Method, device and medium for thrombus detection

Abstract

The invention discloses a thrombus detection method and device and a medium. The method is executed by at least one processor and comprises the steps that a 3D medical image at least comprising CTA of a blood vessel containing a target part of a patient is obtained, and the blood vessel contains thrombus; obtaining a blood vessel segmentation of a thrombus-related blood vessel; extracting an image layer containing thrombus-related blood vessels from the 3D medical image based on blood vessel segmentation in combination with an image layer expansion condition; based on the image layer containing thrombus-related blood vessels, performing thrombus preliminary prediction by using the trained thrombus detection network; based on the blood vessel segmentation, the priori knowledge associated with the target part and the priori knowledge associated with the thrombus, generating a review template of thrombus detection; and re-checking the preliminary thrombus prediction result by using a re-checking template to generate a re-checked thrombus prediction result, and presenting the re-checked thrombus prediction result to the user. According to the method, two types of prior knowledge of the target part and thrombus are fully utilized, and an accurate and credible thrombus detection result is efficiently and automatically generated for a user.

Description

Technical Field

[0001] This application relates to the field of medical imaging and medical data processing. More specifically, this application relates to methods, apparatus, and media for thrombosis detection. Background Technology

[0002] Cerebrovascular disease is one of the leading causes of death and disability in humans, with thrombosis being a key factor in the development of serious conditions such as stroke. Acute ischemic stroke occurs when a thrombus blocks a blood vessel, preventing the delivery of oxygen and other nutrients to downstream vessels, leading to ischemia and hypoxia in the perfusion area of ​​those vessels, and ultimately causing brain damage. Acute ischemic stroke is a leading cause of death and disability worldwide, making early diagnosis crucial. According to stroke treatment guidelines, the door-to-needle time (DNT) / door-to-puncture time (DPT) should be minimized for patients with acute ischemic stroke to achieve recanalization. The AHA / ASA (American Heart Association / American Stroke Association) sets a primary goal of DNT ≤ 1 hour for ≥50% of patients undergoing intravenous thrombolysis. For patients transported directly, endovascular treatment should be controlled within ≤1.5 hours for 50% of patients.

[0003] Before thrombectomy, patients undergo CTA (CT Angiography) scans to help doctors determine the location of the thrombus and the condition of the blood vessels, serving as a reference for surgical planning. While CTA angiography images can quickly assess the condition of cerebral blood vessels, visually determining the location of a thrombus (vascular occlusion) requires a high level of expertise from the interpreter and may not be able to provide comprehensive and accurate thrombus identification in complex situations, potentially delaying treatment. Therefore, there is an urgent clinical need for computer algorithm models that can provide doctors with comprehensive, accurate, and automated thrombus detection in a short time. In the existing technology, computer-aided thrombosis detection methods mainly rely on deep learning models to directly predict medical images. However, such methods often have the following limitations: (1) Currently, fully automatic thrombosis detection methods are usually based on the full CTA image for detection. However, the thrombosis structure is much smaller than the image size, and the detection model is prone to ignoring the small features of the thrombosis. In addition, the excessively large detection area is also prone to unnecessary thrombosis false detection. (2) In most automatic thrombosis detection models, they often rely solely on image features and fail to consider medical prior knowledge such as the relationship between thrombosis and blood vessels. Therefore, the detected thrombosis may be falsely detected in a way that contradicts prior knowledge. (3) Some existing technologies use MIP projection (Maximum Intensity Projection) to detect thrombosis areas in 2D regions. Since it only retains the voxels with the highest density in the line of sight, it may miss low-density or small-volume thrombosis. In addition, the 2D detection results cannot intuitively show the three-dimensional spatial relationship between thrombosis and blood vessels, which may also affect subsequent surgical planning.

[0004] Therefore, there is currently no thrombus detection method that can conveniently, efficiently, and accurately perform automatic thrombus detection based on 3D medical images such as CTA, thereby effectively assisting doctors in shortening diagnosis time, striving to enter the treatment process as soon as possible, and effectively supporting subsequent surgical planning. Summary of the Invention

[0005] This application addresses the aforementioned deficiencies in the prior art. There is a need for a method, apparatus, and medium for thrombus detection, capable of efficiently and accurately detecting thrombi in patients based on 3D medical images containing blood vessels at the target site. The detected thrombi should match prior knowledge of the target site and the thrombus itself, thereby effectively assisting physicians in shortening diagnostic time, minimizing the need for thrombolysis / thrombectomy, and providing accurate and reliable thrombus detection results as a reference for subsequent vascular recanalization procedures.

[0006] According to a first aspect of this application, a method for thrombosis detection is provided, the method being executed by at least one processor, the at least one processor performing the following steps: acquiring a 3D medical image containing blood vessels at a target site in a patient, wherein the blood vessels contain thrombosis, and the 3D medical image includes at least a CTA image; acquiring vascular segmentation of thrombosis-related blood vessels; extracting an image layer containing thrombosis-related blood vessels from the 3D medical image based on the vascular segmentation and in conjunction with image layer extension conditions; performing preliminary thrombosis prediction using a trained thrombosis detection network based on the image layer containing thrombosis-related blood vessels; generating a thrombosis detection verification template based on the vascular segmentation, prior knowledge associated with the target site, and prior knowledge associated with thrombosis; and verifying the preliminary thrombosis prediction result using the verification template to generate a verified thrombosis prediction result and presenting it to a user.

[0007] According to another aspect of this application, an apparatus for thrombosis detection is provided, comprising: an interface configured to: acquire a 3D medical image of a blood vessel including a target site of a patient, or acquire a 3D medical image of a blood vessel including a target site of a patient and a segmentation of a thrombosis-related blood vessel; and at least one processor configured to perform the steps of the method for thrombosis detection according to various embodiments of this application.

