A method and device for recommending a liver cancer embolism placement point planning

Abstract

The application provides a method and device for recommending a hepatic cancer embolism placement point planning, wherein the method comprises the following steps: obtaining input data; segmenting the input data to obtain a segmentation result of the input data, and classifying the segmentation result to determine whether there is a high-risk tumor in the liver; if there is a high-risk tumor in the liver, calculating an embolism placement point according to the segmentation result, generating an embolism placement point planning result; synthesizing the embolism placement point planning result with an original medical image to generate a recommended scheme, and outputting the recommended scheme. Through deep learning, machine learning and traditional image processing methods, various organs such as patient arterial vessels, livers, tumors and bones are automatically segmented, and according to the automatically generated segmentation information, recommended information is generated to recommend the embolism placement point of each tumor, so as to assist doctors in comprehensively and deeply understanding the condition of the patient.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and apparatus for planning and recommending embolization sites for liver cancer. Background Technology

[0002] The number of patients with liver diseases has been steadily increasing year by year, seriously endangering human health. For early-stage liver cancer, the main treatment is surgical resection. However, liver cancer has an insidious onset and is highly malignant, and most patients are already in the middle or late stages when they seek medical attention. The success rate of direct surgical resection is low, so non-surgical treatment is often used.

[0003] In recent years, with the development of interventional radiology, interventional therapy has been recognized as the preferred treatment for intermediate and advanced liver cancer. One of the most commonly used interventional treatment methods is hepatic artery chemoembolization (HACE). The procedure involves inserting a needle into the femoral artery under real-time medical imaging monitoring. A catheter sheath is then introduced through the needle, followed by the placement of a catheter and guidewire. Guided by the guidewire, the catheter is selectively inserted into the arterial blood supply to the liver tumor. Chemotherapy drugs and embolic agents are then injected into this blood supply vessel through the catheter, completely blocking it and preventing the tumor from absorbing nutrients, thereby inhibiting and killing liver cancer tissue. Combined with other treatment options, HACE may become an alternative to surgery, bringing new hope to patients with unresectable tumors. However, interventional therapy currently faces the problem of the embolization area not completely corresponding to the actual tumor kill area, leading to residual tumor, recurrence, and metastasis. In addition, hepatic artery chemoembolization has a long operation time, especially when there are many embolization sites. The contrast agent used in arterial angiography enters the bloodstream through blood circulation, so a large dose of contrast agent is required, which is highly harmful to the patient's body.

[0004] Currently, computer-aided techniques for hepatic artery chemoembolization are still very scarce, making it difficult to provide effective support for this procedure. Summary of the Invention

[0005] The present invention aims to provide a method and apparatus for planning and recommending liver cancer embolization placement sites to overcome or at least partially solve the above-mentioned problems.

[0006] To achieve the above objectives, the technical solution of the present invention is specifically implemented as follows:

[0007] One aspect of the present invention provides a method for planning and recommending embolization placement sites for liver cancer, comprising: acquiring input data; segmenting the input data to obtain segmentation results, classifying the segmentation results, and determining whether a high-risk liver tumor exists based on the classification; if a high-risk liver tumor exists, calculating embolization placement sites based on the segmentation results to generate embolization placement site planning results; combining the embolization placement site planning results with original medical images to generate a recommended scheme, and outputting the recommended scheme.

[0008] The process of acquiring input data includes: receiving raw input data from the user through an input device; and preprocessing the raw data to obtain the input data.

[0009] The original input data includes, but is not limited to: CT images, MRI images, digital subtraction angiography images, and ultrasound images; the input devices include, but are not limited to: computers, mobile phones, tablets, CT scanners, MRI scanners, and X-ray machines.

[0010] The preprocessing of the raw data includes: determining the data type of the raw data based on the header information of the raw data, and performing preprocessing corresponding to the type of the raw data; the preprocessing includes, but is not limited to: pixel size normalization, pixel value normalization, image size normalization, header processing, window width and window level changes, image dimension adjustment, and image orientation adjustment.

[0011] The process involves segmenting the input data to obtain segmentation results and classifying these results. This includes using image processing methods to segment the required organs and tissues from the input data and classifying the segmented lesions into high-risk and low-risk lesions. In multi-phase image types, multi-phase image fusion is used to determine tumors.

[0012] Image processing methods include, but are not limited to, machine learning and traditional image processing algorithms.

[0013] Machine learning includes, but is not limited to, semantic segmentation, semi-supervised learning, and unsupervised learning; traditional image processing algorithms include, but are not limited to, active contour models, graph cut algorithms, and region growing algorithms.

[0014] The process of calculating embolization placement points based on segmentation results and generating embolization placement point planning results includes: directly calculating embolization placement points for vascular regions that meet preset conditions based on segmentation results and generating embolization placement point planning results; or performing hepatic artery processing and embolization placement point calculation to generate embolization placement point planning results.

[0015] The process of directly calculating embolization placement points for vascular regions that meet preset conditions based on segmentation results and generating embolization placement point planning results includes: using each tumor segmentation result as a unit, using machine learning to calculate the embolization placement point for each tumor and vascular segmentation result, and generating embolization placement point planning results; or using each tumor segmentation result as a unit, using traditional image processing algorithms to obtain embolization placement points for vascular regions that meet the conditions through distance thresholds, and generating embolization placement point planning results.

[0016] The hepatic artery processing includes: using a skeleton extraction algorithm to extract the vascular skeleton of the intrahepatic arteries, refining the vessels to the width of one voxel; based on the intersection of the liver segmentation results and the arterial vascular skeleton, determining all the locations where extrahepatic arteries enter the liver; constructing a connected graph based on the intrahepatic arterial vascular skeleton, where each voxel of the intrahepatic artery corresponds to a vertex in the connected graph, designating the vertex in the connected graph corresponding to the voxel where the artery enters the liver as the root node of the tree, and using a spanning tree algorithm starting from this root node to generate a complete arterial vascular tree structure; when a node has multiple child nodes, this node is a node at a vascular branch, merging all nodes between two adjacent vascular branches that satisfy the condition of having only one child node with the next higher level node to generate the final tree structure, and removing nodes with too few voxels in the final tree structure. In the final tree structure, the root node represents the arterial branch entering the liver, each child node represents a vascular branch, and the leaf nodes represent the vascular branch where each intrahepatic artery terminal is located, with the blood flow direction from the root node to the leaf node.

[0017] The processing of hepatic artery vessels includes: using a deep learning semantic segmentation model to divide the vessel branches into vessels of different levels and performing overall segmentation of the vessel tree; or using traditional image processing algorithms to extract each vessel branch and calculate the blood flow direction.

[0018] The calculation of embolization placement points includes: using a deep learning segmentation model to perform binary classification segmentation on the vascular tree, dividing it into vascular branches that need embolization and vascular branches that do not need embolization; or using traditional image processing algorithms to calculate all embolization placement points that meet the placement conditions for each tumor.

[0019] The calculation of embolization placement points includes: detecting arteriovenous fistulas and planning embolization placement points.

[0020] The detection of arteriovenous fistulas includes: extracting voxels from the intrahepatic arteries and a certain range around the lesion; using the K-means algorithm to cluster and classify the extracted voxels; and determining whether an arteriovenous fistula exists based on an empirical threshold for arteriovenous fistulas.

