AI assisted detection of vascular abnormalities in medical images

Abstract

The invention relates to AI-assisted detection of vascular abnormalities in medical images. The invention relates to a computer-implemented training data preparation method (100, 200), comprising: receiving (102, 202) an input medical image of a blood vessel of a patient; determining (104, 204) a vessel segmentation from the input medical image; identifying and annotating (110, 210) anatomical landmarks in the blood vessel segmentation to produce an annotated blood vessel segmentation; the annotated vessel segments are stored (112, 212) as training data. The invention also relates to a training method for training a neural network on the basis of training data, and to a medical diagnostic method applying the AI model thus trained.

Description

Technical Field

[0001] The present invention relates to a method and system for AI (artificial intelligence)-assisted vascular anomaly detection in medical imaging, and in particular to a method for training an AI model and applying the AI ​​model to medical imaging. Background Art

[0002] Neurovascular abnormalities—including but not limited to occlusions, stenosis, and aneurysms—present significant health risks, making timely and accurate detection of neurovascular abnormalities essential for effective therapeutic intervention and improved patient outcomes. Advanced medical imaging technologies such as computed tomography angiography (CTA) and magnetic resonance imaging (e.g., magnetic resonance angiography (MRA), time-of-flight magnetic resonance angiography (MRTOF), and post-contrast T1-weighted MRI) have greatly improved our ability to visualize and assess neurovascular structures. However, the interpretation of these complex images typically requires highly trained experts and is subject to human error, high variability, and prolonged turnaround time.

[0003] AI-based methods have been proposed to automatically interpret medical images to facilitate decision-making and support intervention—for example, in cases where a patient is suspected of having a stroke—thus reducing the time to treatment. However, such conventional AI-based methods for achieving continuous vessel segmentation suffer from reduced robustness, particularly in the presence of signal loss, noise, vessel tortuosity, calcification, and proximity to bone or bifurcations. Furthermore, such conventional AI-based methods for detecting large vessel occlusions (LVOs) cannot identify the exact location of the occlusion within the vascularity model.

[0004] To address these challenges, various methods have been proposed. EP 4 160 529 A1 discloses a method for tracing vessel trees and detecting LVOs in medical imaging. The proposed method employs a probabilistic approach to generate a vessel tree based on anatomical landmarks and vessel centerlines identified and / or determined in medical images of a patient's vessels.

[0005] Deep learning has driven substantial progress in the fields of vascular tree segmentation, a key step in constructing vascular trees, and neurovascular anomaly detection. For this task, cutting-edge methods utilize architectures similar to U-Net to generate binary vascular masks from medical images, achieving excellent Dice overlap. However, these techniques are often hampered by their inability to generalize effectively to a wide range of pathological scenarios, limiting their applicability.

[0006] Despite significant progress, a robust, efficient, and accurate pipeline that can both accurately identify blood vessels and accurately identify anomalies in a single system is still lacking. Summary of the Invention

[0007] It is an object of the present invention to provide methods and apparatus for training and employing AI models to identify vascular abnormalities with increased accuracy.

[0008] To achieve this object, a training data preparation method, a medical data analysis method, a device for carrying out these methods, and a computer program product are proposed according to the independent claims. Advantageous embodiments are the subject matter of the dependent claims.

[0009] In a training data preparation method, an input medical image of a patient's blood vessels is received. A blood vessel segmentation is determined based on the input medical image. Anatomical landmarks in the blood vessel segmentation are identified and annotated to generate an annotated blood vessel segmentation. The annotated blood vessel segmentation is stored as training data.

[0010] Such annotated training data allows for improved training of AI models to not only recognize vascular trees from medical images, but also to simultaneously identify landmarks in the vascular trees.

[0011] In some embodiments, prior to storing the annotated vessel segmentation as training data, the method includes inserting anomalies into the vessel segmentation.

[0012] Training data enhanced in this way produces AI models that are more robust at examining medical image data that includes anomalies.

[0013] In some embodiments, inserting the anomaly into the blood vessel segmentation includes removing a portion of the blood vessel segmentation to simulate an obstruction of the blood vessel.