[0008] According to another aspect of this application, a non-transitory computer-readable storage medium is provided, having stored thereon computer-executable instructions, wherein when the computer-executable instructions are executed by a processor, the steps of the method for thrombosis detection according to various embodiments of this application are performed.

[0009] The methods, apparatus, and media for thrombus detection provided in the various embodiments of this application, on the one hand, extract 3D vascular segments related to thrombi from 3D medical images, remove layers outside the vascular region, and narrow the thrombus detection range. This makes thrombus detection more efficient and easier to focus on the subtle features of thrombi, while also reducing the false detection of thrombi in non-vascular regions and improving the accuracy of automatic thrombus detection. On the other hand, in the process of using a thrombus detection network to predict thrombi, the thrombus detection network, as a deep learning model, has inherent black-box characteristics, which may produce thrombus prediction results that contradict prior knowledge. Especially when the distribution characteristics of the training data do not completely match the data distribution in the real application scenario of the thrombus detection network, the problem of the prediction results violating the physical laws of prior knowledge may be aggravated. Based on vascular segmentation, this application further combines prior knowledge of the target site of thrombus presence and prior knowledge associated with the thrombus itself to generate a 3D verification template that can automatically verify the initial thrombus detection results. This removes thrombus predictions from thrombi automatically detected by the thrombus detection network that contradict the two types of prior knowledge, minimizing the possibility of predictions that do not conform to prior knowledge due to the thrombus detection network acting as a black box. The verified, more reliable, accurate, and easily viewable 3D thrombus prediction results are provided to the user, thereby improving the efficiency and accuracy of thrombus diagnosis, significantly shortening DNT / DPT time, and providing a more reliable basis for subsequent thrombus recanalization surgery planning.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application.

[0011] It should be understood that the foregoing general description and the following detailed description are merely illustrative and explanatory, and are not intended to limit the scope of the claimed invention. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating a method for thrombosis detection according to an embodiment of this application is shown.

[0014] Figure 2This diagram illustrates the steps for generating a review template when the target part is the head, according to an embodiment of this application.

[0015] Figure 3 This document illustrates an example of a specific manner in which the reviewed thrombosis prediction results are presented to the user according to an embodiment of this application.

[0016] Figure 4 This paper illustrates another specific example of how a reviewed thrombosis prediction result is presented to a user according to an embodiment of this application.

[0017] Figure 5 Several examples of user operations according to embodiments of this application are shown.

[0018] Figure 6 A schematic diagram of the composition of a device for thrombosis detection according to an embodiment of this application is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0020] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. Words such as "comprising" or "including" mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, without excluding other elements or objects.

[0021] The terms "first," "second," and similar words used in this application do not indicate any order, quantity, or importance, but are merely used for distinction. Words such as "including" or "comprising" mean that the element preceding the word encompasses the elements listed after it, and do not exclude the possibility of encompassing other elements as well. The execution order of the steps in the method described in conjunction with the accompanying drawings in this application is not intended to be limiting. As long as the logical relationship between the steps is not affected, several steps can be integrated into a single step, a single step can be decomposed into multiple steps, and the execution order of the steps can be changed according to specific needs.

[0022] It should also be understood that the term "and / or" in this application is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Furthermore, the character " / " in this application generally indicates that the preceding and following related objects have an "or" relationship.

[0023] To keep the following description of the embodiments of this application clear and concise, detailed descriptions of known functions and known components are omitted.

[0024] According to embodiments of this application, a method for thrombosis detection is provided, the method being executed by at least one processor. Figure 1 A flowchart illustrating a method for thrombosis detection according to an embodiment of this application is shown.

[0025] like Figure 1 As shown, firstly, in step 101, at least one processor may acquire a 3D medical image containing blood vessels at the patient's target site, wherein the blood vessels contain thrombi, and the 3D medical image includes at least a CTA image.

[0026] In some embodiments, the target site may be, for example, the head / neck, or the legs, lungs, etc., and is particularly suitable for anatomical sites with symmetrical structures in the vascular region. Without loss of generality, the following embodiments will use the head as the target site for specific description of the method of this application. However, it should be understood that when the target site is other human anatomical sites, the specific meaning of the prior knowledge involved in the method may be different, and the specific implementation method may need to be adapted, which will not be elaborated here.

[0027] In some embodiments, the 3D medical images should at least include CTA angiography images. In other embodiments, they may also include other 3D medical images containing blood vessels in different target areas of the patient, such as head and neck CTA images / CTV images / CTP images, head and neck MRA images, or head and neck CTA images / CTV images / CTP images generated based on NCCT, as well as leg CTA images, etc., which will not be listed here.

[0028] In step 102, at least one processor acquires the segmentation of the thrombosis-related vessels. Taking head CTA as an example, the thrombosis-related vessel is the Circle of Willis. When segmenting the Circle of Willis, for example, any well-trained vessel segmentation algorithm / model suitable for this anatomical structure can be used, based on the head CTA image, or the segmentation of the Circle of Willis can be acquired with the assistance of other medical images of the same patient. In other embodiments, the segmentation of the thrombosis-related vessels corresponding to the 3D medical image obtained in step 101, using an external model / algorithm, may also be acquired simultaneously. That is, steps 102 and 101 are not sequential and can be completed concurrently. Since the above vessel segmentation is based on 3D medical images, the vessel segmentation in each embodiment of this application is 3D vessel segmentation, which will not be described again below.

