A method for detecting at least one lesion in a patient's pancreas in at least one medical image.

Abstract

The present invention relates to a method implemented by computer means for detecting at least one lesion in a patient's pancreas in at least one medical image, for example, a portal vein computed tomography (CT) scan. A segmentation network is trained in a five-fold cross-validation process. The output of this network is then post-processed to extract imaging features such as normalized lesion risk, predicted lesion diameter, and MPD diameter in the pancreatic head, body, and tail. A logistic regression model is calibrated to predict the presence of a lesion based on these features.

Description

Technical Field

[0001] The present invention relates to a method implemented by computer means for detecting at least one lesion of a patient's pancreas in at least one medical image.

Background Art

[0002] Pancreatic cancer is currently the 11th most common cancer in the world and the 7th leading cause of cancer-related death. It is predicted that the incidence rate will increase by up to 78% during the period from 2018 to 2040, which is a major and growing healthcare problem. Due to the increasing incidence rate of pancreatic cancer, combined with its extremely low 5-year survival rate of 9%, this disease may become the 3rd leading cause of cancer-related death by 2025.

[0003] Most patients with early-stage pancreatic cancer present non-specific symptoms and thus, when examined, often undergo a normal computed tomography (CT) examination performed in the portal venous phase. In the initial stage, pancreatic lesions tend to be small (less than 2 cm) and isodense, and the reported sensitivity ranges from 58% to 77%, so image diagnosis can be difficult. In addition, the heavy workload of radiologists as well as specialized skills and experience can further affect the interpretation of CT scans. Since this disease progresses rapidly, most pancreatic cancers are detected during the terminal stage when there are limited treatment options available, thus explaining the observed low 5-year survival rate. To date, only 10% of patients have undergone pancreatectomy, the only radical treatment.

[0004] However, the proportion of patients diagnosed with stage 1A pancreatic cancer has been increasing in recent years. Generally, the description of stage 1A pancreatic cancer includes the fact that the cancer is confined to the pancreas, is less than 2 cm (0.8 inches) in size, and has not metastasized to nearby lymph nodes or distant sites. In this stage 1A, patients are more frequently eligible for pancreatectomy and adjuvant chemotherapy, and as a result, the 5-year survival rate for these patients is now over 80%, which highlights the importance of detecting pancreatic cancer as early as possible. To identify findings that can alert radiologists to the potential presence of pancreatic cancer, studies have retrospectively analyzed CT scans of pancreatic cancer patients prior to histopathological diagnosis. There has been agreement that subtle secondary signs, such as main pancreatic duct (MPD) dilation, were often observed up to one year before a pancreatic cancer diagnosis. This is due to the fact that pancreatic cancer is primarily a ductal adenocarcinoma, and the malignant tumor causes stenosis of the MPD, and therefore causes upstream dilation. Dilation is typically defined as a conduit larger than 3 mm in the head of the pancreas and larger than 2 mm in the body and tail of the pancreas. In addition, dilation upstream of stenosis, reported as local disappearance of the lumen of MPD, is also a pathological pattern.

[0005] Given this situation, deep learning (DL) methods have the potential to play a crucial role in supporting the daily practice of radiologists by alerting patients at risk of developing pancreatic cancer. Promising results have been obtained for some pathologies, such as breast cancer, where DL models significantly reduced false positive and false negative rates in two large datasets, while simultaneously significantly reducing the workload of radiologists. This effort, along with many attempts to use DL for lesion detection, is also being directed towards pancreatic cancer. These studies have proposed DL models for detecting pancreatic tumors and validated them in independent patient databases. While they showed promising results, these methods relied purely on DL and did not address the need to identify radiological findings that predict pancreatic cancer, which could potentially improve the detection of early lesions. [Preliminary Technology Documents] [Non-licensed literature]

[0006] [License 1] Galloway, Mary M (1975). "Texture analysis using gray level run lengths". Computer Graphics and Image Processing. 4(2): pp. 172~179. doi:10.1016 / S0146-664X(75)80008-6. [License 2] Pentland AP (June 1984). "Fractal-based description of natural scenes". IEEE Transactions on Pattern Analysis and Machine Intelligence. 6(6): pages 661~74. doi:10.1109 / TPAMI.1984.4767591.PMID 22499648.S2CID 17415943. [License 3] Amadasun M, King R (1989). "Textural features corresponding to textural properties". IEEE Transactions on Systems, Man, and Cybernetics. 19(5): pages 1264~1274. doi:10.1109 / 21.44046. [License 4] Thibault G, Angulo J, Meyer F (March 2014). "Advanced statistical matrices for texture characterization: application to cell classification". IEEE Transactions on Bio-Medical Engineering. 61(3): pp. 630-7. doi:10.1109 / TBME.2013.2284600.PMID24108747.S2CID11319154. [Overview of the project] [Means for solving the problem]

[0007] This paper proposes and evaluates a method that can predict patients at risk of pancreatic cancer.

[0008] For this purpose, this document provides a method implemented by computer means for detecting at least one lesion in a patient's pancreas in at least one medical image, the method being: (a) The step of predicting at least one first probability map representing the probability that each pixel or voxel in the image is part of a pancreatic lesion, (b) The step of predicting at least one second probability map representing the probability that each pixel or voxel in the image is part of the pancreas, (c) A step of predicting at least one third probability map representing the probability that each pixel or voxel in the image is part of the main pancreatic duct, (d) A step of segmenting the pancreas, main pancreatic duct, and at least one lesion in the image based on the probability map, (e) The step of determining the pancreatic head, pancreatic body, and pancreatic tail of the segmented pancreas, (f) A step of determining a first feature representing the risk of lesions based on the first probability map, (g) A step to determine a second characteristic representing the size of the lesion, (h) A step of determining a third feature that represents the maximum size of the main pancreatic duct in the head of the pancreas, (i) A step of determining a fourth feature that represents the maximum size of the main pancreatic duct in the body of the pancreas, (j) A step to determine a fifth feature that represents the maximum size of the main pancreatic duct in the tail of the pancreas, (k) A step of determining a first score based on at least the first, second, and one of the third, fourth, and fifth features, wherein the first score represents the probability that the patient has a pancreatic lesion. Includes an inference phase.

[0009] Medical images can be defined as visual representations of internal structures or functions acquired using various imaging techniques. These images are generated through the use of hardware and software systems that capture and process data to produce two-dimensional (2D) or three-dimensional (3D) images useful for diagnosing, planning, and monitoring various medical conditions. Medical images typically contain information about tissue density, composition, and function and are interpreted by radiologists, physicians, or other healthcare professionals who are specially trained in image analysis.