[0021] The detection of arteriovenous fistulas includes using machine learning or traditional algorithms to detect features in the liver region in order to determine if an arteriovenous fistula is present.

[0022] The embolization site planning process includes: calculating the shortest distance from each branch of the intrahepatic artery to the target tumor; marking the branch when its distance to the target tumor is 1 voxel; extracting all arterial terminals based on blood flow direction and the complete arterial tree structure, and extending the vessels in a small range along the terminal direction; marking the branch to which the terminal belongs if the extension directly contacts the target tumor; after detecting all embolization branches, determining whether all embolization branches need to be merged; and when all sub-branches of a certain vessel... When all blood vessels supply blood to the tumor, it is not necessary to embolize all sub-branches of that blood vessel separately; the embolization point can be directly set on that blood vessel. Alternatively, based on the complete arterial tree structure, starting from the leaf node, the child nodes corresponding to the next higher level blood vessel are determined to see if all child nodes of that node need to be embolized. If so, the marking of the blood vessel branches corresponding to all child nodes of that node is canceled, and the blood vessel branch corresponding to that node is marked. After traversing all child nodes, all marked blood vessel branches that need to be embolized are summarized, and the blood vessel branches that need to be embolized for the target tumor are output.

[0023] The recommended protocol includes: recommended medication, recommended dosage, the vascular branch that needs to be embolized, and marking the tail end of the vascular branch as the recommended embolization placement point according to the blood flow direction.

[0024] The recommended approach also includes: if an arteriovenous fistula is detected, marking the area where voxels of high CT values ​​are located to indicate the presence of an arteriovenous fistula.

[0025] The output recommendation scheme includes: displaying the recommendation scheme in a preset output format through a display device; wherein: the display device includes, but is not limited to, mobile phones, tablets and computers; the output format includes, but is not limited to, 2D images, 3D images and time-series images.

[0026] The method also includes allowing users to replace all intermediate results.

[0027] Another aspect of the present invention provides an apparatus for planning and recommending embolization placement sites for liver cancer, comprising: an acquisition module for acquiring input data; a segmentation module for segmenting the input data to obtain segmentation results, classifying the segmentation results, and determining whether a high-risk liver tumor exists based on the classification; a calculation module for calculating embolization placement sites based on the segmentation results if a high-risk liver tumor exists, and generating embolization placement site planning results; and an output module for combining the embolization placement site planning results with original medical images to generate a recommended scheme and output the recommended scheme.

[0028] The acquisition module acquires input data in the following ways: it receives raw input data from the user through the input device; and it preprocesses the raw data to obtain the input data.

[0029] The original input data includes, but is not limited to: CT images, MRI images, digital subtraction angiography images, and ultrasound images; the input devices include, but are not limited to: computers, mobile phones, tablets, CT scanners, MRI scanners, and X-ray machines.

[0030] The acquisition module preprocesses the raw data in the following ways: it determines the data type of the raw data based on the header information of the raw data, and performs preprocessing corresponding to the type of the raw data. The preprocessing includes, but is not limited to: pixel size normalization, pixel value normalization, image size normalization, header processing, window width and window level changes, image dimension adjustment, and image orientation adjustment.

[0031] The segmentation module segments the input data in the following ways to obtain the segmentation results and classifies the segmentation results: it uses image processing methods to segment the required organs and tissues from the input data and classifies the segmented lesions into high-risk lesions and low-risk lesions; in multi-phase image types, multi-phase image fusion is used to determine tumors.

[0032] Image processing methods include, but are not limited to, machine learning and traditional image processing algorithms.

[0033] Machine learning includes, but is not limited to, semantic segmentation, semi-supervised learning, and unsupervised learning; traditional image processing algorithms include, but are not limited to, active contour models, graph cut algorithms, and region growing algorithms.

[0034] The calculation module calculates embolization placement points based on the segmentation results in the following ways: directly calculating embolization placement points for vascular regions that meet preset conditions based on the segmentation results, and generating embolization placement point planning results; or performing hepatic artery processing and embolization placement point calculation, and generating embolization placement point planning results.

[0035] The calculation module calculates embolization placement points for vascular regions that meet preset conditions based on the segmentation results in the following ways: taking each tumor segmentation result as a unit, it uses machine learning to calculate the embolization placement point for each tumor and vascular segmentation result and generates embolization placement point planning results; or taking each tumor segmentation result as a unit, it uses traditional image processing algorithms to obtain embolization placement points for vascular regions that meet the conditions through distance thresholds and generates embolization placement point planning results.

[0036] The computation module processes hepatic artery blood vessels as follows: A skeleton extraction algorithm is used to extract the vascular skeleton of the intrahepatic arteries, refining the vessels to the width of one voxel. Based on the intersection of the liver segmentation results and the arterial skeleton, the locations of all extrahepatic arteries entering the liver are determined. A connected graph is constructed based on the intrahepatic arterial skeleton, with each voxel of the intrahepatic artery corresponding to a vertex in the connected graph. The vertex in the connected graph corresponding to the voxel indicating the location of the artery entering the liver is designated as the root node of the tree. A spanning tree algorithm is used to generate a complete tree structure of the arteries, starting from this root node. When a node has multiple child nodes, this node is a node at a branch of the blood vessel. All nodes between two adjacent branches that have only one child node are merged into the next higher-level node to generate the final tree structure. Nodes with too few voxels in the final tree structure are removed. In the final tree structure, the root node represents the branch of the artery entering the liver, each child node represents a branch of the blood vessel, and the leaf nodes represent the branch where each intrahepatic artery terminal is located. The blood flow direction is from the root node to the leaf node.

[0037] The computation module processes hepatic artery blood vessels in the following ways: it uses a deep learning semantic segmentation model to divide blood vessel branches into different levels of blood vessels and performs overall segmentation of the blood vessel tree; or it uses traditional image processing algorithms to extract each blood vessel branch and calculate the blood flow direction.

[0038] The calculation module calculates embolization placement points in the following ways: it uses a deep learning segmentation model to perform binary classification of the vascular tree, dividing it into vascular branches that need embolization and those that do not; or it uses traditional image processing algorithms to calculate all embolization placement points that meet the placement conditions for each tumor.

[0039] The calculation module calculates the embolization placement point in the following way: it detects arteriovenous fistulas and plans the embolization placement point.

[0040] The calculation module detects arteriovenous fistulas in the following way: extracting voxels from the intrahepatic arteries and a certain range around the lesion; using the K-means algorithm to cluster and classify the extracted voxels; and determining whether an arteriovenous fistula exists based on an empirical threshold for arteriovenous fistulas.

[0041] The calculation module detects arteriovenous fistulas by using machine learning or traditional algorithms to detect features in the liver region and thus determine the arteriovenous fistula.