[0029] In step 103, at least one processor extracts an image layer containing thrombus-related vessels from the 3D medical image based on the vessel segmentation and in conjunction with image layer expansion conditions. After completing vessel segmentation, the next step is to extract the image layer containing thrombus-related vessels and remove image layers unrelated to thrombus outside the vessel region. However, in the practice of this application, it has been found that when a vessel region contains thrombus, the vessel segmentation may be affected by the thrombus and fail to include the complete vessel region, especially when the thrombus is near the vessel terminal, making the vessel structure more likely to be incomplete due to the presence of the thrombus. Therefore, in some embodiments, the range of the image layers to be extracted can be appropriately broadened by combining preset image layer expansion conditions. In this way, the scope of subsequent thrombus identification can be limited to the vessel region containing the thrombus as much as possible, making the small features of the thrombus easier to focus on and thus more accurate in identification, without losing the image layer containing the original vessel structure. As an example only, the image layer expansion conditions may include including a specified number of image layers along the direction of each vessel in the vessel segmentation into the image layer of the thrombus-related vessels in the direction of the vessel terminal. In other embodiments, a corresponding number of image layers may be extended in each direction of the vascular structures such as the Willis ring obtained by vascular segmentation. The specific extension method and degree can be set as needed. This application does not impose specific restrictions on this, as long as the extracted image layers can completely contain the thrombosis-related blood vessels.

[0030] Next, in step 104, at least one processor performs a preliminary thrombus prediction based on the image layer containing the thrombus-related vessels using a trained thrombus detection network. In some embodiments, the obtained preliminary thrombus prediction result may be, for example, a label of the predicted thrombus value represented by a mask within a 3D region of the same size as the 3D medical image or the image layer containing the thrombus-related vessels, or it may be represented in other ways such as the predicted thrombus value plus location information. This application does not specifically limit this, as long as it is consistent with the ground truth thrombus label of the training data when training the thrombus detection network.

[0031] In some embodiments, the thrombosis detection network may employ existing deep learning techniques such as convolutional neural networks, such as Unet-based deep learning network models (e.g., nnUnet), or a suitable version of the YOLO model or other real-time object detection models, or image detection / segmentation algorithm frameworks with higher detection accuracy such as Faster R-CNN / Transformer, and be trained using a training dataset with ground truth annotations for thrombosis. The specific algorithm is not limited in this application. Preferably, to improve the accuracy and precision of the trained thrombosis detection network in predicting thrombosis, the training should include training data for the target site to which the model will be applied. For example, if the target site is the head, then at least training data containing Willis loops should be used to train the thrombosis detection network.

[0032] Then, in step 105, at least one processor generates a verification template for thrombosis detection based on the vessel segmentation, prior knowledge associated with the target site, and prior knowledge associated with thrombosis. As mentioned earlier, the vessel segmentation is 3D, therefore the verification template generated here is also a 3D verification template.

[0033] Finally, in step 106, at least one processor uses the verification template to verify the preliminary thrombosis prediction result to generate a verified thrombosis prediction result and present it to the user. The verified thrombosis prediction result is also 3D, which will not be elaborated on later.

[0034] As mentioned earlier, even if the thrombus detection network has been sufficiently trained using training data from the same location, the black-box nature of deep learning networks—meaning they essentially fit statistical patterns in the training data—means that these patterns don't always perfectly align with general physical rules (i.e., prior knowledge). Therefore, when limited training data cannot cover all boundary conditions, the model's predictions are more likely to contradict prior knowledge. Unlike some existing technologies that require manual verification of the authenticity and accuracy of automatically identified thrombi, this embodiment of the application uses templates of the two types of prior knowledge described above. These templates are then automatically applied as a whole to the preliminary thrombus prediction results obtained in step 104. This includes, but is not limited to, efficiently removing erroneous thrombus detections by multiplying the verification template with the preliminary thrombus prediction results by voxels. This makes the verified thrombus detection more consistent with prior knowledge, providing a more reliable basis for doctors' diagnosis, surgical planning, and implementation.

[0035] The method for thrombosis detection according to the embodiments of this application extends the vascular segmentation to a certain extent, preserving the complete thrombosis-related vascular structure while effectively removing layers outside the vascular region. When using a thrombosis detection network for detection, it narrows the thrombosis detection range, enabling the thrombosis detection network to focus on the fine features of thrombosis and improve thrombosis detection performance (accuracy reaches 91.9%, sensitivity reaches 91.9%, experimental results are detailed below). At the same time, it can also avoid detecting thrombosis in non-vascular areas, thus avoiding providing users with false thrombosis prediction results. Furthermore, this application proposes combining prior knowledge of vascular segmentation with the target location of thrombus presence, as well as prior knowledge associated with the thrombus itself, to generate a verification template that can automatically review the preliminary thrombus prediction results. This reduces the workload of medical personnel in verifying each detected thrombus, lowers their workload and the professional threshold required for rapid and accurate interpretation of medical images. The reviewed 3D thrombus prediction results are more reliable and accurate, which can not only significantly improve the efficiency and effectiveness of thrombus diagnosis and greatly shorten the DNT / DPT time, but also more intuitively display the three-dimensional spatial relationship between thrombus and blood vessel in 3D space, providing a more reliable basis for subsequent thrombus recanalization surgery planning.

[0036] As an example only, the prior knowledge associated with thrombosis may include, for example, that thrombosis should be located inside a blood vessel. Therefore, when the detected thrombosis does not intersect with the segmented blood vessel, that is, when all voxels contained in the thrombosis are located outside the segmented blood vessel area, the detected thrombosis is considered to be inconsistent with the prior knowledge of thrombosis and can be judged as an incorrect thrombosis detection.

[0037] On the other hand, when the target site is the head, the prior knowledge associated with the target site may include, for example, that the blood vessels in the head should be symmetrical about the midsagittal plane of the head. Therefore, based on the blood vessel segmentation, the prior knowledge associated with the target site, and the prior knowledge associated with thrombosis, a verification template for thrombosis detection can be generated according to... Figure 2 Steps 201-204 shown are executed.