[0014] Such training data produces AI models capable of identifying and / or locating obstructions in medical images.

[0015] In some embodiments, inserting the anomaly into the vessel segmentation includes adding a section to the vessel segmentation to simulate an aneurysm.

[0016] Such training data then produces an AI model capable of identifying and / or localizing aneurysms in medical images.

[0017] In some embodiments, the method includes: training a first artificial intelligence model using stored training data to generate a vessel segmentation based on an input medical image; and training a second artificial intelligence model using the stored training data to determine anatomical landmarks in the input medical image, wherein the training step is performed by assigning higher weights to regions where anomalies are inserted.

[0018] As a result, trained AI models will more accurately and reliably identify anomalies in medical image data.

[0019] An implementation of the medical image data analysis method includes: receiving an input medical image of a patient's blood vessels; determining blood vessel segmentation based on the input medical image with the help of a first artificial intelligence model trained as described above; and determining anatomical landmarks of the blood vessels based on the input medical image with the help of a second artificial intelligence model trained as described above.

[0020] The first and second artificial intelligence models trained thereby provide blood vessel segmentation and landmark detection with highly improved accuracy.

[0021] In some embodiments, a method includes: determining a semantic tree of a blood vessel; and determining a location where a portion of the semantic tree of the blood vessel is missing from the blood vessel segmentation, and determining the location as an abnormal location.

[0022] Such automated detection speeds up the process from acquiring a medical image to confirming a diagnosis. If a large blood vessel blockage is detected by this method, treatment can be started quickly, potentially saving lives.

[0023] In some embodiments, the method includes: generating a surface model based on blood vessel segmentation; calculating, for multiple segments, a local blood vessel radius as the distance between the segment and a centerline of the surface model; and determining a position where the difference in the local blood vessel radius between two segments exceeds a threshold as an abnormal position.

[0024] In this way, aneurysms can be detected quickly and treatment can therefore be started immediately.

[0025] In some embodiments, a U-Net segmentation network is used as the first artificial intelligence model and / or the second artificial intelligence model, wherein the training data is prepared such that the vessel landmark region is labeled as foreground.

[0026] Such a network is particularly robust for detecting biological features and can be well trained to accurately detect landmark regions. Applying another U-Net to the landmark regions can produce very accurate vessel segmentation in those regions.

[0027] In some embodiments, a U-Net network is used as the first artificial intelligence model and / or the second artificial intelligence model, wherein the U-Net network is trained based on training data to detect objects of interest and simultaneously perform at least one auxiliary task.

[0028] In this way, the same network can, for example, detect anomalies while also providing vessel segmentation.

[0029] In some embodiments, a method includes: determining a semantic tree of the blood vessel; and tracing a path from an entry point to an anomaly along the semantic tree of the blood vessel.

[0030] It will be appreciated that the method steps as disclosed below, above with respect to the embodiments of the present disclosure and / or identified in the claims may be implemented according to a dedicated processing device or by a processor adapted to perform these steps in a computer-implemented manner. Therefore, the object is also solved by an apparatus comprising means for performing the above-mentioned method.

[0031] Furthermore, a computer program is proposed to solve the problem, the computer program comprising instructions which, when executed by a computer, cause the computer to perform the steps of any of the above methods.

[0032] This will allow the computer to achieve the above-mentioned advantages of the executed method.

[0033] The computer program may be provided as or on a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to perform the steps of any of the above-mentioned methods.