[0042] The calculation module plans the embolization placement points as follows: It calculates the shortest distance from each branch of the intrahepatic artery to the target tumor; when the distance from a branch to the target tumor is 1 voxel, the branch is marked; based on the blood flow direction and the complete arterial tree structure, all arterial terminals are extracted, and the vessels are extended in a small range along the terminal direction. If the extension can directly contact the target tumor, the branch to which the terminal belongs is marked; after detecting all the vascular branches that need to be embolized, it checks whether all the vascular branches that need to be embolized need to be merged. When a certain vessel... If all sub-branch vessels supply blood to the tumor, then it is not necessary to embolize each sub-branch of that vessel individually; the embolization point can be directly set on that vessel. Alternatively, based on the complete arterial tree structure, starting from the leaf node, determine whether all child nodes of the previous level vessel need to be embolized. If so, unmark the vessel branches corresponding to all child nodes of that node and mark the vessel branch corresponding to that node. After traversing all child nodes, summarize all marked vessel branches that need to be embolized and output the vessel branches that need to be embolized for the target tumor.

[0043] The recommended protocol includes: recommended medication, recommended dosage, the vascular branch that needs to be embolized, and marking the tail end of the vascular branch as the recommended embolization placement point according to the blood flow direction.

[0044] The recommended approach also includes: if an arteriovenous fistula is detected, marking the area where voxels of high CT values ​​are located to indicate the presence of an arteriovenous fistula.

[0045] The output module outputs the recommended solution in the following way: the recommended solution is displayed on a display device in a preset output format; wherein: the display device includes, but is not limited to, mobile phones, tablets and computers; the output format includes, but is not limited to, 2D images, 3D images and time-series images.

[0046] The intermediate results obtained from the segmentation and calculation modules can be replaced by the user.

[0047] Therefore, the method and apparatus for planning and recommending liver cancer embolization placement points provided by this invention automatically segment and generate multiple organs such as patient arteries, liver, tumors, and bones through deep learning, machine learning, and traditional image processing methods. Based on this automatically generated segmentation information, recommendation information is generated to recommend embolization placement points for each tumor, thereby assisting doctors in having a more comprehensive and in-depth understanding of the patient's condition and improving surgical efficiency. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A diagram showing the relationship between the various parts of the method and apparatus for planning and recommending liver cancer embolization placement sites provided in an embodiment of the present invention;

[0050] Figure 2 A flowchart illustrating a method for planning and recommending liver cancer embolization placement sites provided in an embodiment of the present invention;

[0051] Figure 3 A schematic diagram illustrating the differences between the one-step and two-step methods in the liver cancer embolization site planning recommendation method provided in this embodiment of the invention;

[0052] Figure 4 This is a schematic diagram of the embolization placement point planning process provided in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the device for planning and recommending liver cancer embolization placement sites provided in an embodiment of the present invention. Detailed Implementation

[0054] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0055] The present invention aims to provide a method and device for recommending embolization placement sites for liver cancer, wherein the recommendation information can help doctors understand all embolization placement sites for all tumors, and to plan or perform surgery more quickly and effectively.

[0056] The relationship between the various parts of the method and device for planning and recommending liver cancer embolization placement sites proposed in this invention is as follows: Figure 1 As shown, the middle section illustrates the overall structure of the device for planning and recommending liver cancer embolization placement sites, the left section shows the flow of the liver cancer embolization placement site planning and recommendation method, and the right section shows the user's interaction with the system. The main flow of the liver cancer embolization placement site planning and recommendation method proposed in this invention can be divided into:

[0057] (1) Input data: The user sends medical data to the storage unit of the liver cancer embolization placement site planning and recommendation device of the present invention through the interactive device, and then transmits it to the liver cancer embolization placement site planning and recommendation system.

[0058] (2) Data preprocessing: After receiving the data, the system activates the computing unit to preprocess the data, unifying the data attributes and format into the style required by the system. This step is optional.

[0059] (3) Organ and tissue segmentation: After receiving the preprocessed data, the data is segmented to obtain the segmentation results of organs and tissues such as liver, lesions, and arteries.

[0060] (4) Embolization placement planning: Based on the segmentation results, all embolization placement points, drugs, and doses for each high-risk tumor are calculated using deep learning or traditional image processing algorithms, and the presence of arteriovenous fistulas is detected.

[0061] (5) Output results: Based on the planning results, generate recommended solutions, store them in the storage unit, and then send them to the user.

[0062] Figure 1 A flowchart illustrating the method for planning and recommending liver cancer embolization placement sites provided by an embodiment of the present invention is shown. See also... Figure 1 The method for planning and recommending liver cancer embolization placement sites provided in this embodiment of the invention includes:

[0063] S1, Obtain input data.

[0064] As an optional implementation of this invention, obtaining input data includes: receiving raw input data input by a user through an input device; and preprocessing the raw data to obtain the input data.

[0065] in:

[0066] Raw input data includes, but is not limited to: CT images, MRI images, digital subtraction angiography images, and ultrasound images; input devices include, but are not limited to: computers, mobile phones, tablets, CT scanners, MRI scanners, and X-ray machines.

[0067] Preprocessing includes, but is not limited to: pixel size normalization, pixel value normalization, image size normalization, header file processing, window width and window level changes, image dimension adjustment, and image orientation adjustment.

[0068] In practice:

[0069] First, the liver cancer embolization placement site planning and recommendation system receives raw input data sent by the user. The types of input data include, but are not limited to, medical images such as CT images, MRI images, digital subtraction angiography images, and ultrasound images. The input devices include, but are not limited to, computers, mobile phones, tablets, CT scanners, MRI scanners, and X-ray machines.

[0070] Next, the raw data undergoes preprocessing. Specifically, the data type is determined based on the header information, and corresponding preprocessing is performed. For example, for CT and MRI images, preprocessing techniques include voxel value normalization, voxel size normalization, 3D image size normalization, and 3D image filtering. For ultrasound and X-ray images, specific threshold region extraction and image filtering are available. Furthermore, specific image enhancement techniques can be applied based on the image phase and imaging sequence. Ultimately, the raw input data is transformed into data that meets system requirements, possessing a unified format and attributes to ensure the stability of subsequent calculations. The data preprocessing methods here include, but are not limited to, pixel size normalization, pixel value normalization, image size normalization, header processing, window width and level adjustments, image dimension adjustment, and image orientation adjustment. The appropriate combination of preprocessing steps can be selected based on the actual usage. If the raw data source is stable, this preprocessing step can be omitted.

[0071] S2, segment the input data to obtain the segmentation results, classify the segmentation results, and determine whether there are high-risk liver tumors based on the classification.

[0072] Specifically, the input data segmentation function uses image processing methods to segment the required organs and tissues from the data, including the liver, hepatic artery, tumor, hepatic vein, portal vein, and other liver and adjacent tissue structures.

[0073] As an optional implementation of this invention, the input data is segmented to obtain segmentation results, and the segmentation results are classified, including: segmenting the required organs and tissues from the input data using image processing methods, and classifying the segmented lesions into high-risk lesions and low-risk lesions; wherein, in multi-phase image types, a multi-phase image fusion method is used to determine tumors. The image processing methods include, but are not limited to, machine learning and traditional image processing algorithms. Machine learning includes, but is not limited to, semantic segmentation, and traditional image processing algorithms include, but are not limited to, active contour models, graph cut algorithms, and region growing algorithms.