[0038] First, in step 201, the midsagittal plane of the patient's head is detected based on the 3D medical image. In some embodiments, specific methods for detecting the midsagittal plane of the patient's head based on 3D medical images such as CTA include using binarization processing and dilation algorithms to remove the skull and extract a three-dimensional mask of the brain parenchyma, and determining the midsagittal plane based on the major axis coordinates and rotation angle of the ellipse; or using SVM (Support Vector Machine) and Hough transform to detect the location of the longitudinal fissure of the brain, determine the initial midsagittal plane, and then select multiple undetermined planes around the initial midsagittal plane, and determine the final midsagittal plane based on the symmetry parameters and grayscale parameters of these undetermined planes; or using the distance relationship between the centroid of the head's centrally symmetrical tissues and the symmetrical points of the left and right symmetrical tissues and the head's midsagittal plane, a more accurate head midsagittal plane can be obtained through multiple rounds of surface fitting. This application does not impose specific limitations on the head midsagittal plane detection algorithm, as long as its detection time can be controlled to, for example, within 1 second or not exceeding 10 seconds, and does not introduce excessive additional delays to thrombus detection. In addition, the head midsagittal plane detection time and the time required for automatic thrombus detection are also related to the hardware configuration used. In some experiments of the embodiments of this application, when using a GPU, the head midsagittal plane detection time is about 1 second and the overall time for automatic thrombus detection is less than 10 seconds; while when using a CPU, the head midsagittal plane detection time is about 2 seconds and the automatic thrombus detection time is about 20 seconds.

[0039] Then, in step 202, considering that the target site is the head, the prior knowledge associated with the anatomical structure of the target site includes that the head vessels are symmetrical about the midsagittal plane of the head. Therefore, the head vessels can be divided into a first side and a second side that are symmetrical about the midsagittal plane of the head, based on the midsagittal plane of the head detected in step 201.

[0040] In step 203, the vascular structures of the first and second sides in the vascular segmentation are completed by comparing the first and second sides. In some embodiments, due to the presence of thrombus or due to image quality issues, at least one side of the two blood vessels may have incomplete vascular segmentation. To facilitate subsequent thrombus verification and the presentation of thrombus prediction results, the symmetry of the two blood vessels can be utilized to complete the missing vascular structure on either side. To distinguish between the existing vascular structure and the vascular structure completed using symmetry, the vascular structure completed in step 203 can be distinguished by marking it differently. For example, the center line of the completed blood vessel can be shown only with a dashed line, etc. This application does not specifically limit the method of distinction.

[0041] In step 204, a verification template for thrombosis detection is generated based on the vascular segmentation after vascular structure completion and prior knowledge associated with thrombosis. The portion of the verification template outside the connected domain of the vascular segmentation after vascular structure completion is set as an image value representing a low probability of thrombosis. As an example, the value 0 can be set as the aforementioned "image value representing a low probability of thrombosis" (the image value within the connected domain of the vascular segmentation in the verification template can be set to 1, for example). Then, when the verification template is applied to the preliminary thrombosis prediction result by multiplying the voxel values ​​one voxel at a time, thrombosis detections completely outside the connected domain of the vascular segmentation will be removed entirely at once. It is understood that if the detected thrombosis partially intersects with the connected domain of the vascular segmentation, such detected thrombosis will not be removed by the verification template and will be fully retained in the verified thrombosis detection result. In some embodiments, to avoid errors of several voxels between vessel segmentation results and thrombus detection results due to image volume effects, the connected regions of the segmented vessels are dilated by several voxels in practice. This avoids the risk of incorrectly removing real thrombi due to such errors. In other embodiments, a value greater than 0 and less than 1 can be used to represent a low probability of thrombus occurrence. In this way, it is equivalent to using prior knowledge related to thrombi to reduce the probability value of thrombus prediction outside the connected regions of the segmented vessels, thereby improving the accuracy and reliability of thrombus detection results.

[0042] The above example only uses the head as the target area. When the target area is another area, such as the leg, prior knowledge about the anatomy of the leg can be used in combination with prior knowledge related to thrombosis to generate a verification template for thrombosis detection. Specific methods are not listed here.

[0043] In other embodiments, other prior knowledge about thrombosis can also be utilized. For example, prior knowledge associated with thrombosis may include that in a symmetrical vascular structure, only one side of the vessel contains a thrombus, which can be called the diseased side vessel, while the other side usually does not contain a thrombus and can therefore be called the healthy side vessel. Additionally, locations in the vessel with extremely high blood flow velocity (such as the origin of the carotid artery) should generally not contain thrombi, which can also be used as prior knowledge associated with thrombosis. Therefore, based on the vessel segmentation after vascular structure completion and the prior knowledge associated with thrombosis, generating a verification template for thrombosis detection can further include: determining the healthy side and the diseased side in the first and second sides, and setting the portion of the vessel on the healthy side in the verification template as an image value representing a low probability of thrombosis; alternatively or additionally, the portion of the vessel where the blood flow velocity exceeds a preset threshold can be set as an image value representing a low probability of thrombosis. Similarly, a low probability of thrombosis can be represented by 0 or a decimal other than 0, which will not be elaborated here.

[0044] In addition, there is some prior knowledge associated with thrombosis. For example, the distal end of the vessel where the thrombus is detected should show no blood flow or weak blood flow on the CTA image. Therefore, if the distal vessel of the vessel where the thrombus is detected in the preliminary thrombosis prediction results shows good filling, it is considered that a hemodynamic discrepancy may occur. In this case, the image value of the detected thrombus portion should be set as an additional prompt requiring special attention to prompt the user to further identify the cause of this discrepancy, such as determining whether the thrombus detection is erroneous or whether there is a compensatory vessel along another pathway. The specific image value used to represent the additional prompt, and the corresponding operation performed by the user based on different image values, can be defined as needed, and this application does not specify them in detail.