[0034] In this way, a computer program may advantageously be provided. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Further embodiments and advantages can be gathered from the accompanying drawings which schematically illustrate embodiments of the invention. In particular:

[0036] Figure 1 A 3D representation of the semantic vessel tree is shown;

[0037] Figure 2 A schematic flow chart illustrating an embodiment of a method for preparing training data;

[0038] Figure 3 A 3D representation of the annotated vessel segmentation is shown;

[0039] Figure 4 A schematic flow chart illustrating an embodiment of a method for preparing training data;

[0040] Figure 5 A schematic flow chart illustrating an embodiment of a method for training an AI module;

[0041] Figure 6 A schematic flow chart illustrating an embodiment of a method for determining an intervention path;

[0042] Figure 7 shows example path planning results from CTA images; and

[0043] Figure 8 Embodiments of a device for preparing training data and / or for analyzing medical image data are shown. DETAILED DESCRIPTION

[0044] The present invention generally relates to methods and systems for vascular anomaly detection in medical imaging. Specifically, the present invention relates to methods and systems for AI-assisted vessel segmentation and centerline detection. Furthermore, the present invention relates to methods for preparing training data for at least one artificial neural network and to methods for training such an artificial neural network.

[0045] Embodiments of the present invention are described herein to provide a visual understanding of such methods and systems. A digital image typically includes a digital representation of one or more objects (or shapes). Digital representations of objects are generally described herein in terms of identifying and manipulating objects. Such manipulations are virtual manipulations performed in the memory or other circuitry / hardware of a computer system. Therefore, it should be understood that embodiments of the present invention can be performed within a computer system using data stored within the computer system.

[0046] Embodiments described herein provide for vascular anomaly detection in medical imaging. Embodiments described herein use semantic knowledge of anatomical landmarks of a vessel tree, combined with vessel centerlines identified using a deep learning model trained as described herein, to detect vascular anomalies such as aneurysms and / or large vessel occlusions (LVOs). Furthermore, embodiments described herein use such semantic knowledge to automatically localize LVOs and aneurysms and calculate 3D paths for intervention planning.

[0047] In fact, all humans have a vascular tree that follows a common structure10, e.g. Figure 1The vascular tree shown in . When a medical image of the vascular system is acquired by a medical image capture device (such as computed tomography angiography (CTA) and magnetic resonance imaging (e.g., magnetic resonance angiography (MRA), time-of-flight magnetic resonance angiography (MRTOF), and post-contrast T1-weighted MRI devices)), a three-dimensional representation of the effects measured by the device is returned as a medical image. Therefore, the expression "image" is not limited to a two-dimensional representation. An image may generally include pixels (picture elements), wherein each pixel represents a physical value measured at a location associated with the pixel. Such physical values ​​may specifically represent, but are not limited to, brightness, color, reflectivity, degree of transmittance, or magnetic resonance intensity. In some image capture devices, such as CT devices, an image in the traditional sense may not be generated immediately by the diagnostic process. Instead, multiple items of raw data may be combined by an algorithm to form or reconstruct a medical image. In such an embodiment, a pixel in the resulting calculated image may not be associated with only one direct surface or volume element of the recorded item or person, but its value may be the result of a combination of multiple measurement points.

[0048] Figure 2 A method 100 for generating training data for training a machine learning network from an input medical image according to one or more embodiments is shown. The steps of the method 100 may be performed by one or more suitable computing devices such as, for example, a general-purpose computer.

[0049] At step 102 of the method, an input medical image of a patient's blood vessels is received, such as the medical image described above.

[0050] At step 104, an initial vessel segmentation is determined based on the input medical image. In some embodiments, step 104 is performed by applying a threshold-based approach, where pixels of the medical image that fall within a certain value range are considered to be part of the vessel tree. Other such approaches are known, such as the method described in the initially mentioned EP 4 160 529 A1, which has the advantage of improved accuracy over threshold-based approaches. Some U-Nets have been trained to determine vessel segmentation in specific pathology scenarios.

[0051] Furthermore, some computer applications such as 3D Slicer have an interactive interface for configuring an automated implementation of step 104, which generates a rough initial vessel segmentation. However, this initial segmentation typically includes many false positives that are non-vessel segmentations.

[0052] In step 106 , the segmentation quality is improved.

[0053] In step 108, anatomical landmarks of the vessels are determined based on the vessel segmentation. For this purpose, an AI model can be employed. Similarly, an extension to the 3D Slicer application (named the "Vascular Modeling Toolkit" released at (https: / / github.com / vmtk / SlicerExtension-VMTK)) can be employed to automatically detect anatomical landmarks. After the centerline is automatically determined, the operator can modify or correct the anatomical landmarks to improve their accuracy.