[0074] In practice, it can be implemented using the following methods:

[0075] Method 1: The preprocessed data is fed into an existing segmentation model to obtain segmentation results for organs such as arteries, lesions, and the liver. High-risk and low-risk lesions are categorized separately. For example, if a high-risk liver tumor is present, the segmentation result is passed to step S3. The segmentation model here is a deep learning semantic segmentation model trained on a dataset of labeled medical images. Alternatively, semi-supervised learning can be used to generate a segmentation model by combining labeled data with a large amount of unlabeled data. Specifically, an initial model is trained using labeled data, and then the initial model is used to calculate the segmentation results of the unlabeled data. The resulting unlabeled data is used as the labeled data. The labeled data and the model-labeled data are then fed into the model to update it. After the model update, the segmentation results for the unlabeled data can be regenerated until the model reaches its optimal state. In addition, a large amount of unlabeled data can be used to enhance the image results of the data as the labeling results. The unlabeled data can be used as the training set to train the baseline model, so that the weights of the baseline model have the ability to extract the features of the training data. Then, these weights are passed to the segmentation model, and the segmentation model is trained again using the already labeled data to obtain the final segmentation model.

[0076] Method 2: Traditional image processing algorithms, such as the graph cut algorithm or active contour model algorithm, are used to obtain liver segmentation results. Based on the liver segmentation results and image features in the enhanced CT image, region growing or clustering algorithms are used to segment tumors and arteries respectively. Finally, based on radiomics or machine learning, the benign or malignant nature of the tumor is determined. If a high-risk liver tumor is found, the segmentation result is passed to step S3. The radiomics or machine learning model here is trained on a dataset of annotated medical images within the tissue. Furthermore, in the process of determining the benign or malignant nature of the tumor, for enhanced CT images and MRI images, this invention employs a multi-phase image fusion method, extracting features from each phase image separately, and then fusing these features in a positional addition manner. The final classification result is generated based on the fused features.

[0077] S3. If a high-risk liver tumor is present, the embolization placement point is calculated based on the segmentation results, and the embolization placement point planning results are generated.

[0078] Specifically, the embolization placement point calculation process aims to plan all embolization placement points for each tumor based on the segmentation results. It also detects arteriovenous fistulas and facilitates the calculation of drugs and dosages. The drug and dosage calculation can be based on factors such as tumor size, vascular blood supply, and arteriovenous fistula status to determine recommended drugs and dosages. This can be achieved through machine learning, including but not limited to random forest algorithms and boosting algorithms; traditional image processing algorithms include, but are not limited to, thresholding methods.

[0079] As an optional implementation of this invention, calculating embolization placement points based on segmentation results and generating embolization placement point planning results includes: directly calculating embolization placement points for vascular regions that meet preset conditions based on segmentation results, and generating embolization placement point planning results; or performing hepatic artery processing and embolization placement point calculation, and generating embolization placement point planning results.

[0080] As an optional implementation of this invention, the embolization placement points are calculated directly based on the segmentation results for vascular regions that meet preset conditions, generating embolization placement point planning results. This includes: using each tumor segmentation result as a unit, calculating the embolization placement point for each tumor and vascular segmentation result using machine learning, and generating embolization placement point planning results; or using each tumor segmentation result as a unit, using traditional image processing algorithms to obtain embolization placement points for arterial vessels that meet the conditions through distance thresholds, and generating embolization placement point planning results. Machine learning can directly calculate the embolization placement point for each tumor and vascular segmentation result using segmentation models, classification models, or object detection models, etc.

[0081] The hepatic artery processing includes: segmenting the vascular branches into different levels using a deep learning semantic segmentation model to perform overall segmentation of the vascular tree; or extracting each vascular branch and calculating the blood flow direction using traditional image processing algorithms; or using a skeleton extraction algorithm to extract the vascular skeleton of the intrahepatic arteries, refining the vessels to the width of one voxel; determining the positions of all extrahepatic arteries entering the liver based on the intersection of the liver segmentation results and the arterial skeleton; constructing a connected graph based on the intrahepatic arterial skeleton, where each voxel of the intrahepatic artery corresponds to a vertex in the connected graph, specifying the voxel corresponding to the position of the artery entering the liver in the connected graph. The vertex is the root node of the tree, and a spanning tree algorithm is used to generate a complete arterial vascular tree structure starting from this root node. When a node has multiple child nodes, this node is a node at a vascular branch. All nodes between two adjacent vascular branch nodes that satisfy the condition of having only one child node are merged to the next higher level node to generate the final tree structure. Nodes with too few voxels in the final tree structure are then removed. In the final tree structure, the root node represents the arterial branch entering the liver, each child node represents a vascular branch, and the leaf nodes represent the vascular branch where each intrahepatic arterial terminal is located. The blood flow direction is from the root node to the leaf node. Specifically, the hepatic artery vascular processing involves extracting each vascular branch and calculating the blood flow direction. Deep learning semantic segmentation models can classify vascular branches into different levels of vessels and perform overall segmentation of the vascular tree. Traditional image processing algorithms include, but are not limited to, spanning trees and region growing algorithms.

[0082] The calculation of embolization placement points includes: using a deep learning segmentation model to perform binary classification of the vascular tree, dividing it into vascular branches that require embolization and those that do not; or using traditional image processing algorithms to calculate all embolization placement points that meet the placement criteria for each tumor; or the calculation of embolization placement points includes: detecting arteriovenous fistulas and planning embolization placement points. Specifically, the function of embolization placement point calculation is to calculate all embolization placement points that meet the criteria for each tumor. The deep learning segmentation model can perform binary classification of the vascular tree, dividing it into vascular branches that require embolization and those that do not; traditional image processing algorithms can determine whether each vascular branch needs embolization based on a distance threshold.

[0083] As an optional implementation of this invention, the detection of arteriovenous fistulas includes: extracting voxels from the intrahepatic arteries and a certain range around the lesion; clustering and classifying the extracted voxels using the K-means algorithm; determining the presence of an arteriovenous fistula based on an empirical threshold; or using machine learning or traditional algorithms to detect features in the liver region to determine the presence of an arteriovenous fistula. Specifically, machine learning includes, but is not limited to, semantic segmentation models, object detection models, and classification models, while traditional algorithms include, but are not limited to, the K-means algorithm.

[0084] The embolization site planning process includes: calculating the shortest distance from each branch of the intrahepatic artery to the target tumor; marking the branch when its distance to the target tumor is 1 voxel; extracting all arterial terminals based on blood flow direction and the complete arterial tree structure, and extending the vessels in a small range along the terminal direction; marking the branch to which the terminal belongs if the extension directly contacts the target tumor; after detecting all embolization branches, determining whether all embolization branches need to be merged; and when all sub-branches of a certain vessel... When all blood vessels supply blood to the tumor, it is not necessary to embolize all sub-branches of that blood vessel separately; the embolization point can be directly set on that blood vessel. Alternatively, based on the complete arterial tree structure, starting from the leaf node, the child nodes corresponding to the next higher level blood vessel are determined to see if all child nodes of that node need to be embolized. If so, the marking of the blood vessel branches corresponding to all child nodes of that node is canceled, and the blood vessel branch corresponding to that node is marked. After traversing all child nodes, all marked blood vessel branches that need to be embolized are summarized, and the blood vessel branches that need to be embolized for the target tumor are output.