[0045] In some embodiments, the image values ​​corresponding to each image unit in the review template include, but are not limited to, the binary representations such as "0" or "1" commonly used in the prior art. They may not even be limited to numerical representations. Instead, additional information related to thrombus detection that can be identified during the automatic thrombus detection process can be assigned to the image units in the review template associated with the thrombus detection. In this way, when generating the review thrombus prediction results based on the review template, richer additional information can be provided for each detected thrombus for doctors to refer to. In particular, it can provide more detailed evidence when doctors decide whether to directly adopt each detected thrombus or whether further correction is needed.

[0046] In other embodiments, using the verification template to verify the preliminary thrombosis prediction results to generate verified thrombosis prediction results and present them to the user may further include: if no thrombosis is detected in the affected vessel after verifying the preliminary thrombosis prediction results using the verification template, then an abnormal thrombosis detection prompt is presented to the user in the verified thrombosis prediction results. This can prevent patients with thrombosis symptoms from having their thrombosis missed due to image quality or other reasons, thus delaying treatment. Therefore, if the affected vessel is identified using methods such as comparing vessel length, but no thrombosis is actually detected on the affected side, a missed detection should be considered, and the user should be notified of this abnormality so that the user can determine whether to combine other clinical information for further judgment, etc., without specific limitations here.

[0047] The following is combined Figure 3 and Figure 4 This document details the specific method by which the reviewed thrombosis prediction results are presented to the user after reviewing the preliminary thrombosis prediction results using the aforementioned review template and generating reviewed thrombosis prediction results. It should be noted that the following presentation methods can be used individually or in combination where applicable and not contradictory; this application does not impose any restrictions on this.

[0048] In some embodiments, the preliminary thrombosis prediction results and the reviewed thrombosis prediction results may be presented to the user in a switching or comparison manner for the user to confirm and generate the final thrombosis prediction result. Figure 3 For example, the preliminary thrombosis prediction results are presented on the left, and the verified thrombosis prediction results are presented on the right, in a comparative manner. Figure 3 As shown, in the initial thrombosis prediction results, two thrombi, thrombi 31 and thrombi 32, were detected. However, after verification using the verification template generated by vessel segmentation 30, thrombi 32, which was outside the connected domain of vessel segmentation 30, were removed. Therefore, in the thrombosis prediction results verified on the right, only one thrombi was detected, namely thrombi 31, which is located within the connected domain of vessel segmentation 30. In some cases, users can verify the rationality and accuracy of prior knowledge or further verification operations on thrombosis prediction results generated by user actions by comparing the thrombosis prediction results before and after verification.

[0049] Figure 4 The upper part is a two-dimensional view of the reviewed thrombosis prediction results, and the lower part is a three-dimensional view of the reviewed thrombosis prediction results. (Example) Figure 4 As shown, when the target site is the head, the detected head midsagittal plane 40 can be presented to the user and overlaid on the preliminary thrombosis prediction result and / or the reviewed thrombosis prediction result. Figure 4Taking the reviewed thrombosis prediction results as an example, but not limited to this, it allows users to use the midsagittal plane of the head (40°) as a reference interface to analyze blood vessels on both sides, for example... Figure 4 The healthy side vessel 41 and the diseased side vessel 42 shown in the figure are compared and observed, and the thrombosis prediction result 43 is more easily analyzed and judged from the perspective of symmetry.

[0050] In other embodiments, the preliminary thrombosis prediction results and the reviewed thrombosis prediction results are presented in contrast, and / or the midsagittal plane is overlaid in the thrombosis prediction results. These presentation methods have proven in practice to provide experienced physicians with a more convenient benchmark and reference for discovering abnormalities such as the possible absence or variation of certain vascular segments in specific patients. This allows them to efficiently and conveniently discover errors or deviations in vascular segmentation, thrombosis detection, etc., that may result from these abnormalities with less attention, thereby making the final thrombosis prediction results more consistent with the actual situation.

[0051] The method according to the embodiments of this application can improve the accuracy of thrombus detection from multiple perspectives. In particular, after reviewing the thrombus prediction results by combining various anatomical structures and prior knowledge of thrombus-related expertise, the accuracy of thrombus detection basically meets the requirements of medical personnel. However, in the process of filing this application, the inventors of this application found that the reviewed thrombus detection results may still deviate from the judgment of the doctor user in some specific situations. Or, although the basic attributes such as the location and size of the automatically detected thrombus are correct, the doctor user may still want to make specific annotations, etc. Therefore, the method according to the embodiments of this application can also receive user operations based on the preliminary thrombus prediction results and / or the reviewed thrombus prediction results, update the review template based on the user operations, and use the updated review template to review the preliminary thrombus prediction results to update the reviewed thrombus prediction results. It is worth noting that when updating the review template, the user operation has a higher priority than the prior knowledge associated with thrombosis. This allows the user to operate on the thrombosis prediction results at each stage, so as to achieve the iteration of thrombosis detection results in a direction that is more accurate and more in line with user needs.

[0052] Figure 5 Several examples of user operations according to embodiments of this application are shown. For example... Figure 5 As shown, in accordance with Figure 3 and Figure 4 Based on the presentation of preliminary / reviewed thrombosis prediction results to the user, as shown, the user can perform the following actions. Figure 5 The user operation 501-user operation 505 shown, or a combination thereof.

[0053] like Figure 5 As shown, user operation 501 means that the user can retain at least one detected thrombus in the preliminary thrombus prediction result. In some embodiments, the detected thrombus in the preliminary thrombus prediction result may have been removed during the review process, but the doctor, based on experience or other clinical images and clinical information of the patient, believes that the detected thrombus in the preliminary thrombus prediction result should be retained. In this case, the thrombus can be retained in the preliminary thrombus prediction result through user operation 501 to restore the initially detected thrombus that was mistakenly removed.

[0054] In other embodiments, the physician user can also specify a location point in the preliminary or reviewed thrombus prediction results as the center point of the newly added thrombus based on experience or other patient information, through user operation 502, thereby reducing the missed detection of real thrombi in the final thrombus detection results.