[0054] Other computer applications that allow similar automatic anatomical landmark detection may exist and may be used for the purpose of improving step 108 .

[0055] Likewise, in step 110, the vessel tree 10 may be identified and annotated such as Figure 3 . The vascular landmarks 14 shown in . The vascular or anatomical landmarks may include the centerline of the blood vessel, the carotid frontal, the carotid artery merge, the middle cerebral artery, the basilar artery branch within the skull, and the vertebral artery merge, etc. Step 110 may also include using the method described in EP4160529A1.

[0056] In step 112, the training data package is stored as training data. The training data package may include one or more of the following: the input medical image received in step 102, vessel segmentation, anatomical landmarks such as centerlines, and / or annotations.

[0057] Applying the method 100 to a plurality of input medical images produces a corpus of annotated training data.

[0058] In order to improve the detection accuracy, the Figure 4 Steps 102, 104, 106, 108, 110, and 112 of method 100 are equivalent to steps 202, 204, 206, 208, 210, and 212 of method 200, respectively.

[0059] However, the method 200 comprises an additional step 214, wherein an anomaly is inserted into the vessel segmentation.The anomaly to be inserted may be any possible anomaly that is known to occur.

[0060] For example, in some embodiments of step 214, a large vessel occlusion (LVO) may be inserted (particularly randomly) as an anomaly into one of the segments on the tree. To this end, essentially, portions of the vessel segmentation are removed to simulate the occlusion of the vessel, since the vessel segmentation represents a blood-filled vessel.

[0061] As would be the case in a real-world LVO, the downstream segmentation and centerline annotations in the corresponding segment will be removed. No blood would flow through a real LVO, so this will simulate a cessation of blood flow. If the LVO segment selected as the location of the random anomaly is the only blood supply to a downstream segment of the vascular tree, the corresponding downstream segment will also be removed. To this end, a representation of a healthy vascular tree that includes downstream relationships can be provided.

[0062] For example, if an LVO is inserted in the ICA, the downstream ICA segment will be removed, but M1 and M2 will be preserved due to the presence of alternative blood supply. However, if an LVO is inserted in M1, the remaining portion of M1 along with M2 will be removed to simulate real-world conditions. In CTA, MRA, or MRTOF scans, the intensity of the corresponding removed region will be nonlinearly transformed to match the intensity of the surrounding brain tissue.

[0063] In some embodiments of step 214, aneurysms may be inserted (particularly randomly) into the vessel segmentation. Since an aneurysm will make the vessel appear larger, a segment is added to the vessel segmentation to simulate the aneurysm.

[0064] Aneurysms typically appear as a limited enlargement of the vascular tree or a small surface mass. Potential methods for detecting such abnormalities are to measure the local radius or distance to the centerline. However, due to the small size of aneurysms in the early stages of the disease, such methods may be insensitive due to mis-segmentation or displacement of the automatic centerline.

[0065] Because the vessel segmentation digitally represents the physical vessel form, this form can be algorithmically altered. An aneurysm can be inserted into the vessel segmentation, for example, as a small mass or localized enlargement on the vessel surface. Both its size and location will be random. While the vessel segmentation is modified, the potential centerline annotation remains unchanged, allowing the AI ​​system to identify the true centerline in such pathologies.

[0066] In a system trained using the resulting training data, this will improve the detection of aneurysms, particularly at early stages.

[0067] To prepare training data based on medical input images from CTA / MRA / MRTOF scans, the intensity of the inserted region will be set to that of the attached vessel, with a small amount of smoothing on the edges of the simulated lesion.

[0068] In some embodiments, a visual inspection step may be provided before storing the annotated vessel segmentation data in step 212. In the visual inspection step, an operator may, for example, visually inspect the enhanced vessel segmentation data to ensure its quality so that the interpolation does not result in an obviously incorrect result.

[0069] If an anomaly is inserted in step 214 or the input medical image received in step 102 , step 202 already includes an anomaly, the type of the anomaly and the location within the vessel segmentation may be stored with the training data package in step 112 , step 212 .