[0085] In the above optional implementations, directly calculating the embolization placement point for vascular regions that meet preset conditions based on the segmentation results is a one-step method. Performing hepatic artery processing and embolization placement point calculation to generate embolization placement point planning results is a two-step method. The two-step method can obtain more patient information. The specific differences between the one-step and two-step methods are as follows: Figure 3 As shown.

[0086] In practice, it can be implemented in the following ways:

[0087] Method 1: Using each tumor segmentation result as a unit, traditional image processing algorithms are used to obtain suitable arterial embolization placement points through distance thresholds, and embolization placement point planning results are generated.

[0088] Specifically, this method is a one-step approach. Using each tumor segmentation result as a unit, a threshold method is used to select all vascular regions whose minimum distance to the tumor is less than a threshold. All eligible vascular regions are then separated into different vascular branches according to connected components. The point closest to the blood vessel in each connected component is selected as the embolization placement point. This process is repeated for all tumors to generate the final embolization placement point plan. Alternatively, all tumors can be input at once to directly calculate all vascular regions requiring embolization; however, this approach is not conducive to subsequent targeted treatment of specific tumors by physicians.

[0089] Method 2: Using each tumor segmentation result as a unit, machine learning is used to calculate the embolization placement point based on each tumor and blood vessel segmentation result, and embolization placement point planning results are generated.

[0090] Specifically, this method is also a one-step approach. It uses each tumor segmentation result as a unit, and performs binary classification or segmentation on all vascular regions using a classification model or semantic segmentation model, classifying or segmenting them into vascular branches requiring embolization and vascular regions that do not. This process iterates through all tumors to generate the final embolization placement plan. Alternatively, all tumors can be input at once to directly calculate all vascular regions requiring embolization, but this approach is not conducive to subsequent targeted treatment of specific tumors by physicians. The deep learning model used here is trained on a dataset of medical images annotated within the tissue.

[0091] Method 3: This method is a two-step process, consisting of hepatic artery vessel treatment and embolization placement point calculation.

[0092] Specifically, the first step involves processing the hepatic artery vessels. A publicly available skeleton extraction algorithm is used to extract the vascular skeleton of the intrahepatic arteries. This skeleton extraction means refining the vessels to a width of one voxel without losing structural information. Then, based on the vascular skeleton, a spanning tree algorithm from graph theory is used to obtain the blood flow direction and extract each vascular branch. Since the intrahepatic arteries themselves exhibit a tree-like structure, the spanning tree algorithm is highly suitable for constructing the intrahepatic vascular hierarchy and determining the blood flow direction.

[0093] First, based on the intersection of the liver segmentation results and the arterial vascular skeleton, the locations of all extrahepatic arteries entering the liver are determined. Then, a connected graph is constructed based on the intrahepatic arterial vascular skeleton, with each voxel of the intrahepatic artery corresponding to a vertex in the connected graph. The vertex in the connected graph corresponding to the voxel indicating the location of the artery entering the liver is designated as the root node of the tree. A spanning tree algorithm is then used to generate a complete tree structure of the arterial vessels, starting from this root node. Because the width of each vessel in the vascular skeleton is one voxel, in the spanning tree, a node at that level will only have multiple child nodes when encountering a vessel branch. In the generated intrahepatic arterial vascular tree structure, nodes with only one child node belong to a branch of the vessel, while nodes with multiple child nodes are located at the intersection of multiple vessel branches. Based on this criterion, all nodes between two vessel branches that satisfy the condition of having only one child node are merged to generate the final tree structure, and nodes with too few voxels in the final tree structure are removed.

[0094] In the final tree structure, the root node represents the arterial branch entering the liver, each child node represents a vascular branch, and the leaf nodes represent the vascular branch where the terminal end of each intrahepatic artery is located. The direction of blood flow is from the root node to the leaf node, and the extraction of vascular branches is completed based on this.

[0095] In the process of processing hepatic artery blood vessels, it can also be achieved in two other ways: by using a deep learning semantic segmentation model to divide blood vessel branches into blood vessels of different levels and to perform overall segmentation of the blood vessel tree; or by using traditional image processing algorithms to extract each blood vessel branch and calculate the blood flow direction.

[0096] (1) The hepatic artery processing method is replaced with a region growing algorithm. After obtaining the entry point of the artery into the liver through the junction of the artery and the liver, the region growing algorithm grows point by point from the entry point on the vascular skeleton into the liver. Since the vascular skeleton is only one voxel wide, when only one voxel is added each time, the growth direction is the blood flow direction. When more than one voxel is added in one growth, it can be determined that a new vascular branch has been generated and marked. In this way, all vascular branches and blood flow directions are obtained.

[0097] (2) Replace the hepatic artery processing method with deep learning semantic segmentation. Obtain the opposing segmentation results of each blood vessel branch through the semantic segmentation model, and then calculate the connection relationship of each blood vessel branch according to the location of the liver entry point.

[0098] The second step involves calculating the embolization placement point. First, arteriovenous fistula (AVC) detection is performed. Since AVC causes abnormally high CT values ​​in the local liver parenchyma and often occurs around arteries and lesions, the detection steps in this invention are as follows: 1) Extract voxels from a certain range around the intrahepatic arteries and lesions, i.e., images of the intrahepatic arteries and liver parenchyma around the lesions. 2) Use the K-means algorithm to cluster and classify the extracted voxels. 3) Based on the differences between cluster centers of the K-means algorithm in a large number of medical images of intrahepatic arteries and liver parenchyma around lesions, and the number of voxels in each cluster, an empirical threshold for AVC is obtained to determine whether an AVC exists.

[0099] The K-means algorithm mentioned here is an iterative clustering algorithm. Its steps are: first, randomly select K objects as initial cluster centers; then, calculate the distance between each voxel and each cluster center; assign each voxel to its nearest cluster center; then, update the cluster centers based on the voxels within each cluster; repeat the above process until the cluster centers no longer change or meet specific conditions. In this invention, the distance between each cluster center and voxel is affected by two factors: the pixel intensity difference between the voxel and the cluster center, and the Euclidean distance from the voxel to the cluster center. For voxel C(xc,yc,zc,ic), the subscripts xc,yc,zc represent its coordinates, and the subscript ic represents its CT value. For cluster center Kq(xk,yk,zk,ik), the subscript q indicates that this is the q-th cluster center. The parameter in parentheses has the same meaning as voxel C. The formula for calculating the distance dis between voxel C and cluster center Kq is as follows:

[0100]

[0101] Here, α and β are two constants used to balance the influence of two factors on the clustering results. The distance between voxel C(xc,yc,zc,ic) and all cluster centers is calculated, and voxel C(xc,yc,zc,ic) is assigned to the cluster with the smallest dis. Each iteration performs the above process for all voxels, and then calculates the mean of x, y, z, i for each voxel in each cluster. This mean is used as the new cluster center parameter for each cluster. Iteration stops when the cluster centers no longer change.