[0010] The aforementioned image may be a 3D image.

[0011] Medical images can take many different forms depending on the imaging technique used and the type of medical condition being investigated. However, according to this document, the image may be a CT scan image, for example, a portal vein CT scan image.

[0012] CT scans, also known as computed tomography, are 3D images created by combining a series of X-ray images taken from different angles around the body. CT scans provide detailed information about the internal structure and composition of organs, bones, and tissues.

[0013] The main difference between portal vein CT scan images and CT scan images lies in the focus of the imaging. CT scan images are a type of medical imaging that uses a computed tomography (CT) scanner to produce detailed images of any part of the body. CT scans can be used to evaluate many different organs and tissues, such as the brain, chest, abdomen, and pelvis, and can be performed with or without contrast agents.

[0014] On the other hand, portal vein CT scans specifically focus on imaging the portal venous system, including the gastrointestinal tract, spleen, and the veins that drain blood from the pancreas to the liver. Portal vein CT scans are a type of CT scan specifically designed to image the portal venous system in the abdomen, using a dedicated protocol that optimizes imaging of the portal venous system.

[0015] Portal vein CT scans can help diagnose and monitor a variety of conditions affecting the liver and pancreas, including tumors, infections, and inflammation. Portal vein CT scans can also help assess the response to treatment and guide further management of the condition. The portal vein system is a vital component of blood supply to the liver and other abdominal organs, and portal vein CT scans can provide valuable information about its function and anatomy.

[0016] In machine learning, the inference phase refers to the stage where a trained model is used to make predictions about new or unknown data. During this phase, the model receives input data and generates outputs based on patterns learned from the training data.

[0017] In other words, the inference phase is the process of applying a trained model to real-world data in order to make predictions or decisions.

[0018] A probability map (also known as a probability density map or probability distribution map) can refer to a representation of the likelihood or probability of a particular event occurring at different locations or regions within an input space.

[0019] More specifically, in the context of image processing, a probability map can be a 2D grid, 3D grid, or heatmap that assigns probability values to each pixel or voxel in an image, indicating the likelihood that the pixel or voxel belongs to a particular class or category. For example, in substance detection, a probability map can be used to identify the location and extent of an object in an image by assigning higher probabilities to pixels or regions that are more likely to belong to the object of interest.

[0020] Probability maps are commonly used in machine learning to represent the uncertainty or variability associated with different outcomes, and can also be used to make predictions, classify inputs, or guide decision-making processes.

[0021] Segmentation may refer to the process of dividing or partitioning an input image or data into multiple segments or regions (sets of pixels or voxels) based on specific criteria such as color, texture, luminance, shape, or other features. More precisely, image segmentation is the process of assigning labels, classes, or categories to all pixels or voxels in an image such that pixels with the same label share specific characteristics. The purpose of segmentation can be to identify and separate different objects or regions of interest within an image or data so that they can be analyzed, processed, or classified individually.

[0022] Several algorithms and techniques can be used for segmentation, including thresholding, clustering, edge detection, region growing, and deep learning-based methods.

[0023] In digital image processing and computer vision, features are information about the content of an image, typically information about whether a particular region of an image has a particular characteristic. Features can be specific structures within an image, such as points, edges, or objects. Features can also be the result of a general nearest neighbor operation or feature detection applied to an image. Other examples of features relate to motion within an image sequence, or shapes defined with respect to curves or boundaries between different image regions.

[0024] In a broader sense, a feature is any information relevant to solving a computational task specific to a particular application. This is the same meaning as a feature in machine learning and pattern recognition in general, although image processing has a much more sophisticated collection of features.

[0025] The pancreas is a glandular organ located in the abdomen that plays a vital role in the digestive system and in regulating blood glucose levels. The head of the pancreas is the widest and rightmost part of the pancreas, located next to the duodenum (the first part of the small intestine) and the bile duct. The tail of the pancreas is the narrower left end of the pancreas, extending toward the spleen. The body of the pancreas is the central part of the pancreas, connecting the head and tail, and is located behind the stomach.

[0026] Because different diseases and conditions can affect specific parts of an organ, it can be important to identify these three parts of the pancreas for diagnostic and therapeutic purposes.

[0027] Pancreatic lesions may refer to any abnormal growth, mass, or area of ​​tissue within the pancreas that differs in appearance or texture from the surrounding tissue. This can include cysts, tumors, or other abnormal tissue growths and may be benign or malignant. Pancreatic lesions can be detected through medical imaging tests such as ultrasound, CT scans, or MRI, and may require further evaluation via biopsy or other diagnostic procedures to determine their nature and potential impact on the patient's health. Management and treatment of pancreatic lesions depend on factors such as the size, location, and type of the lesion, as well as the individual's overall health and medical history.

[0028] The first score described above can quantify the likelihood that the detected area within the pancreas is a lesion (tumor, cyst, or other abnormality), whether malignant or benign. The purpose of the first score is to provide clinicians with an objective tool to help guide management decisions, such as recommending further diagnostic tests or proceeding with treatment.

[0029] The first score may be displayed directly within the interface, or it may be displayed in an interpreted manner that notifies the user of the presence or absence of a lesion based on a threshold. For example, if the first score is greater than the threshold, the interface may display a message informing the user of the presence or risk of a pancreatic lesion. In other words, a classification (e.g., lesion / non-lesion) may be derived from the first score.

[0030] The first score may be determined based on only one of the first, second, and third, fourth, and fifth features.

[0031] The first score may be based on the maximum values ​​of at least the first and second features, as well as the third, fourth, and fifth features. In such cases, it should be noted that the maximum values ​​of the third, fourth, and fifth features are based on the third, fourth, and fifth features, since the third, fourth, and fifth features must be calculated to determine their maximum values.