[0070] In such Figure 5 At step 302 of the training method 300 shown in FIG, a first AI model is trained using the training data obtained by method 100 and / or method 200 to determine blood vessel segmentation from medical image data. At step 304, a second AI model is trained using the training data obtained by method 100 and / or method 200 to determine anatomical landmarks such as centerlines from medical image data.

[0071] like Figure 6 The medical analysis method 400 shown in FIG4 includes step 402, wherein an input medical image of a patient's blood vessels is received. In step 404, the input medical image is analyzed with the aid of a first AI model to determine a blood vessel segmentation. In step 406, the input medical image is analyzed with the aid of a second AI model to determine anatomical landmarks such as a centerline.

[0072] In step 408, atypical changes in the distance between the vessel surface and the associated centerline are determined. An aneurysm is determined to be located where such atypical changes occur. In some embodiments, the distance between a segment of the vessel surface model and the centerline can be calculated for multiple segments. An anomaly can be determined to be located where the difference in the distance between two segments exceeds a threshold.

[0073] A semantic tree is constructed based on the segmentation and / or centerline in step 410. Large vessel occlusions are determined to be located where portions of the semantic tree and / or centerline are missing.

[0074] In step 412, a path is traced along the semantic vessel tree from the selected entry point to the anomaly located in step 408 or step 410. For example Figure 7As shown in , visualization 30 of the vascular tree can be enhanced with an intervention path 32, for example, in preparation for an intervention to remove an anomaly. The intervention path 32 can then guide surgical instruments through the patient's vascular system from the entry point to the anomaly, simplifying the intervention and avoiding time-consuming errors.

[0075] In some implementations, steps 408 and 410 may be replaced and / or integrated into steps 404 and 406 to utilize enhanced training data to detect anomalies.

[0076] In some such embodiments, a U-Net segmentation network can be applied as the first AI model and / or the second AI model. Landmarks (such as those annotated in the training data) can be detected by the U-Net. The U-Net is trained so that the landmark detection problem is reformatted as a segmentation problem. In the training data, a ground truth label map is generated in which the areas around the ground truth landmarks are labeled as foreground. The U-Net is then trained to produce segmentations of these landmark areas. Centerline and vessel segmentations can be input to the neural network as additional features (channels) of the training data to assist in the learning process.

[0077] In some other embodiments, a Retina U-Net as described in Bagcilar, Omer, et al., “Automated LVO detection and collateral scoring on CTA using a 3D self-configuring object detection network: a multi-center study.” (Scientific Reports 13.1 (2023): 8834) is used as the first AI model and / or the second AI model. In this model, a U-Net-like network is trained to detect objects of interest while also performing an auxiliary task (e.g., blood vessel segmentation). By jointly training the model with the two tasks, it has been shown that the rich voxel-wise information from blood vessel segmentation can be effectively utilized to improve detection accuracy.

[0078] The AI ​​system proposed in this paper possesses several advantages over previous state-of-the-art approaches, such as robustness, efficiency, and / or comprehensiveness.

[0079] Regarding robustness, AI models for vessel segmentation and centerline annotation trained using the proposed method are expected to exhibit enhanced robustness in pathological scenarios. The training method results in higher sensitivity to pathological changes, thereby facilitating more accurate detection of disease-related changes and increasing diagnostic accuracy.

[0080] Regarding efficiency, given that the proposed method is based on deep neural networks—known for their efficiency during test time—the runtime of the proposed pipeline during deployment is expected to be small. This is important, as certain acute vascular diseases require a rapid response. Furthermore, using the proposed manual segmentation pipeline will also reduce the time required to prepare manual datasets for pipeline development.

[0081] Regarding the last of the aforementioned advantages – comprehensiveness – the output of the AI ​​pipeline covers all stages of the clinical workflow, including abnormality detection, vasculature visualization, and treatment planning. This makes the system versatile and potentially suitable for a wide range of applications.