[0102] In the process of calculating the embolization placement point, the arteriovenous fistula detection method can be replaced by the following approach: Segmenting the hepatic veins (hepatic vein and portal vein) in the arterial phase medical image using a deep learning model; determining the presence of locally high-intensity venous segments based on the voxel intensity distribution of the segmented hepatic veins; if present, it indicates that contrast agent has prematurely entered the hepatic vein through local abnormal connections, thus suggesting the possible presence of an arteriovenous fistula. Alternatively, target detection and semantic segmentation can be performed directly on the liver region of the raw data to extract the arteriovenous fistula result.

[0103] When implementing the planning for embolization placement, you can refer to... Figure 4 As shown.

[0104] S4 combines the embolization placement site planning results with the original medical images to generate a recommended plan, and outputs the recommended plan.

[0105] As an optional implementation of this invention, the recommended scheme includes: recommending a drug, a recommended dosage, the vascular branch requiring embolization, and marking the tail end of the vascular branch as the recommended embolization placement point according to the blood flow direction. The recommended scheme also includes: if an arteriovenous fistula is detected, marking the area containing voxels with high CT values ​​to indicate the presence of an arteriovenous fistula. The output recommended scheme includes: displaying the recommended scheme in a preset output format via a display device; wherein: the display device includes, but is not limited to, mobile phones, tablets, and computers; the output format includes, but is not limited to, 2D images, 3D images, and time-series images.

[0106] Specifically, after obtaining the embolization site planning results, all results are combined with the original medical images to generate a treatment plan recommendation, including recommended drugs, recommended dosages, the vascular branches requiring embolization, and marking the tail end of the vascular branch as the recommended embolization site according to the blood flow direction. Furthermore, if an arteriovenous fistula is detected, the voxel regions of classes with high CT values ​​in the K-means algorithm results are marked, indicating the presence of an arteriovenous fistula. Treatment of the arteriovenous fistula must be performed before hepatic artery chemoembolization can be performed. The results are stored in a storage unit and then sent to the user's display device. The output display devices include, but are not limited to, mobile phones, tablets, and computers. The output formats include, but are not limited to, 2D images, 3D images, and time-series images.

[0107] As an optional implementation of this invention, the method for planning and recommending liver cancer embolization placement sites provided by this invention further includes: all intermediate results can be replaced by the user. Specifically, all steps in this invention allow the user to modify the results. For example, if the user does not wish to use the automated organ and tissue segmentation results, the operator can draw the segmentation results themselves and replace the automated segmentation results through the interactive interface. Then, this method performs subsequent calculations based on the user's segmentation results.

[0108] Therefore, the method for planning and recommending embolization sites for liver cancer provided in this invention generates an overall picture of the patient's body based on CT scans or other medical images, and automatically calculates all suitable embolization sites for each tumor. This helps doctors gain a comprehensive understanding of the tumors requiring embolization, the number of arterial blood vessels supplying each tumor, the location of each embolization site, the overall course of intrahepatic arteries, and the presence of arteriovenous fistulas. This allows doctors to have a clear understanding of all embolization sites for all tumors, enabling faster and more effective surgical planning and execution.

[0109] Figure 5 This diagram illustrates the structure of a device for planning and recommending liver cancer embolization placement sites according to an embodiment of the present invention. This device applies the aforementioned method. The following is only a brief description of the structure of the device; for other matters not covered herein, please refer to the relevant descriptions in the aforementioned method for planning and recommending liver cancer embolization placement sites. Figure 5 The device for planning and recommending liver cancer embolization placement sites provided in this embodiment of the invention includes:

[0110] The acquisition module is used to acquire input data;

[0111] The segmentation module is used to segment the input data, obtain the segmentation results, classify the segmentation results, and determine whether there are high-risk liver tumors based on the classification.

[0112] The calculation module is used to calculate the embolization placement point based on the segmentation results if a high-risk liver tumor is present, and generate the embolization placement point planning result.

[0113] The output module is used to synthesize the embolization placement site planning results with the original medical images, generate a recommended plan, and output the recommended plan.

[0114] As an optional implementation of this invention, the acquisition module acquires input data in the following manner: receiving raw input data input by the user through an input device; preprocessing the raw data to obtain the input data.

[0115] As an optional implementation of this invention, the original input data includes, but is not limited to: CT images, MRI images, digital subtraction angiography images, and ultrasound images; the input devices include, but are not limited to, computers, mobile phones, tablets, CT scanners, MRI scanners, and X-ray machines.

[0116] As an optional implementation of this invention, the acquisition module preprocesses the raw data in the following way: it determines the data type of the raw data according to the header information of the raw data, and performs preprocessing corresponding to the type of the raw data; the preprocessing includes but is not limited to: pixel size normalization, pixel value normalization, image size normalization, header processing, window width and window level changes, image dimension adjustment and image orientation adjustment.

[0117] As an optional implementation of this invention, the segmentation module segments the input data in the following manner to obtain the segmentation result of the input data, and classifies the segmentation result: the required organs and tissues are segmented from the input data using image processing methods, and the segmented lesions are divided into high-risk lesions and low-risk lesions; wherein, in multi-phase image type images, the tumor is determined by multi-phase image fusion.

[0118] As an optional implementation of the present invention, the image processing method includes, but is not limited to, machine learning and traditional image processing algorithms.

[0119] As an optional implementation of this invention, machine learning includes, but is not limited to, semantic segmentation, semi-supervised learning, and unsupervised learning, while traditional image processing algorithms include, but are not limited to, active contour models, graph cut algorithms, and region growing algorithms.

[0120] As an optional implementation of this invention, the calculation module calculates the embolization placement point based on the segmentation results in the following ways to generate embolization placement point planning results: directly calculates the embolization placement point for the vascular region that meets the preset conditions based on the segmentation results, and generates embolization placement point planning results; or performs hepatic artery processing and embolization placement point calculation to generate embolization placement point planning results.

[0121] As an optional implementation of this invention, the calculation module directly calculates embolization placement points for vascular regions that meet preset conditions based on the segmentation results in the following ways, generating embolization placement point planning results: taking each tumor segmentation result as a unit, the embolization placement point is calculated based on each tumor and vascular segmentation result using machine learning, and embolization placement point planning results are generated; or taking each tumor segmentation result as a unit, the embolization placement point for vascular regions that meet the conditions is obtained by using traditional image processing algorithms through distance thresholds, and embolization placement point planning results are generated.

[0122] As an optional implementation of this invention, the calculation module processes the hepatic artery vessels in the following manner: using a skeleton extraction algorithm, the intrahepatic artery vessels are extracted to refine the vessels to the width of one voxel; based on the intersection of the liver segmentation results and the artery vessel skeleton, the positions of all extrahepatic arteries entering the liver are determined; a connected graph is constructed based on the intrahepatic artery vessel skeleton, where each voxel of the intrahepatic artery vessels corresponds to a vertex in the connected graph, and the vertex in the connected graph corresponding to the voxel where the artery vessel enters the liver is designated as the root node of the tree, and a spanning tree algorithm is used to generate a complete artery vessel tree structure starting from this root node; when a node has multiple child nodes, this node is a node at a vessel branch, and all nodes between two adjacent vessel branch nodes that satisfy the condition of having only one child node are merged to the next higher level node to generate the final tree structure, and nodes with too few voxels in the final tree structure are removed. In the final tree structure, the root node represents the artery vessel branch entering the liver, each child node represents a vessel branch, and the leaf nodes represent the vessel branch where each intrahepatic artery terminal is located, with the blood flow direction from the root node to the leaf node.

