] or ensembling methods, which will increase resilience to differences in images (20);
• Multimodal analysis: combining CT data with data from other imaging methods (MRI, ultrasound and X-ray images), utilize 3D models, as well as clinical and biochemical parameters for more accurate diagnosis (21).

Practical applicability of the model

For practical applicability, the algorithms output is presented as a monochromatic image, with the region containing the presumed cancerous tumor highlighted in red (Figure 5). The choice of color to indicate the segmented classes is arbitrary. In future developments, alternative color schemes or multimodal visualizations can be used to increase information content and improve visual interpretation.

Figure 5 In the CT image of the abdominal organs with contrast enhancement (venous phase), the neural network algorithm identifies the area with a pancreatic tumor (shown in red). The normal structure of the pancreas is shown in green. CT, computed tomography.

Figure 5 shows the segmentation performed by a modified neural network, which makes it possible to accurately determine the area of the pancreatic tumor (red) and normal tissue (green). It is important to note that for practical use, it is necessary not only to determine the area of interest, but also to measure its size. In the process of adapting the final algorithm for medical practice, a technical module was introduced to indicate the tumor size in centimetres, which is important for assessing the stage of the disease and planning treatment. In the future, it is possible to introduce tools for automatic assessment of tumor volume and shape (22). Additionally, it is possible to introduce a feedback mechanism from medical professionals to improve the algorithm based on clinical observations and results, which will contribute to its further development and optimization.

Prospects for development and integration

Further development will involve expanding the functionality of the diagnostic model. This includes adding analysis of other tumor types and organs, integrating with electronic medical records, automating report generation, and implementing automatic calibration and real-time training as new data becomes available (23). The introduction of the diagnostic system into clinical practice will significantly speed up the diagnostic process, reduce subjective errors and improve the quality of diagnosis of malignant pancreatic tumors.

Conclusions

• The findings demonstrate the potential of neural network algorithms in analysing CT images for detecting pancreatic cancer. The proposed algorithm exhibited a segmentation accuracy of 70% according to the Dice metric. In addition, after further training, the algorithm achieved an impressive accuracy of 88% in the classification of pancreatic cancer cases, demonstrating sensitivity of 98% and specificity of 98%. Notably, subjecting the algorithm to iterative retraining using an expanded dataset significantly elevates the reliability of pancreatic cancer diagnosis. This underscores the algorithm’s capacity for progressive refinement to augment its diagnostic performance.
• As part of the work, a functional database of medical images was developed, tested and enhanced. This medical image database is tailored to record, store, process, and transmit data in PNG format. PostgreSQL was chosen as the database system, due to its reliability, high scalability, and compliance with the requirements for storing and processing medical data. The future roadmap involves expanding the database and integrating it into a single software product alongside the algorithm for detecting pancreatic cancer and inflammatory diseases, complemented by a client interface. This will create a single software package to support medical decisions.
• The application of neural network algorithms based on texture features holds significant potential in the realm of clinical decision support systems. Employing diagnostic methods based on computer neural networks will contribute to their inclusion in clinical decision-making protocols. The incorporation of textural signatures for training neural networks will enhance the accuracy of diagnosing malignant pancreatic neoplasms and play a substantial role in determining patient management strategies and monitoring the effectiveness of prescribed treatments. We will consider applying the proposed methods to improve the diagnostic model in future research.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://cco.amegroups.com/article/view/10.21037/cco-25-17/rc

Data Sharing Statement: Available at https://cco.amegroups.com/article/view/10.21037/cco-25-17/dss

Peer Review File: Available at https://cco.amegroups.com/article/view/10.21037/cco-25-17/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://cco.amegroups.com/article/view/10.21037/cco-25-17/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Regional Clinical Hospital of the Kaliningrad (protocol No. 7-23 dated 05.10.2023).Informed consents (consent to participate and consent to publish) were obtained from all participants (all participants are over 18 years old).