[0032] Several types of pancreatic lesions can occur, some of which are benign (non-cancerous) and some of which are malignant (cancerous). The following are some examples of potential pancreatic lesions. Pancreatic cysts: These are fluid-filled cysts that can occur in the pancreas. Most pancreatic cysts are benign, but some are precancerous or cancerous. Pancreatic pseudocysts: These are fluid and debris accumulations that can form after the onset of acute pancreatitis. Pseudocysts are usually benign, but they can become infected or rupture. Serous cystadenoma: These are benign pancreatic tumors filled with a clear, watery fluid. Mucinous cystic neoplasms: These are precancerous pancreatic tumors filled with thick, sticky mucus. Intraductal papillary mucinous neoplasms: These are precancerous pancreatic tumors that grow within the pancreatic duct and produce mucus. Solid papillary tumors: These are rare pancreatic tumors that are usually benign but have the potential to become malignant. Pancreatic endocrine tumors: These are rare tumors that arise from hormone-producing cells in the pancreas. Most pancreatic endocrine tumors are non-cancerous, but some may be malignant. Pancreatic adenocarcinoma: This is the most common type of pancreatic cancer and usually originates from the cells that line the inside of the pancreatic duct. Pancreatic lymphoma: This is a rare type of pancreatic cancer that originates from the lymphatic tissue of the pancreas. Metastatic pancreatic tumors: These are tumors that originate in other parts of the body, such as the lungs, breasts, or colon, and have metastasized to the pancreas.

[0033] The first feature representing the risk of lesions based on the first probability map can be a number between 0 and 1.

[0034] To calculate this first feature, candidate pixels or voxels outside the segmented pancreas (e.g., all pixels or voxels associated with a probability higher than 0) can be excluded. Then, for each remaining connected component, the lesion risk that falls between 0 and 1 can be calculated by averaging the probabilities of all its pixels or voxels.

[0035] Linked components with a lesion risk of less than 0.05 can be automatically removed.

[0036] In other words, step (f) is the following substep, namely, A step of removing all pixels or voxels that are not connected to the segmented pancreas, For each remaining connected component, the lesion risk is calculated by averaging the probabilities of all its pixels or voxels, and connected components with a lesion risk below a threshold are removed. It may include.

[0037] The threshold may be, for example, less than 0.05.

[0038] A connected component is a set of pixels that are connected to each other through neighbor relationships. A connected component can be a group of pixels that form a consistent object or region within an image.

[0039] A common method for identifying connected components in an image is to use connected component labeling algorithms. These algorithms analyze the relationships between pixels in an image and assign a unique label to each connected component. Once connected components are labeled, various actions can be performed on them, such as calculating their size, shape, or orientation, or applying filters to enhance or remove them.

[0040] A second feature representing the size of the lesion may be the maximum diameter of the corresponding segmented pancreatic lesion. To calculate the maximum diameter of the lesion, if the lesion is segmented, its diameter can be measured along the axial view for each slice in the case of a 3D image, for example, using the skimage library. The 2D Ferret diameter for each slice can be calculated. If the lesion is not segmented, the diameter can be automatically set to 0.

[0041] A second characteristic indicating the size of the lesion can also be determined by the following: Extracting the contour of a lesion using edge detection algorithms or morphological calculations. The contour is the boundary of the lesion and can be used to measure its size. Measuring the maximum diameter of a lesion by finding the longest distance between two points on the lesion's contour. This can be done using distance transformation, skeletonization, or other geometric measurements.

[0042] The first, second, and third probability maps can be obtained using at least one convolutional neural network or model, for example, a fully convolutional neural network.

[0043] The aforementioned convolutional neural network could be nnUNet.

[0044] nnUNet automatically designs a segmentation pipeline based on the UNet architecture by relying on heuristics applied to the data, which enable the estimation of key parameters. Dataset characteristics are estimated to automatically perform preprocessing steps. This is followed by the automatic definition of model design choices (number of layers, size of convolutional kernels, convolutional blocks, etc.). Training procedures (data augmentation, scheduled learning rate, etc.) are also performed.

[0045] The first, second, and third probability maps may be obtained using multiple models or convolutional neural networks, or an ensemble of models or convolutional neural networks, each model capable of predicting the first, second, and third probability maps.

[0046] The aforementioned ensemble of models (e.g., an ensemble of k models) can be trained using a k-fold cross-validation process. The output of each model (probability maps and segmentations for the pancreas, pancreatic lesions, and main pancreatic duct) is then averaged pixel by pixel or voxel by voxel to generate a single final probability map. The final segmentation is obtained by assigning each pixel or voxel to the most likely class (i.e., the class with the highest probability). The lesion probability map is obtained by examining the probability that each pixel or voxel is a lesion.

[0047] More specifically, k-fold cross-validation is a well-known technique used in machine learning to evaluate the performance of a model and prevent overfitting. In k-fold cross-validation, the original dataset is divided into k subsets or folds of approximately equal size. The model is then trained on k-1 of these folds and validated on the remaining folds.

[0048] The k-fold cross-validation process is repeated k times, with each fold being used once as part of the validation set. The model's performance is then averaged across all k iterations to obtain a more reliable estimate of its performance.

[0049] During training, image preprocessing, i.e., resampling and intensity normalization, can be entirely determined and performed by nnUNet.

[0050] Identification of the pancreatic head, body, and tail of a segmented pancreas can be obtained by performing the following steps. Along the axial view, extract the morphological skeleton of the pancreatic segmentation slice by slice, for example, using the skeletonize function from the skimage (or scikit-image) library. Obtain a 3D skeleton of the pancreas, and then, for example, use a network library to convert the obtained 3D pancreatic skeleton into a graph. Consider the point in the image located at the lower right anterior part of the abdomen, and identify the point on the graph furthest from it. Therefore, this point can be considered to be the tail end of the pancreas. The head of the pancreas is identified by examining the point in the graph that is furthest from the tail of the pancreas. Once the pancreatic head and tail are identified, the shortest path between them is calculated (for example, using Dijkstra's algorithm) to ultimately obtain a centerline that passes through the pancreas and connects its two ends. The midline is divided into three parts: the head of the pancreas, the body of the pancreas, and the tail of the pancreas. The first 25% can be considered the tail of the pancreas, the next 50% may be the body of the pancreas, and the last 25% may be the head of the pancreas. For each voxel in the pancreatic segmentation, find the nearest point on the midline and assign the corresponding location or portion of the pancreatic head, body, and tail.

[0051] When the pancreas is subsegmented, the main pancreatic duct may also be subsegmented. For each voxel of the segmented main pancreatic duct, its nearest point on the midline can be calculated. The positions of the pancreatic head, body, and tail of the segmented pancreas can then be assigned to the voxels according to the position of their nearest points on the midline.

[0052] Once this is done, the diameter of the main pancreatic duct can be measured in each part (pancreatic head, pancreatic body, and pancreatic tail). Along the axial view, for each slice, the diameter of the main pancreatic duct can be calculated, for example, using the IMEA library. In addition, it is possible to check whether the diameter was calculated in the pancreatic head, pancreatic body, or pancreatic tail.