[0123] As an optional implementation of this invention, the computing module performs hepatic artery vascular processing in the following manner: using a deep learning semantic segmentation model to divide the vascular branches into different levels of vessels, and performing overall segmentation of the vascular tree; or using traditional image processing algorithms to extract each vascular branch and calculate the blood flow direction.

[0124] As an optional implementation of this invention, the calculation module calculates the embolization placement points in the following ways: using a deep learning segmentation model to perform binary classification segmentation on the vascular tree, dividing it into vascular branches that need embolization and vascular branches that do not need embolization; or using a traditional image processing algorithm to calculate all embolization placement points that meet the placement conditions for each tumor.

[0125] As an optional implementation of this invention, the calculation module calculates the embolization placement point in the following manner: it detects arteriovenous fistulas and plans the embolization placement point.

[0126] As an optional implementation of this invention, the calculation module detects arteriovenous fistulas in the following manner: extracting voxels from the intrahepatic artery and a certain range around the lesion; using the K-means algorithm to cluster and classify the extracted voxels; and determining whether an arteriovenous fistula exists based on an empirical threshold for arteriovenous fistulas.

[0127] As an optional implementation of this invention, the calculation module detects arteriovenous fistulas by: using machine learning or traditional algorithms to detect features of the liver region to determine the arteriovenous fistula.

[0128] As an optional implementation of this invention, the calculation module plans the embolization placement point in the following manner: calculating the shortest distance from each branch of the intrahepatic artery to the target tumor; when the distance from a branch to the target tumor is 1 voxel, the branch is marked; based on the blood flow direction and the complete arterial tree structure, all arterial capillaries are extracted, and the vessels are extended in a small range along the capillaries; if the extension can directly contact the target tumor, the branch to which the capillary belongs is marked; after detecting all the vascular branches that need to be embolized, it is determined whether all the vascular branches that need to be embolized need to be combined with... Furthermore, when all sub-branches of a certain blood vessel supply blood to the tumor, it is not necessary to embolize each sub-branchose of that blood vessel individually; the embolization point can be directly set on that blood vessel. Alternatively, based on the complete arterial tree structure, starting from the leaf node, the child nodes corresponding to the next higher level blood vessel are determined to see if all child nodes of that node need to be embolized. If so, the marking of the blood vessel branches corresponding to all child nodes of that node is canceled, and the blood vessel branch corresponding to that node is marked. After traversing all child nodes, all marked blood vessel branches that need to be embolized are summarized, and the blood vessel branches that need to be embolized for the target tumor are output.

[0129] As an optional implementation of the present invention, the recommended scheme includes: a recommended drug, a recommended dosage, a vascular branch that needs to be embolized, and marking the tail end of the vascular branch as the recommended embolization placement point according to the blood flow direction.

[0130] As an optional implementation of the present invention, the recommended approach further includes: if an arteriovenous fistula is detected, marking the area where the voxel of the class with high CT value is located to indicate the presence of an arteriovenous fistula.

[0131] As an optional implementation of this invention, the output module outputs the recommended solution in the following manner: the recommended solution is displayed on a display device in a preset output format; wherein: the display device includes, but is not limited to, mobile phones, tablets and computers; the output format includes, but is not limited to, 2D images, 3D images and time-series images.

[0132] As an optional implementation of this invention, all intermediate results obtained by the segmentation module and the calculation module can be replaced by the user.

[0133] Therefore, the device for planning and recommending liver cancer embolization sites provided in this invention generates an overall picture of the patient's body based on CT scans or other medical images, and automatically calculates all suitable embolization sites for each tumor. This helps doctors gain a comprehensive understanding of the tumors requiring embolization, the number of arterial blood vessels supplying each tumor, the location of each embolization site, the overall course of the intrahepatic arteries, and the presence of arteriovenous fistulas. This allows doctors to have a clear understanding of all embolization sites for all tumors, enabling faster and more effective surgical planning and execution.

[0134] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for planning and recommending embolization sites for liver cancer, characterized in that, include: Get the input data; The input data is segmented to obtain the segmentation results, and the segmentation results are classified to determine whether there are high-risk liver tumors based on the classification. If a high-risk liver tumor is present, the embolization placement point is calculated based on the segmentation results, and the embolization placement point planning results are generated. The embolization placement site planning results are combined with the original medical images to generate a recommended plan, which is then output. in: The step of calculating the embolization placement point based on the segmentation result and generating the embolization placement point planning result includes: Based on the segmentation results, the embolization placement points are directly calculated for vascular regions that meet the preset conditions, generating embolization placement point planning results; or Perform hepatic artery vessel processing and embolization point calculation, and generate embolization point planning results; The step of directly calculating embolization placement points for vascular regions that meet preset conditions based on the segmentation results, and generating embolization placement point planning results, includes: Using each tumor segmentation result as a unit, machine learning is employed to calculate the embolization placement point based on each tumor and blood vessel segmentation result, and embolization placement point planning results are generated; or Using each tumor segmentation result as a unit, traditional image processing algorithms are used to obtain suitable arterial embolization placement points based on distance thresholds, and embolization placement point planning results are generated. The hepatic artery vascular treatment includes: Using a skeleton extraction algorithm, the vascular skeleton of the intrahepatic artery is extracted, refining the blood vessels to the width of 1 voxel. Based on the intersection of the liver segmentation results and the arterial vascular skeleton, the locations of all extrahepatic arteries entering the liver are determined; A connected graph is constructed based on the intrahepatic arterial vascular skeleton. Each voxel of the intrahepatic artery corresponds to a vertex on the connected graph. The vertex in the connected graph corresponding to the voxel where the artery enters the liver is designated as the root node of the tree. A spanning tree algorithm is then used to generate a complete tree structure of the artery vascular system, starting from this root node. When a node has multiple child nodes, this node is a node at a blood vessel branch. All nodes between two adjacent blood vessel branches that satisfy the condition of having only one child node are merged into the next higher level node to generate the final tree structure. Nodes with too few voxels in the final tree structure are then removed. In the final tree structure, the root node represents the arterial branch entering the liver, each child node represents a blood vessel branch, and the leaf nodes represent the blood vessel branch where each intrahepatic arterial terminal is located. The blood flow direction is from the root node to the leaf node. The hepatic artery vascular treatment includes: Using a deep learning semantic segmentation model, vascular branches are divided into vessels of different levels, and the vascular tree is segmented as a whole; or Traditional image processing algorithms are used to extract individual blood vessel branches and calculate blood flow direction; The calculation of the embolization placement point includes: The vascular tree is divided into binary classification using a deep learning segmentation model, into vascular branches requiring embolization and those not requiring embolization; or Traditional image processing algorithms are used to calculate all embolization sites that meet the placement criteria for each tumor; or Detect arteriovenous fistulas and plan embolization placement sites; or Calculate the shortest distance from each branch of the intrahepatic artery to the target tumor. When the distance from a branch to the target tumor is 1 voxel, mark that branch. Based on the blood flow direction and the complete arterial vascular tree structure, all arterial vascular terminals are extracted, and the vessels are extended in a small range along the direction of the vascular terminals. If the extension can directly contact the target tumor, the vascular branch to which the vascular terminal belongs is marked. After identifying all vascular branches that require embolization, it is determined whether all embolization branches need to be merged. If all sub-branches of a certain blood vessel supply blood to the tumor, it is not necessary to embolize each sub-branchose of that blood vessel separately; the embolization point can be directly set on that blood vessel. Alternatively, based on the complete arterial tree structure, starting from the leaf node, the child nodes corresponding to the previous level blood vessel are determined to see if all child nodes of that node need to be embolized. If so, the marking of the vascular branches corresponding to all child nodes of that node is canceled, and the vascular branch corresponding to that node is marked. After traversing all child nodes, summarize all marked vascular branches that need to be embolized, and output the vascular branches that need to be embolized for the target tumor.