[0055] In other embodiments, user operation 503 can also modify the attribute parameters of the detected thrombus in the preliminary thrombus prediction result or the reviewed thrombus prediction result, or configure attribute parameters for newly added thrombi, wherein the attribute parameters include at least location and size.

[0056] In other embodiments, user operation 504 can further annotate at least one detected thrombus in the preliminary or reviewed thrombus prediction results. This specific annotation may include, for example, at least one of the following: thrombus type (including cardiogenic or atherosclerotic), thrombus CT value, thrombus permeability, thrombus time window, thrombus treatment status, and thrombus treatment history. These annotations will be presented in the final thrombus prediction results, helping doctors in scenarios such as emergency green channels to quickly identify possible false positives by comparing medical images, thrombus detection, and thrombus annotations, thereby improving diagnostic accuracy and avoiding surgical errors. Furthermore, doctors can combine the annotated thrombus CT value and thrombus permeability to more accurately determine the thrombus type, and can also combine multiple annotation information such as thrombus type and thrombus time window to consider whether the patient is suitable for surgery and / or select a matching surgical plan, while also providing valuable information for surgical implementation. As an example only, the CT value of a thrombus can be used to determine whether it is a thrombus with high-density arterial signs and magnetically sensitive vessels. These types of thrombi are often rich in red blood cells, making thrombectomy easier and recanalization more effective; stent thrombectomy and recanalization are even more effective. Conversely, low-density thrombi rich in fibrin with lower CT values ​​are better recanalized using contact thrombectomy, and so on. The type, specific options, and content of the annotation information can be defined and added by the user as needed; this application does not impose specific restrictions on this.

[0057] In addition, users may have performed multiple rounds of user actions. In this case, user action 505 can be used to select to view the history of at least partially reviewed thrombosis prediction results, present multiple history records for comparison, and select the most suitable one as the final thrombosis prediction result.

[0058] The aforementioned user operations can be completed by the user drawing on an image containing preliminary and / or verified thrombosis prediction results. Alternatively or additionally, these user operations can also be completed based on the parsing of rich text input by the user or imported. For example, a user can provide their usual rules for manually examining thrombosis in text / rich text format, which can be newly generated or referenced from existing medical records, diagnostic reports, and other clinical data. Furthermore, to facilitate accurate parsing by the thrombosis detection network, the network can be pre-trained using the rules the user is likely to use. Alternatively, during application, historical user operations performed by a specific user can be recorded and accumulated as that user's personal habits and used for progressive training of the thrombosis detection network. In this way, each user can gradually acquire a thrombosis detection network that better suits their usage habits and, due to accumulated personal experience, becomes increasingly accurate in its detection results.

[0059] According to embodiments of this application, an apparatus for thrombosis detection is also provided. Figure 6 A schematic diagram of the composition of a device for thrombosis detection according to an embodiment of this application is shown.

[0060] like Figure 6 As shown, the device 600 includes at least an interface 601 and at least one processor 602. The interface 601 may be configured, for example, to acquire a 3D medical image containing blood vessels at a target site in the patient. In other embodiments, the interface 601 may also acquire both the 3D medical image containing the blood vessels at the target site and segmentation of thrombosis-related blood vessels. In still other embodiments, the interface 601 may also receive a trained thrombosis detection network, etc., from a model training device (not shown), which are not listed here.

[0061] In other embodiments, the device 600 may also include a storage area (not shown), which may be used, for example, to store a trained thrombosis detection network, and in some specific cases, to store a trained vascular segmentation model for vascular segmentation, so that when only a 3D medical image of the blood vessels at the patient's target site is acquired through the interface 601, the vascular segmentation model can be used to obtain the vascular segmentation of the thrombosis-related blood vessels.

[0062] In other embodiments, the storage area may also store a model training network for progressively supplementing the thrombosis detection network, so as to perform online progressive supplementing training of the thrombosis detection network based on user operations, enabling the thrombosis detection network to effectively improve the model prediction performance metrics by using user operations as supervision during application.

[0063] Interface 601 may include, for example, a network cable connector, a cable connector, a serial connector, a USB connector, a parallel connector, a high-speed data transmission adapter such as fiber optic, USB 3.0, or Xunlei, a wireless network adapter such as a WiFi adapter, or a telecommunications (3G, 4G / LTE, etc.) adapter. In some embodiments, interface 601 may, for example, directly receive 3D medical images containing blood vessels at a patient's target site from a 3D medical image acquisition device (not shown), such as a computed tomography (CT) scanner or magnetic resonance imaging (MRI) device, and may also acquire 3D medical images containing blood vessels at a patient's target site from other storage media, as well as vascular segmentation of thrombosis-related vessels based on the 3D medical images.

[0064] In some embodiments, at least one processor 602 may be configured, for example, to perform the steps of the methods for thrombosis detection according to various embodiments of this application, and finally output and present the thrombosis prediction results to the user on a display device (not shown). Specific methods and steps have been described in detail with reference to Figures xx-xx, and will not be repeated here.

[0065] In some embodiments, at least one processor 602 may be a processing device that includes more than one general-purpose processing device, such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), etc. More specifically, the processor may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor that runs other instruction sets, or a processor that runs a combination of instruction sets. The processor may also be more than one special-purpose processing device, such as an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), a system-on-a-chip (SoC), etc.

[0066] According to embodiments of this application, a non-transitory computer-readable storage medium is also provided, having stored thereon computer-executable instructions, wherein when the computer-executable instructions are executed by a processor, various steps of the method for thrombosis detection according to various embodiments of this application are performed.

[0067] Furthermore, this storage medium can also be used to store 3D medical images containing blood vessels at the patient's target site, received via an interface from 3D medical imaging devices such as CTA, or 3D medical images containing blood vessels at the patient's target site and segmented blood vessels related to thrombosis, acquired via an interface. In other embodiments, it can also receive rich text containing review rules provided by the user to assist in thrombosis detection, as well as other patient-related clinical data, etc., which are not listed here.