[0053] Based on this, many indicators such as the minimum, maximum, mean, median, and percentile of the main pancreatic duct diameter in the pancreatic head, body, and tail can be extracted. To calculate the minimum, maximum, mean, median, and percentile for the entire main pancreatic duct, the indicators from the three parts can be aggregated.

[0054] Another advantage of subsegmenting the pancreas into the tail, body, and head is that the location of the lesion can also be extracted by calculating the distance between the lesion voxel and the midline. Once this is done, the portion where the majority of the lesion voxels are located (the head, body, and tail) can be determined and assigned as the lesion location.

[0055] Overall, subsegmentation of the pancreas can be used to identify the location of other anatomical structures within the pancreas and to extract local features.

[0056] The calculations can be performed using the IMEA library.

[0057] The maximum diameter of the main pancreatic duct in the pancreas can also be calculated by taking the maximum diameter between the head, body, and tail of the pancreas. For each part, if there is no segmentation, the main pancreatic duct diameter is set to 0.

[0058] The determination of the first score can be performed by a first logistic regression model.

[0059] The method described herein may also include a step of determining a second score based on at least the third, fourth, and fifth features, the second score representing the probability that the main pancreatic duct is dilated.

[0060] The second score may be displayed directly within the interface, or it may be displayed in an interpreted manner that informs the user of the dilated or non-dilated state of the main pancreatic duct based on a threshold. For example, if the second score is greater than the threshold, the interface may display a message informing the user of the dilated state of the main pancreatic duct. In other words, a classification (e.g., dilated / non-dilated) may be derived from the second score.

[0061] The second score may be determined based solely on the third, fourth, and fifth features.

[0062] The determination of the second score described above can be performed by a second logistic regression model.

[0063] The first and / or second logistic regression models can be trained using the Scikit-Learn library with default hyperparameters.

[0064] In addition to the features described above, the first and / or second scores may also be determined based on at least one other feature from the following list. Primary statistics of pancreatic diameter, lesion diameter, main pancreatic duct diameter, and / or common bile duct diameter. These features can be regionized for each part of the pancreas (pancreatic head, pancreatic body, pancreatic tail) using pancreatic subsegmentation, thus creating new features. The location of the lesion, for example, the location of the lesion in the head, body, and / or tail of the pancreas. Radiomic features of the pancreas, lesions, main pancreatic duct, and / or common bile duct. These features can be regionized into each part of the pancreas (pancreatic head, pancreatic body, pancreatic tail) using subsegmentation of the pancreas, thus creating new features. 3D radiomic features of the pancreas, lesions, main pancreatic duct, and / or common bile duct. These features can be regionized into individual parts of the pancreas (pancreatic head, pancreatic body, pancreatic tail) using subsegmentation of the pancreas. Two-dimensional radiomic features of the pancreas, main pancreatic duct, lesions, and / or common bile duct, and / or primary statistics of these features. These features can be regionized for each part of the pancreas (pancreatic head, pancreatic body, pancreatic tail) using subsegmentation of the pancreas.

[0065] In the medical field, radiomics is a method of extracting numerous features from medical images using data feature identification algorithms. These features, called radiomic features, have the potential to reveal patterns and characteristics on images that are not perceptible to the naked eye.

[0066] Primary statistics, also known as descriptive statistics, are summary metrics that describe the basic characteristics of a dataset without making assumptions about the underlying distribution. These statistics include indicators of central tendency, such as the mean, median, and mode, providing information about typical values ​​or the mean of the data. They also include indicators of variance or variability, such as range, variance, and standard deviation, providing information about how spread the data is. Other primary statistics include indicators of skewness and kurtosis, which describe the shape of the distribution. These statistics are widely used in data analysis and can help provide insights into the characteristics of a dataset.

[0067] Radiomic features are the group, namely, Primary features 3D shape features 2D shape features Gray-Level Co-occurrence Matrix (GLCM) Characteristics Gray Level Size Zone Matrix (GLSZM) Features Gray Level Run Length Matrix (GLRLM) Characteristics Features of the Neighboring Gray Tone Difference Matrix (NGTDM) Gray Level Dependence Matrix (GLDM) Characteristics It can be divided into:

[0068] More specifically, The primary feature describes the distribution of voxel intensity within the image region defined by the mask through commonly used basic metrics. The primary feature may include at least one of the following features: energy, total energy, entropy, minimum, 10th percentile, 90th percentile, maximum, mean, median, interquartile range, range, mean absolute deviation (MAD), robust mean absolute deviation (rMAD), root mean squared (RMS), standard deviation, skewness, kurtosis, variance, and uniformity. 3D shape features are a description of the three-dimensional size and shape of the ROI. These features are independent of the gray level intensity distribution in the ROI and are therefore calculated only for underived images and masks. The 3D shape features may include at least one of the following features: mesh volume, voxel volume, surface area, surface area-to-volume ratio, sphericity, compactness 1, compactness 2, spherical disproportion, maximum 3D diameter, maximum 2D diameter (slice, row, or column), major axis length, minor axis length, minimum axis length, elongation, and flatness. 2D shape features are a two-dimensional description of the size and shape of the ROI. These features are independent of the gray level intensity distribution in the ROI and are therefore calculated only for underived images and masks. The 2D shape features may include at least one of the following features: mesh surface area, pixel surface area, perimeter, perimeter-to-surface area ratio, sphericity, spherical disproportion, maximum 2D diameter, major axis length, minor axis length, and elongation.

[0069] All of these radiomic features are defined at the URL https: / / pyradiomics.readthedocs.io / en / latest / features.html# and can be implemented through the use of the pyradiomics library.

[0070] The mathematical definitions of these features are independent of the imaging modality and can be found in the literature, i.e., Galloway, Mary M (1975). "Texture analysis using gray level run lengths". Computer Graphics and Image Processing. 4(2): pp. 172-179. doi:10.1016 / S0146-664X(75)80008-6. Pentland AP (June 1984). "Fractal-based description of natural scenes". IEEE Transactions on Pattern Analysis and Machine Intelligence. 6(6): pp. 661-74. doi:10.1109 / TPAMI.1984.4767591. PMID 22499648. S2CID 17415943. Amadasun M, King R (1989). "Textural features corresponding to textural properties". IEEE Transactions on Systems, Man, and Cybernetics. 19(5): pp. 1264-1274. doi: 10.1109 / 21.44046. Thibault G, Angulo J, Meyer F (March 2014). "Advanced statistical matrices for texture characterization: application to cell classification". IEEE Transactions on Bio-Medical Engineering. 61(3): pp. 630-7. doi:10.1109 / TBME.2013.2284600.PMID24108747.S2CID11319154. It can also be seen there.