[0082] Figure 8 An apparatus 500 is shown illustrating an embodiment of an apparatus for training data preparation and / or for medical image data analysis according to an embodiment of the present invention. The apparatus 500 is configured to perform a method for training data preparation and / or for medical image data analysis according to an embodiment of the present invention.

[0083] Device 500 may be or include a (personal) computer, a workstation, a virtual machine running on host hardware, a microcontroller, or an integrated circuit. Alternatively, device 500 may be a real or virtual computer group (the technical term for a real computer group is a "cluster," and the technical term for a virtual computer group is a "cloud").

[0084] The device 500 may include an interface 502, a computing unit 504, and a memory unit 506. The interface 502 may be a hardware interface or a software interface (e.g., a PCI bus, USB, or FireWire). The computing unit 504 may include hardware and software elements, such as a microprocessor, a CPU (abbreviated as "central processing unit"), a GPU (abbreviated as "graphics processing unit"), a field programmable gate array (abbreviated as "FPGA"), or an ASIC (abbreviated as "application-specific integrated circuit"). The computing unit 504 may be configured for multithreading, that is, the computing unit may host different computing processes simultaneously, execute active and passive computing processes in parallel, or switch between active and passive computing processes. In particular, the computing unit 504 may be represented as a processor. The memory unit 506 may include one or more databases. Each of the interface 502, the computing unit 504, and the memory unit 506 may include several subunits configured to perform different tasks and / or be spatially separated.

[0085] The device 500 can be connected to one or more databases via a network. The network can be implemented as a LAN (short for "local area network"), in particular a WiFi network, or any other local connection. Alternatively, the network can be the Internet. In particular, the network can be implemented as a VPN (short for "virtual private network"). Alternatively, the database can also be integrated into the device 500, for example, the database can be stored in the memory unit 506 of the device 500. In this case, the database is connected via an internal connection.

[0086] The systems, devices, and methods described herein can be implemented using digital circuitry or one or more computers utilizing known computer processors, memory units, storage devices, computer software, and other components. Typically, a computer includes a processor for executing instructions and one or more memories for storing instructions and data. A computer may also include or be coupled to one or more mass storage devices such as one or more magnetic disks, internal hard disks and removable disks, magneto-optical disks, optical disks, and the like.

[0087] The systems, devices, and methods described herein can be implemented using computers operating in a client-server relationship. Typically, in such a system, the client computer is remotely located from the server computer and interacts via a network. The client-server relationship can be defined and controlled by computer programs running on the respective client and server computers.

[0088] The systems, devices, and methods described herein can be implemented within a network-based cloud computing system. In such a network-based cloud computing system, a server or another processor connected to the network communicates with one or more client computers via the network. For example, a client computer can communicate with a server via a web browser application resident on and operating on the client computer. The client computer can store data on the server and access the data via the network. The client computer can send a request for data or a request for an online service to the server via the network. The server can perform the requested service and provide the data to the client computer(s). The server can also send data suitable for causing the client computer to perform a specified function (e.g., perform a calculation, display specified data on a screen, etc.). For example, the server can send steps or functions suitable for causing the client computer to perform the methods and workflows described herein (including Figure 2 、 Figure 4 、 Figure 5 or Figure 6 Certain steps or functions of the methods and workflows described herein (including Figure 2 、 Figure 4 、 Figure 5 or Figure 6 One or more of the steps or functions of the method and workflow described herein may be performed by a server or by another processor in a network-based cloud computing system. Figure 2 、 Figure 4 、 Figure 5 or Figure 6 One or more of the steps of the method and workflow described herein may be performed by a client computer in a network-based cloud computing system. Figure 2 、 Figure 4 、 Figure 5 or Figure 6 One or more of the steps of ) may be performed in any combination by a server and / or by a client computer in a network-based cloud computing system.