[0068] In some embodiments, the aforementioned non-transitory computer-readable storage medium may be, for example, read-only memory (ROM), random access memory (RAM), phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), electrically erasable programmable read-only memory (EEPROM), other types of random access memory (RAM), flash drives or other forms of flash memory, cache, registers, static memory, optical disc read-only memory (CD-ROM), digital versatile optical disc (DVD) or other optical storage, magnetic tape cassette or other magnetic storage devices, or any other possible non-transitory medium used to store information or instructions that can be accessed by a computer device.

[0069] To verify the performance of the method for thrombus detection according to the embodiments of this application, experimental validation was conducted on an experimental dataset consisting of head CT arterial phase angiography (CTA) images of 260 patients with ischemic stroke. 73 images in the experimental dataset came from a public dataset (IEEE Data:STROKE, provided by Mohamed ABDELRADY et al., data link: https: / / ieee-dataport.org / open-access / stroke-ct-series-patients-m1-thrombus-thrombectomy), and the other 187 images came from private data from a hospital. The aforementioned public dataset contained NCCT / CTA images of 88 patients with ischemic stroke, with ground truth annotations for thrombi represented by a mask ground truth (GT). Data with poor image quality was removed, and 73 images meeting the experimental requirements were selected and used as training data in the training set. The 187 CTA images in the hospital's private dataset were also annotated with thrombus ground truth using the mask GT method, with 150 of them used as training data and the other 37 as test data.

[0070] Table 1 shows a performance comparison between the thrombosis detection method according to the embodiments of this application and a prior art thrombosis detection model based on the nnUnet model architecture + PyTorch deep learning framework that does not utilize vessel segmentation and prior knowledge of the target site / thrombosis. To maintain consistency in the comparison benchmark, the thrombosis detection method of this application also uses the nnUnet model architecture + PyTorch deep learning framework; the difference lies in combining vessel segmentation, prior knowledge associated with the target site, and prior knowledge associated with the thrombosis.

[0071] Table 1. Comparison of performance indicators of the thrombosis detection method of this application with existing technologies.

[0072] In Table 1, the precision index reflects the proportion of true positives among the samples predicted as positive by the model, focusing on assessing false positives. The higher the value, the better, with a maximum of 100%. The sensitivity index measures the model's ability to identify positive samples, focusing on assessing false negatives. The higher the value, the better, with a maximum of 100%. The calculation method is shown in Equation (1).

[0073] Precision = TP / (TP + FP) Sensitivity = TP / (TP + FN) (1) Among them, TP (true positive) means that the patient actually has the disease and the test is positive, FP (false positive) means that the patient does not actually have the disease but the test is positive, TN (true negative) means that the patient does not actually have the disease and the test is negative, and FN (false negative) means that the patient actually has the disease but the test is negative.

[0074] As can be seen from Table 1, this application significantly outperforms the prior art in both accuracy and sensitivity, two key performance indicators. This demonstrates that by combining vascular segmentation with prior knowledge associated with the target site and prior knowledge associated with thrombus into the thrombus detection method, this application significantly improves the performance of automatic thrombus detection in terms of both false positives and false negatives, and can efficiently provide users with more accurate and reliable automatic thrombus detection results.

[0075] The methods, apparatus, and media for thrombosis detection according to various embodiments of this application can accurately screen image layers by combining vascular segmentation, making the automatic thrombosis detection using the thrombosis detection network more focused on the small features of the thrombosis, improving the detection rate and accuracy of thrombosis. Furthermore, by combining vascular segmentation, prior knowledge of the target location of the thrombosis, and prior knowledge associated with the thrombosis itself to generate a verification template, efficient, automatic, and overall verification of the initial thrombosis detection results is achieved, thereby removing thrombosis that contradicts prior knowledge, making the automatic thrombosis prediction results after verification more accurate and reliable. In addition, the presentation of thrombosis detection results provides users with convenient conditions for further verification, such as overlaying the midsagittal plane of symmetrical structures with the thrombosis detection results, and switching / comparing the thrombosis detection results before and after verification. This allows radiologists and clinicians to apply their rich field experience to the verification and judgment of automatically generated thrombosis prediction results more conveniently, with less attention and workload. Furthermore, it greatly reduces the professional threshold required for medical personnel to quickly and accurately interpret thrombosis based on medical images. Building upon this foundation, this application also provides users with a visualized and convenient interactive means to add, delete, modify, and annotate thrombosis detection data. Furthermore, the user's operation history can be collected and used as training data to supplement the training of the thrombosis detection network. This allows the thrombosis detection network to achieve performance improvement through progressive training and to learn the domain experience and operating habits of specific users. Consequently, the final output thrombosis prediction results are more reliable and accurate, with lower false positive and false negative rates than existing technologies. At the same time, it better meets the personalized needs of users, significantly improving the efficiency and effectiveness of thrombosis diagnosis, greatly shortening DNT / DPT time, and providing a more reliable basis for subsequent thrombosis recanalization surgery planning.

[0076] Furthermore, although exemplary embodiments have been described herein, their scope includes any and all embodiments based on this application that have equivalent elements, modifications, omissions, combinations (e.g., schemes involving intersections of various embodiments), adaptations, or alterations. Elements in the claims will be interpreted broadly based on the language used in the claims and are not limited to the examples described in this specification or during the implementation of this application, which will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered illustrative only, and the true scope and spirit are indicated by the full scope of the claims and their equivalents.

[0077] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. Other embodiments may be used by those skilled in the art upon reading the above description. Furthermore, in the above detailed description, various features may be grouped together to simplify the application. This should not be construed as an intention that a disclosed feature not claimed is necessary for any claim. Rather, the subject matter of the application may be less than all the features of a particular disclosed embodiment. Thus, the claims are incorporated herein by reference as examples or embodiments, wherein each claim is an independent, separate embodiment, and these embodiments are contemplated as being able to be combined with each other in various combinations or arrangements. The scope of this application should be determined by reference to the claims and the full scope of their equivalents.