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Gupta N, Yelamanchi R. Pancreatic adenocarcinoma: A review of recent paradigms and advances in epidemiology, clinical diagnosis and management. World J Gastroenterol 2021;27:3158-81. [Crossref] [PubMed]
2. Pekarek L, Fraile-Martinez O, Garcia-Montero C, et al. Towards an updated view on the clinical management of pancreatic adenocarcinoma: Current and future perspectives. Oncol Lett 2021;22:809. [Crossref] [PubMed]
3. Kaczor-Urbanowicz KE, Cheng J, King JC, et al. Reviews on Current Liquid Biopsy for Detection and Management of Pancreatic Cancers. Pancreas 2020;49:1141-52. [Crossref] [PubMed]
4. Rizzo S, Botta F, Raimondi S, et al. Radiomics: the facts and the challenges of image analysis. Eur Radiol Exp 2018;2:36. [Crossref] [PubMed]
5. Avanzo M, Stancanello J, Pirrone G, et al. Radiomics and deep learning in lung cancer. Strahlenther Onkol 2020;196:879-87. [Crossref] [PubMed]
6. Paramzin FN, Kakotkin VV, Burkin DA, et al. Radiomics and artificial intelligence in the differential diagnosis of tumor and non-tumor diseases of the pancreas (review). Surgical practice (Russia) 2023;8:53-65.
7. Heckler M, Hackert T. Surgery for locally advanced pancreatic ductal adenocarcinoma-is it only about the vessels. J Gastrointest Oncol 2021;12:2503-11. [Crossref] [PubMed]
8. Bartoli M, Barat M, Dohan A, et al. CT and MRI of pancreatic tumors: an update in the era of radiomics. Jpn J Radiol 2020;38:1111-24. [Crossref] [PubMed]
9. Schlick K, Kiem D, Greil R. Recent Advances in Pancreatic Cancer: Novel Prognostic Biomarkers and Targeted Therapy-A Review of the Literature. Biomolecules 2021;11:1469. [Crossref] [PubMed]
10. Vaiyapuri T, Dutta AK, Punithavathi ISH, et al. Intelligent Deep-Learning-Enabled Decision-Making Medical System for Pancreatic Tumor Classification on CT Images. Healthcare (Basel) 2022;10:677. [Crossref] [PubMed]
11. Gandikota HP. S A, M SK. CT scan pancreatic cancer segmentation and classification using deep learning and the tunicate swarm algorithm. PLoS One 2023;18:e0292785. [Crossref] [PubMed]
12. Bagheri MH, Roth H, Kovacs W, et al. Technical and Clinical Factors Affecting Success Rate of a Deep Learning Method for Pancreas Segmentation on CT. Acad Radiol 2020;27:689-95. [Crossref] [PubMed]
13. Khdhir R, Belghith A, Othmen S. Pancreatic Cancer Segmentation and Classification in CT Imaging using Antlion Optimization and Deep Learning Mechanism. International Journal of Advanced Computer Science and Applications 2023; [Crossref]
14. Nishio M, Noguchi S, Fujimoto K. Automatic pancreas segmentation using coarse-scaled 2d model of deep learning: usefulness of data augmentation and deep U-net. Appl Sci 2020;10:3360.
15. Yang M, Zhang Y, Chen H, et al. AX-Unet: A Deep Learning Framework for Image Segmentation to Assist Pancreatic Tumor Diagnosis. Front Oncol 2022;12:894970. [Crossref] [PubMed]
16. Zhao G, Chen X, Zhu M, et al. Exploring the application and future outlook of Artificial intelligence in pancreatic cancer. Front Oncol 2024;14:1345810. [Crossref] [PubMed]
17. Medical Segmentation Decathlon. [Accessed: 15 February 2023]. Available online: http://medicaldecathlon.com
18. Ghorpade H, Kolhar S, Jagtap J, et al. An optimized two stage U-Net approach for segmentation of pancreas and pancreatic tumor. MethodsX 2024;13:102995. [Crossref] [PubMed]
19. Yang E, Kim JH, Min JH, et al. nnU-Net-Based Pancreas Segmentation and Volume Measurement on CT Imaging in Patients with Pancreatic Cancer. Acad Radiol 2024;31:2784-94. [Crossref] [PubMed]
20. Suri A, Mukherjee P, Pickhardt PJ, et al. A Comparison of CT-Based Pancreatic Segmentation Deep Learning Models. Acad Radiol 2024;31:4538-47. [Crossref] [PubMed]
21. Zhang C, Achuthan A, Himel GMS. State-of-the-art and challenges in pancreatic CT segmentation: a systematic review of U-Net and its variants. IEEE Access 2024;12:78726-42.
22. Nadeem A, Ashraf R, Mahmood T, et al. Automated CAD system for early detection and classification of pancreatic cancer using deep learning model. PLoS One 2025;20:e0307900. [Crossref] [PubMed]
23. Podină N, Gheorghe EC, Constantin A, et al. Artificial Intelligence in Pancreatic Imaging: A Systematic Review. United European Gastroenterol J 2025;13:55-77. [Crossref] [PubMed]