[0071] The common bile duct, like the main pancreatic duct, can be determined through the use of probability maps and / or segmentation.

[0072] This document also proposes computer software that includes instructions for implementing at least some of the methods described herein when the software is executed by a processor.

[0073] This document is, An input interface for receiving medical images, A memory for storing at least the instructions of a computer program according to the above claims, A processor that accesses memory to read the aforementioned instructions and then executes the method described in this document, An output interface for providing instructions based on a first score and / or a second score. We also propose a computer device equipped with this feature.

[0074] This document also proposes a computer-readable non-temporary recording medium on which computer software is registered to implement the methods described herein when the computer software is executed by a processor.

[0075] Other features, details, and benefits are shown in the detailed description and diagrams below. [Brief explanation of the drawing]

[0076] [Figure 1] This diagram schematically illustrates an example of a computer device as described in this document. [Figure 2] This figure shows the data used to construct the training and test sets. [Figure 3] This figure shows the training phase pipeline for the method described in this document. [Figure 4] This figure shows the pipeline for the inference phase of the method described in this document. [Figure 5] This figure shows three different subsegments of the pancreas. [Figure 6] This diagram shows an example of model segmentation, with the left image representing the input image and the right image showing the segmented elements on the image (pancreas in red, pancreatic lesions in green, and the main pancreatic duct in blue). [Figure 7] This diagram shows an example of model segmentation, with the left image representing the input image and the right image showing the segmented elements on the image (pancreas in red, pancreatic lesions in green, and the main pancreatic duct in blue). [Figure 8] This diagram shows an example of model segmentation, with the left image representing the input image and the right image showing the segmented elements on the image (pancreas in red, pancreatic lesions in green, and the main pancreatic duct in blue). [Figure 9] This diagram shows an example of model segmentation, with the left image representing the input image and the right image showing the segmented elements on the image (pancreas in red, pancreatic lesions in green, and the main pancreatic duct in blue). [Figure 10] This figure shows the model performance on the test set. [Figure 11] This figure shows the model performance on the test set. [Modes for carrying out the invention]

[0077] The attached drawings include meaningful colors. This application will be published in black and white, but colored versions of the attached drawings have already been submitted to the Patent Office.

[0078] Figure 1 schematically shows an example of a computer device 1 according to the present invention. The computer device 1 is Input Interface 2 At least memory 3 for storing computer program instructions, A processor 4 accesses memory 3 to read the aforementioned instructions and executes the method described in this document, Output interface 5 and It is equipped with.

[0079] dataset The data was collected from nine medical centers in Europe, the United States, and Brazil. The inclusion criteria were as follows: (i) Presence of portal vein CT scan (ii) Maximum slice thickness of 3 mm (iii) Patients with confirmed pancreatic tumors were studied before any treatment or surgery. This resulted in a total of 2890 cases, which were further divided into a training set of 2134 subjects and an independent test set of 756 subjects (see Figure 2).

[0080] The training set consisted of portal vein CT scans of 2,134 patients from five facilities, of which 1,692 had pancreatic tumors. Diagnosis was obtained from biopsy reports for 78% of the subjects, and through the C25 code associated with pancreatic tumors in the International Classification of Diseases (ICD-10) for the remaining subjects.

[0081] The training set also included 422 control subjects who, according to their radiology reports, showed no evidence of pancreatic lesions.

[0082] Table 1 below shows the characteristics of the patients.

[0083] [Table 1]

[0084] The 2134 subjects consisted of 1174 women (55%) and 960 men (45%), with a median age of 64 years (range [56,74] years). 1184 patients had pancreatic ductal adenocarcinoma (PDAC), 134 had neuroendocrine tumors (NETs), and 158 had unclassified solid lesions. Additionally, 81 subjects had intraductal papillary mucinous neoplasm (IPMN), 34 had mucinous cystic neoplasm (MCN), 42 had serous cystadenoma (SCA), and 59 had unclassified cystic lesions. Ultimately, 43% (907) of the subjects had dilated pancreatic dysplasia (MPD).

[0085] The independent test set included 756 subjects collected from both public and private data at four different sites (see Figure 2). These sites differed from those used in the training set. The public data included 361 portal vein CT scans, of which 281 were from patients with pancreatic lesions and 80 were from healthy patients. Cancer cases were obtained from the Medical Decathlon Challenge. Control subjects were accessed from the pancreatic CT dataset in the Cancer Imaging Archive. The private dataset consisted of 212 portal vein CT scans from patients with histopathological confirmation of pancreatic tumors, as well as 183 portal vein CT scans from a second medical institution without radiological evidence of pancreatic lesions. Demographic information was not available for subjects from the test set. This database consisted of patients with PDAC (n=360), NET (n=48), and unclassified solid lesions (n=12), as well as patients with IPMN (n=18) and unclassified cystic lesions (n=53) (see Table 1). Ultimately, 276 subjects had expanded MPD.

[0086] Annotation Protocol Each portal vein CT scan was reviewed and annotated by one of a group of nine radiologists. The annotator segmented the pancreas on each image. Where a lesion was identified by a radiologist, that lesion was systematically segmented. The radiologists characterized the tumor type based on the CT scan. Possible types were PDAC, NET, IPMN, MCN, and SCA. Cases where the type of lesion could not be determined were labeled as unclassified. Finally, the radiologists segmented the MPD where it was visible and visually assessed the MPD dilation.

[0087] Detection pipeline segmentation The 3D nnUNet was trained to segment the pancreas, lesions, and MPD in a 5-fold cross-validation process. Image preprocessing, i.e., resampling and intensity normalization, was entirely determined and performed by the nnUNet. When applied to new images, the trained nnUNet generated segmentation of the pancreas, pancreatic lesions, and MPD, as well as a lesion probability map assigning a probability that each voxel is a lesion (see Figure 3).

[0088] In the inference phase, five nnUNets are used to predict probability maps, each assigning a probability that each voxel is pancreatic, lesional, or MPD. These five lesional probability maps are then averaged per voxel to generate a single final probability map. Final segmentation is obtained by assigning each voxel to the most likely class (i.e., the class with the highest probability). The lesional probability map is obtained by examining the probability that each voxel is lesional.