[0078] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for detecting thrombosis, characterized in that, The method is executed by at least one processor, which performs the following steps during runtime: Acquire 3D medical images containing blood vessels at a target site in a patient, wherein the blood vessels contain thrombi, and the 3D medical images include at least CTA images; Obtain vessel segmentation for thrombosis-related vessels; Based on the blood vessel segmentation and combined with image layer expansion conditions, an image layer containing thrombosis-related blood vessels is extracted from the 3D medical image; Based on the image layer containing thrombosis-related blood vessels, a pre-trained thrombosis detection network is used to perform preliminary thrombosis prediction. Based on the blood vessel segmentation, prior knowledge associated with the target site, and prior knowledge associated with thrombus, a verification template for thrombus detection is generated. The preliminary thrombosis prediction results are reviewed using the review template to generate reviewed thrombosis prediction results and present them to the user.

2. The method according to claim 1, characterized in that, The image layer expansion conditions include at least the following: Along the direction of each blood vessel in the segmented blood vessel, a specified number of image layers in the direction of the vessel terminal are incorporated into the image layer of the thrombus-related blood vessel.

3. The method according to claim 1 or 2, characterized in that, The prior knowledge associated with thrombosis includes that the thrombosis is located inside a blood vessel; and when the target site is the head, the prior knowledge associated with the target site includes that the blood vessels in the head are symmetrical about the midsagittal plane of the head. The step of generating a thrombus detection verification template based on the vessel segmentation, prior knowledge associated with the target site, and prior knowledge associated with thrombus further includes: Based on the 3D medical image, the midsagittal plane of the patient's head is detected. Based on the midsagittal plane of the head, the blood vessels of the head are divided into a first side and a second side. With the first side and the second side compared, the blood vessel structure of the first side and the second side in the blood vessel segmentation is completed. Based on the vascular segmentation after vascular structure completion and prior knowledge associated with thrombosis, a verification template for thrombosis detection is generated. In the verification template, the portion outside the connected domain of the vascular segmentation after vascular structure completion is set as an image value representing a low probability of thrombosis.

4. The method according to claim 3, characterized in that, Based on the vascular segmentation after vascular structure completion and prior knowledge associated with thrombosis, the generated thrombosis detection verification template further includes: Identify the healthy and diseased sides in the first and second sides, and set the portion of the blood vessels on the healthy side in the review template to an image value representing a low probability of thrombosis; and / or, The portion of the blood flow velocity in the blood vessel that exceeds a preset threshold is set as an image value indicating a low probability of thrombosis. And / or, if the distal vessel from which the thrombus was detected is well filled in the preliminary thrombus prediction results, the portion of the detected thrombus in the review template is set as the image value for additional prompts.

5. The method according to claim 4, characterized in that, Using the aforementioned review template, the preliminary thrombosis prediction results are reviewed to generate a reviewed thrombosis prediction result, which is then presented to the user, further including: If no thrombus is detected in the affected blood vessel after reviewing the preliminary thrombus prediction results using the review template, an abnormal thrombus detection prompt will be displayed to the user in the reviewed thrombus prediction results.

6. The method according to claim 1 or 2, characterized in that, Using the aforementioned review template, the preliminary thrombosis prediction results are reviewed to generate reviewed thrombosis prediction results, which are then presented to the user and further include at least one or a combination of the following: The preliminary thrombosis prediction results and the reviewed thrombosis prediction results are presented to the user in a switching or comparison manner, so that the user can confirm and generate the final thrombosis prediction result; When the target site is the head, the detected head midsagittal plane is presented to the user, and the detected head midsagittal plane is superimposed on the preliminary thrombosis prediction result and / or the reviewed thrombosis prediction result.

7. The method according to claim 6, characterized in that, The method further includes: Receive user actions based on the preliminary thrombosis prediction results and / or the reviewed thrombosis prediction results; The review template is updated based on the user operation. The updated review template is used to review the preliminary thrombosis prediction result to update the reviewed thrombosis prediction result. When updating the review template, the user operation has a higher priority than prior knowledge associated with thrombosis.

8. The method according to claim 7, characterized in that, The user operation includes one or a combination of the following: At least one thrombus detected is retained in the preliminary thrombus prediction results; In the preliminary thrombosis prediction result or the reviewed thrombosis prediction result, a location point is specified as the center point of the newly added thrombosis; In the preliminary thrombosis prediction result or the reviewed thrombosis prediction result, modify the attribute parameters of the detected thrombosis, or configure attribute parameters for newly added thrombosis, wherein the attribute parameters include at least location and size; In the preliminary thrombosis prediction result or the reviewed thrombosis prediction result, at least one detected thrombosis is specifically labeled, wherein the specific label includes at least one of the following: thrombosis type, thrombosis CT value, thrombosis permeability, thrombosis time window, thrombosis treatment status, and thrombosis treatment history. Select at least a portion of the historical records of the reviewed thrombosis prediction results for comparison presentation; wherein... The user operation is performed by the user drawing on an image containing preliminary thrombosis prediction results and / or verified thrombosis prediction results; and / or, the user operation is performed based on the parsing of rich text input by the user or imported.

9. A device for thrombosis detection, characterized in that, include: The interface is configured to: acquire a 3D medical image containing blood vessels at the patient's target site, or acquire a 3D medical image containing blood vessels at the patient's target site and vascular segmentation of thrombosis-related blood vessels. At least one processor is configured to perform the steps of the method for thrombosis detection according to any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having stored thereon computer-executable instructions, wherein when the computer-executable instructions are executed by a processor, the steps of the method for thrombosis detection according to any one of claims 1 to 8 are performed.

Images