[0089] The systems, apparatus, and methods described herein may be implemented using a computer program product tangibly embodied in an information carrier (e.g., a non-transitory machine-readable storage device) for execution by a programmable processor; and the methods and workflow steps described herein (including Figure 2 、 Figure 4 、 Figure 5 or Figure 6 One or more of the steps or functions of a program may be implemented using one or more computer programs executable by such a processor. A computer program is a collection of computer program instructions that can be used, directly or indirectly, in a computer to perform a specific activity or produce a specific result. A computer program may be written in any form of programming language, including compiled or interpreted languages, and a computer program may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0090] Designations such as "first," "second," or "third" are used in this document only to distinguish between similar but different items. They do not specify any type of hierarchy.

[0091] The foregoing specific embodiments should be understood to be illustrative and exemplary in all aspects, rather than restrictive, and the scope of the invention disclosed herein is not determined by the specific embodiments, but by the claims interpreted as if they were in accordance with the full scope permitted by patent law. It should be understood that the embodiments shown and described herein are merely illustrative of the principles of the invention, and it should be understood that various modifications can be implemented by those skilled in the art without departing from the scope of the invention. Those skilled in the art can implement various other feature combinations without departing from the scope of the invention.

[0092] Independent of grammatical usage of the term, individuals who identify as male or female are also included within the term.

Claims

1. A computer-implemented training data preparation method (100, 200), comprising: receiving (102, 202) an input medical image of a patient's blood vessels; determining (104, 204) a blood vessel segmentation based on the input medical image; identifying and annotating (110, 210) anatomical landmarks in the vessel segmentation to produce an annotated vessel segmentation; and The annotated vessel segmentations are stored (112, 212) as training data.

2. The method according to claim 1, wherein Before storing the annotated blood vessel segmentation as training data, the method comprises: Anomalies are inserted (214) into the vessel segmentation.

3. The method according to claim 2, wherein: Inserting (214) anomalies into the vessel segmentation includes: The segmented portion of the blood vessel is removed to simulate occlusion of the blood vessel.

4. The method according to claim 2 or 3, wherein: Inserting (214) anomalies into the vessel segmentation includes: Segments are added to the vessel segmentation to simulate aneurysms.

5. The method according to any one of the preceding claims, comprising: using the stored training data to train (302) a first artificial intelligence model to generate a blood vessel segmentation from an input medical image, and A second artificial intelligence model is trained (304) using the stored training data to determine anatomical landmarks in the input medical image, wherein the training steps (302, 304) are performed by assigning higher weights to regions where abnormalities are inserted into the training data.

6. A medical image data analysis method (400), comprising: receiving (402) an input medical image of a patient's blood vessels; determining (404) a blood vessel segmentation from the input medical image by means of a first artificial intelligence model trained according to claim 5; Anatomical landmarks of the blood vessels are determined (406) from the input medical image by means of a second artificial intelligence model trained according to claim 5.

7. The method according to claim 6, comprising: Determine the semantic tree of the vessel, and A location where a portion of the semantic tree of the blood vessel is missing from the blood vessel segmentation is determined (410), and the location is determined to be an abnormal location.

8. The method according to claim 6 or 7, comprising: generating a surface model based on the blood vessel segmentation; For a plurality of segments, calculating a local vessel radius as a distance between the segment and a centerline of the surface model; A location where the difference in local vessel radius between two segments exceeds a threshold is determined ( 408 ) as an abnormal location.

9. The method according to any one of claims 6 to 8, comprising: A U-Net segmentation network is used as the first artificial intelligence model and / or the second artificial intelligence model, wherein the training data is prepared such that a blood vessel landmark region is labeled as a foreground.

10. The method according to any one of claims 6 to 9, comprising: A U-Net network is used as the first artificial intelligence model and / or the second artificial intelligence model, wherein the U-Net network is trained according to the training data to detect objects of interest and simultaneously perform at least one auxiliary task.

11. The method according to any one of claims 7 to 10, comprising: tracing (412) a path from a predefined surgical entry point to the anomaly along a semantic tree of the vessel, and The path is saved as a guide for the surgical procedure.

12. An apparatus comprising means for performing the method according to any one of the preceding claims.

13. A computer program comprising instructions for causing a computer to perform the steps of the method according to any one of claims 1 to 11 when the program is executed by the computer.

14. A computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 11.

Images