[0089] Feature extraction Using the output of nnUNet, the features were extracted as follows: (i) Lesion risk between 0 and 1 calculated based on the 3D probability map provided by nnUNet. To do this, all extrapancreatic candidate lesions were excluded using a segmentation map. Then, for each connected component, the lesion risk between 0 and 1 was calculated by averaging the probabilities of all its voxels. Connected components with a lesion risk of less than 0.05 were automatically removed. (ii) Maximum diameter of a lesion segmented by nnUNet. Once a lesion is segmented, its diameter can be measured slice by slice along the axial view using the skimage library. The 2D Ferret diameter of each slice can be measured, and many values ​​such as minimum, maximum, mean, median, and percentile can be calculated. If the lesion is not segmented, the diameter is automatically set to 0. The common bile duct can also be segmented. Once the segmentation network is trained, the exact same method can be used to calculate the diameter of the common bile duct. (iii) Maximum MPD diameter in the head, body, and tail of the pancreas. To locally measure the MPD diameter, it is first necessary to identify the head, body, and tail of the pancreas. This is done by performing the following steps. 1) Segment the pancreas (using the nnUNet mentioned above). 2) Use the skeletonize function from skimage to extract the morphological skeleton of the pancreatic segmentation slice by slice along the axial view. 3) Obtain a 3D skeleton of the pancreas, which is converted into a graph using a network library. 4) Consider the point in the image located at the far right anterior lower part of the abdomen, and identify the point on the graph furthest from it. Therefore, this point can be considered to be the tail end of the pancreas. 5) Identify the head of the pancreas by examining the point in the graph that is furthest from the tail of the pancreas. 6) Once the pancreatic head and tail are identified, the shortest path between them is calculated using Dijkstra's algorithm to obtain a midline that passes through the pancreas and connects its two ends. 7) Divide the midline into three parts. The first 25% is considered the tail of the pancreas, the next 50% is the body of the pancreas, and the last 25% is the head of the pancreas. 8) For each voxel in the pancreatic segmentation, find the nearest point on the midline and assign its corresponding position.

[0090] Figure 5 illustrates how this method allows for the identification of the pancreas in three sub-parts. This figure shows the results of sub-segmentation (in the lower image) with three different segmented pancreases (in red in the upper image), as well as sub-segmented pancreatic head (green), pancreatic body (blue), and pancreatic tail (red).

[0091] When the pancreas is subsegmented, the MPD can also be subsegmented. For each voxel in the MPD segmentation, its nearest point on the midline is calculated. Then, a location (pancreatic head, pancreatic body, pancreatic tail) is assigned to the voxel according to the position of that nearest point on the midline.

[0092] Once this is done, the diameter of the pancreatic ductus (MPD) can be measured in each section. Along the axial view, the MPD diameter is calculated for each slice using the IMEA library. In addition, it is determined whether the diameter was calculated in the pancreatic head, body, or tail. Based on this, many metrics can be extracted, such as the minimum, maximum, mean, median, and percentile values ​​of the MPD diameter in the pancreatic head, body, and tail.

[0093] To calculate the minimum, maximum, mean, median, and percentile values ​​for the entire MPD, the metrics from three parts can be aggregated.

[0094] Another advantage of subsegmenting the pancreas is that the location of lesions can also be extracted by calculating the distance between the lesion voxel and the midline. Once this is done, it becomes possible to identify where the majority of lesion voxels are located (pancreatic head, pancreatic body, and pancreatic tail) and assign the result as the lesion location.

[0095] Overall, subsegmentation of the pancreas can be used to identify the location of other anatomical structures within the pancreas and to extract localized features.

[0096] The calculations were performed using the IMEA20 library. The maximum MPD diameter in the pancreas was also calculated by taking the maximum diameter between the pancreatic head, body, and tail. For each portion, if there was no segmentation, the MPD diameter was set to 0.

[0097] Logistic regression It is proposed to predict the presence of lesions and MPD expansion in patients using two logistic regression models that depend on previously defined features. To do so, nnUNet was applied to a validation set for each fold, resulting in a total of 2134 3D probability maps and segmentations from which features were extracted. Two logistic regression models with default hyperparameters were trained using the Scikit-Learn library.

[0098] The first logistic regression model has three characteristics, namely, 1) Risk of lesions 2) Lesion diameter 3) Maximum MPD diameter in the pancreas The presence of lesions was predicted based on this.

[0099] A second logistic regression model predicted MPD dilation using MPD diameters in the pancreatic head, body, and tail.

[0100] Once the two logistic regression models are trained, the evaluation of the test subjects is done in three steps, namely, 1) Using nnUNet, calculate the lesion probability map, as well as the segmentation of the pancreas, lesions, and MPD. 2) Feature extraction and, 3) To predict the presence of lesions and MPD expansion, two previously learned logistic regression models are applied (see Figure 4). It is composed of.

[0101] In this example, lesion risk, maximum lesion diameter, and maximum MPD diameter were used to predict the presence of a lesion. Regarding MPD, MPD diameters in the pancreatic head, body, and tail were considered.

[0102] In addition to the features described above, each logistic regression model may also use at least one other feature from the following list as an input feature. Primary statistics of pancreatic diameter, lesion diameter, main pancreatic duct diameter, and / or common bile duct diameter. These features can be regionized for each part of the pancreas (pancreatic head, pancreatic body, pancreatic tail) using pancreatic subsegmentation, and thus new features can be created. The location of the lesion, for example, the location of the lesion in the head, body, and / or tail of the pancreas. Radiomic features of the pancreas, lesions, main pancreatic duct, and / or common bile duct. These features can be regionized for each part of the pancreas (pancreatic head, pancreatic body, pancreatic tail) using subsegmentation of the pancreas, and thus new features can be created. 3D radiomic features of the pancreas, lesions, main pancreatic duct, and / or common bile duct. These features can be regionized into individual parts of the pancreas (pancreatic head, pancreatic body, pancreatic tail) using subsegmentation of the pancreas. Two-dimensional radiomic features of the pancreas, main pancreatic duct, lesions, and / or common bile duct, and / or primary statistics of these features. These features can be regionized for each part of the pancreas (pancreatic head, pancreatic body, pancreatic tail) using subsegmentation of the pancreas.

[0103] statistical analysis To evaluate performance, receiver operating characteristic (ROC) curves were created by plotting sensitivity at different thresholds against the false positive rate. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) at the operating point maximizing balanced accuracy were measured, as well as the area under the curve (AUC). Case-level evaluations were performed for both logistic regressions estimating the presence of lesions and logistic regressions predicting MPD dilation. For example, in lesion detection, sensitivity was defined as the ratio of the number of patients in whom the model correctly detected lesions to the total number of patients with pancreatic lesions. The calculation of other metrics was defined accordingly.

[0104] Bootstrap sampling was used to provide median and 95% confidence interval (CI) values ​​for AUC, sensitivity, specificity, PPV, and NPV.

[0105] Finally, segmentation performance was evaluated by calculating the Dice score and the Normalized Surface Dice (NSD) score between the reference segmentation and the predicted segmentation for each test patient. NSD is suitable for evaluating small structures such as MPDs because it allows for a tolerance between the reference segmentation and the predicted segmentation. In this study, the tolerance was set to 2 mm along each spatial dimension.

[0106] result Detecting patients with pancreatic tumors The model performance in the test set is reported in Figures 10 and 11, and in Table 2 below.

[0107] Figure 10 shows the ROC (Receiver Operating Characteristic) curve (AUC: Area Under Curve) of the logistic regression model predicting the presence of lesions. The central line represents the mean curve, and the shaded area represents the 95% confidence interval. Figure 11 shows the model confusion matrix obtained at the operating point that maximizes balance accuracy.

[0108] [Table 2]

[0109] The model achieved an AUC of 0.98 (95% CI: 0.97, 0.99), a sensitivity of 0.94 (469 out of 493, 95% CI: 0.92, 0.97), and a specificity of 0.95 (246 out of 262, 95% CI: 0.92, 0.98).

[0110] For comparison with other methods, model performance on publicly available data (defined in Figure 2) is also reported in Table 2. The evaluation metrics remained the same as those obtained for the entire test set.

[0111] The model was also evaluated in subjects with specific lesion characteristics and types. (i) Patients with lesions less than 2 cm in diameter, (ii) Patients with lesions of the same concentration, (iii) Patients with PDAC, (iv) Patients with NETs, (v) Patients with IPMN Five subgroups of the cohort were further investigated.

[0112] The performance obtained in these subsets is reported in Table 2. AUC, sensitivity, and specificity remained consistent across subgroups and were comparable to those obtained in the overall test set. The model performed best in the NET subgroup, with a sensitivity of 1.0 (95% CI: 0.98, 1.0). Specificity was slightly lower in smaller lesions compared to other subsets.

[0113] The importance of characteristics in lesion detection sensitivity The logistic regression model used to predict the presence of lesions in patients was based on three features: lesion risk, lesion diameter, and MPD diameter. An ablation study was conducted to evaluate the impact of combining these features on performance. Two additional logistic regression models were trained: the first model used both lesion risk and lesion diameter, while the second model used only lesion risk. Table 3 below reports the sensitivity of these three models in the test set and previously defined subsets.

[0114] [Table 3]

[0115] The use of MPD diameter and lesion diameter systematically improved lesion detection sensitivity across all groups. In the overall test set, adding MPD diameter and lesion diameter resulted in a 4% sensitivity improvement compared to the baseline model using lesion risk only. The impact of these two features was particularly strong in isodense lesions, showing a 10% sensitivity increase compared to models using only one feature.

[0116] MPD Extended Detection Performance Regarding the performance of the MPD extension, a logistic regression model was also evaluated on the test set. The results are reported in Table 4 below.

[0117] [Table 4]

[0118] An AUC of 0.97 (95% CI: 0.96, 0.98), a sensitivity of 0.94 (259 out of 276, 95% CI: 0.89, 0.97), and a specificity of 0.90 (432 out of 480, 95% CI: 0.86, 0.94) were achieved.

[0119] Segmentation performance The segmentation predicted by nnUNet, as described above, was evaluated both quantitatively and qualitatively. The Dice scores between ground truth and nnUNet segmentation for the pancreas, lesions, and MPD are reported in Table 5 below.

[0120] [Table 5]

[0121] NSD scores were calculated only for MPD. Dice scores and NSD scores were measured only in publicly available data to compare the segmentation model with other studies. Qualitative examples of the model segmentation are provided in Figures 6–9. In these figures, the left image shows the input image (corresponding axial portal vein CT slice of the patient - white arrows indicate the location of the lesion in Figures 6–8 and MPD in Figure 9), and the right image shows the segmented elements on the image (pancreas in red, pancreatic lesion in green, main pancreatic duct in blue, -PDAC: pancreatic ductal adenocarcinoma, -IPMN: intraductal papillary mucinous neoplasm, -NET: neuroendocrine tumor, -MPD: main pancreatic duct).

[0122] Across the entire test set, the mean Dice scores were 0.91 (±0.06), 0.69 (±0.34), and 0.58 (±0.37) for pancreas, lesions, and MPD, respectively. The NSD score for MPD was 0.71 (±0.39).

[0123] Consider This paper provides a method for automatically detecting patients with pancreatic lesions (e.g., tumors) and identifying cases with MPD dilation. The proposed method was validated in an independent cohort of 756 subjects. It is shown how using MPD dilation information can improve the sensitivity of lesion detection compared to a baseline method that relies solely on segmentation network output. Finally, the model's ability to accurately localize lesions was also evaluated by assessing its segmentation performance.

[0124] The models were evaluated in patient subgroups based on the characteristics and type of lesions. Similar AUC, sensitivity, and specificity were observed across different subgroups, thus highlighting the robustness of the proposed method. In some subgroups, such as NET (49 out of 312) and IPMN (18 out of 290), greater variability was observed for PPV and NPV, largely due to the imbalance between healthy and diseased subjects.

[0125] The results using publicly available data can be compared to those of competing deep learning models that used this database to test their methods. This method demonstrated significantly higher performance than conventional techniques, with an AUC of 0.99 (95% CI: 0.98, 0.99).

[0126] The improved performance may be largely attributable to the use of multiple specific features to predict the presence of lesions. While conventional methods tend to rely solely on segmentation generated by convolutional neural networks to predict lesions on images, the proposed method combines lesion risk, as well as MPD diameter and lesion diameter, to predict whether a subject has a lesion by logistic regression. To evaluate the effect of combining these features, the sensitivity of logistic regression was measured depending on the features used for training. Logistic regression models using only lesion risk, which are most similar to state-of-the-art methods, have lower sensitivity in all subsets compared to models using two or three features. In particular, using MPD diameter histologically improved sensitivity (+5%), especially for isotonic lesions. An improvement in sensitivity of PDAC detection (+5%) was also observed when using lesion and MPD diameter. A less significant effect was observed with respect to pancreatic NETs (+2% sensitivity). However, since NETs are not associated with MPD dilation, the impact on sensitivity in patients with NETs was not predicted. Ultimately, adding MPD diameter and lesion diameter significantly improved sensitivity in the case of IPMN (+6%), but the small number of cases (18 out of 290) prevented any meaningful results from being obtained.

[0127] The segmentation network was also used to design a logistic regression model that predicts MPD dilation based on the diameters of the pancreatic head, body, and tail. An AUC of 0.97 (95% CI: 0.96, 0.98) was reported. The inventors emphasize that while other deep learning methods may enable MPD segmentation, none of them utilize MPD segmentation to provide alerts for potential dilation, which is an important finding for radiologists when evaluating the pancreas.

[0128] The segmentation performance of the algorithms for the pancreas, lesions, and MPD was also evaluated. For the pancreas, no competing deep learning methods evaluated Dice scores on the same dataset as this study. However, three deep learning models were found that reported an average Dice score of 0.87 on the test set. The segmentation network presented in this paper obtained similar results on the publicly available data, achieving a higher average Dice score across the entire test set. For lesion segmentation, an average Dice score of 0.63 was obtained on the publicly available data. This represents a 9% improvement over the prior art using nnUNet, which reported a Dice score of 0.54 on this dataset. Finally, the inventors were unable to find any Dice scores obtained for MPD by other deep learning methods. However, considering the small size of MPD and compared to the Dice scores obtained for lesions, the model appeared to demonstrate satisfactory performance, confirmed by an NSD score of 0.71 on the test set. [Explanation of Symbols]

[0129] 1. Computer devices 2 Input Interfaces 3 memory 4 processors 5 Output Interfaces

Claims

1. A method implemented by computer means for detecting at least one lesion in a patient's pancreas in at least one medical image, (a) The step of predicting at least one first probability map representing the probability that each pixel or voxel in the image is part of a pancreatic lesion, (b) The step of predicting at least one second probability map representing the probability that each pixel or voxel in the image is part of the pancreas, (c) A step of predicting at least one third probability map representing the probability that each pixel or voxel in the image is part of the main pancreatic duct, (d) A step of segmenting the pancreas, the main pancreatic duct, and at least one lesion in the image based on the probability map, (e) The step of determining the pancreatic head, pancreatic body, and pancreatic tail of the segmented pancreas, (f) A step of determining a first feature representing the risk of lesions based on the first probability map, (g) A step of determining a second feature that represents the maximum size of the lesion, (h) A step of determining a third feature that represents the maximum size of the main pancreatic duct in the head of the pancreas, (i) A step of determining a fourth feature that represents the maximum size of the main pancreatic duct in the body of the pancreas, (j) A step of determining a fifth feature that represents the maximum size of the main pancreatic duct in the tail of the pancreas, A step of determining a first score based on at least the first feature, the second feature, and at least one of the third, fourth, and fifth features, wherein the first score represents the probability that the patient has a pancreatic lesion. A method that includes an inference phase.

2. The method according to claim 1, wherein the medical image is a CT scan image.

3. The method according to claim 2, wherein the image is a portal vein CT scan image.

4. Step (f) is a substep, namely, The steps include removing all pixels or voxels located outside the segmented pancreas, For each remaining connected component, the lesion risk is calculated by averaging the probabilities of all its pixels or voxels, and connected components with a lesion risk below a threshold are removed. The method according to any one of claims 1 to 3, including the method described in any one of claims 1 to 3.

5. The method according to any one of claims 1 to 4, wherein the second feature is the maximum diameter of the corresponding segmented pancreatic lesion.

6. The method according to claim 5, wherein the image is a 3D image, and the maximum diameter of the corresponding segmented pancreatic lesion is determined by measuring the 2D Ferret diameter of the segmented lesion slice by slice along the axial view.

7. The method according to any one of claims 1 to 6, wherein the first, second, and third probability maps are obtained using at least one convolutional neural network or model, e.g., nnUNet.

8. The method according to any one of claims 1 to 7, wherein the first, second, and third probability maps are obtained using a plurality or ensemble of convolutional neural networks or models, each model capable of predicting the first, second, and third probability maps, and the outputs of the models are averaged in pixel or voxel units to generate each of the probability maps.

9. The method according to claim 8, wherein the ensemble of models is trained using a k-fold cross-validation process.

10. The aforementioned image is a 3D image, and step (e) is the following substep, namely, The steps include extracting the morphological skeleton of the pancreatic segmentation slice by slice along the axial view, The steps include obtaining a 3D skeleton of the pancreas and converting the skeleton into a graph, A step of identifying the point on the graph that is furthest from the point in the image located at the rightmost anterior lower part of the abdomen, wherein the point is considered to be the tail end of the pancreas. The steps include identifying the pancreatic head as the point in the graph furthest from the pancreatic tail, The steps include: calculating the shortest path between the pancreatic head and the pancreatic tail in order to obtain a central line that passes through the pancreas and connects the pancreatic head end and the pancreatic tail end; The steps include dividing the aforementioned central line into the pancreatic head, pancreatic body, and pancreatic tail, For each voxel of the pancreatic segmentation, the step of finding the nearest point on the midline and assigning the corresponding pancreatic head, pancreatic body, and pancreatic tail; The method according to any one of claims 1 to 9, including

11. The method according to claim 10, wherein for each voxel of the segmented main pancreatic duct, the nearest point on the midline is calculated, the positions of the pancreatic head, pancreatic body, and pancreatic tail of the segmented pancreas are assigned to the voxel according to the position of the nearest point on the midline, and the diameter of the main pancreatic duct is measured in each of the pancreatic head, pancreatic body, and pancreatic tail.

12. The method according to any one of claims 1 to 11, wherein the first score is determined using a first logistic regression model.

13. The method according to any one of claims 1 to 12, wherein the method comprises the step of determining a second score based on at least the third, fourth, and fifth features, the second score representing the probability that the main pancreatic duct is dilated.

14. The method according to claim 13, wherein the second score is determined using a second logistic regression model.

15. The first score or the second score is from the following list, namely, Primary statistics of the diameter of the pancreas, the diameter of the lesion, the diameter of the main pancreatic duct, and / or the diameter of the common bile duct. Location of the lesion, for example, the location of the lesion in the head, body, and / or tail of the pancreas, The aforementioned pancreas, lesion, main pancreatic duct, and / or common bile duct, at least one radiomic feature, The 3D morphological features of the pancreas, lesion, main pancreatic duct, and / or common bile duct, 2D morphological features of the pancreas, main pancreatic duct, lesions, and / or common bile duct, and / or primary statistics of these features. The method according to any one of claims 1 to 14, determined based on at least one other feature from.

Images