2. The method according to claim 1, characterized in that, Obtaining input data includes: Receive raw data input by the user through an input device; The original data is preprocessed to obtain the input data.

3. The method according to claim 2, characterized in that, The raw data includes, but is not limited to: CT images, MRI images, digital subtraction angiography images, and ultrasound images; The input devices include, but are not limited to, computers, mobile phones, tablets, CT scanners, MRI scanners, and X-ray machines.

4. The method according to claim 2, characterized in that, The preprocessing of the raw data includes: The data type of the original data is determined based on the header information of the original data, and preprocessing is performed according to the type of the original data. The preprocessing includes, but is not limited to: pixel size normalization, pixel value normalization, image size normalization, header processing, window width and window level changes, image dimension adjustment, and image orientation adjustment.

5. The method according to claim 1, characterized in that, The step of segmenting the input data to obtain segmentation results and classifying the segmentation results includes: The required organs and tissues are segmented from the input data using image processing methods, and the segmented lesions are divided into high-risk lesions and low-risk lesions; among them, in multi-phase image types, multi-phase image fusion is used to determine tumors.

6. The method according to claim 5, characterized in that, The image processing methods include, but are not limited to, machine learning and traditional image processing algorithms.

7. The method according to claim 6, characterized in that, The machine learning methods include, but are not limited to, semantic segmentation, semi-supervised learning, and unsupervised learning. The traditional image processing algorithms include, but are not limited to, active contour models, graph cut algorithms, and region growing algorithms.

8. The method according to claim 1, characterized in that, The detection of arteriovenous fistulas includes: Voxels were extracted from intrahepatic arterial vessels and a certain area surrounding the lesion; The extracted voxels were clustered and classified using the K-means algorithm. The presence of an arteriovenous fistula is determined based on the empirical threshold for arteriovenous fistula.

9. The method according to claim 1, characterized in that, The detection of arteriovenous fistulas includes: The characteristics of the liver region are detected using machine learning or traditional algorithms to determine arteriovenous fistulas.

10. The method according to claim 1, characterized in that, The recommended protocol includes: recommended medication, recommended dosage, the vascular branch requiring embolization, and marking the tail end of the vascular branch as the recommended embolization placement point according to the blood flow direction.

11. The method according to claim 1, characterized in that, The recommended approach also includes: if an arteriovenous fistula is detected, marking the area where the voxel of the class with high CT value is located to indicate the presence of an arteriovenous fistula.

12. The method according to claim 1, characterized in that, The output of the recommended solution includes: The recommended solution is displayed on a display device in a preset output format; wherein: The display devices include, but are not limited to, mobile phones, tablets, and computers; The output formats include, but are not limited to, 2D images, 3D images, and time-series images.

13. The method according to claim 1, characterized in that, Also includes: All intermediate results can be replaced by the user.

14. A device for planning and recommending embolization placement sites for liver cancer, characterized in that, include: The acquisition module is used to acquire input data; The segmentation module is used to segment the input data to obtain the segmentation result of the input data, classify the segmentation result, and determine whether there is a high-risk liver tumor based on the classification. The calculation module is used to calculate the embolization placement point based on the segmentation results if a high-risk liver tumor is present, and generate the embolization placement point planning result. The output module is used to synthesize the embolization placement point planning results with the original medical images, generate a recommended plan, and output the recommended plan. in: The calculation module calculates the embolization placement points based on the segmentation results in the following manner, and generates embolization placement point planning results: Based on the segmentation results, the embolization placement points are directly calculated for vascular regions that meet the preset conditions, generating embolization placement point planning results; or Perform hepatic artery vessel processing and embolization point calculation, and generate embolization point planning results; The calculation module directly calculates embolization placement points for vascular regions that meet preset conditions based on the segmentation results, generating embolization placement point planning results as follows: Using each tumor segmentation result as a unit, machine learning is employed to calculate the embolization placement point based on each tumor and blood vessel segmentation result, and embolization placement point planning results are generated; or Using each tumor segmentation result as a unit, traditional image processing algorithms are used to obtain suitable arterial embolization placement points based on distance thresholds, and embolization placement point planning results are generated. The calculation module performs hepatic artery blood vessel processing in the following manner: Using a skeleton extraction algorithm, the vascular skeleton of the intrahepatic artery is extracted, refining the blood vessels to the width of 1 voxel. Based on the intersection of the liver segmentation results and the arterial vascular skeleton, the locations of all extrahepatic arteries entering the liver are determined; A connected graph is constructed based on the intrahepatic arterial vascular skeleton. Each voxel of the intrahepatic artery corresponds to a vertex on the connected graph. The vertex in the connected graph corresponding to the voxel where the artery enters the liver is designated as the root node of the tree. A spanning tree algorithm is then used to generate a complete tree structure of the artery vascular system, starting from this root node. When a node has multiple child nodes, this node is a node at a blood vessel branch. All nodes between two adjacent blood vessel branches that satisfy the condition of having only one child node are merged into the next higher level node to generate the final tree structure. Nodes with too few voxels in the final tree structure are then removed. In the final tree structure, the root node represents the arterial branch entering the liver, each child node represents a blood vessel branch, and the leaf nodes represent the blood vessel branch where each intrahepatic arterial terminal is located. The blood flow direction is from the root node to the leaf node. The calculation module performs hepatic artery blood vessel processing in the following manner: Using a deep learning semantic segmentation model, vascular branches are divided into vessels of different levels, and the vascular tree is segmented as a whole; or Traditional image processing algorithms are used to extract individual blood vessel branches and calculate blood flow direction; The calculation module calculates the embolization placement point in the following manner: The vascular tree is divided into binary classification using a deep learning segmentation model, into vascular branches requiring embolization and those not requiring embolization; or Traditional image processing algorithms are used to calculate all embolization sites that meet the placement criteria for each tumor; or Detect arteriovenous fistulas and plan embolization placement sites; Calculate the shortest distance from each branch of the intrahepatic artery to the target tumor. When the distance from a branch to the target tumor is 1 voxel, mark that branch. Based on the blood flow direction and the complete arterial vascular tree structure, all arterial vascular terminals are extracted, and the vessels are extended in a small range along the direction of the vascular terminals. If the extension can directly contact the target tumor, the vascular branch to which the vascular terminal belongs is marked. After identifying all vascular branches that require embolization, it is determined whether all embolization branches need to be merged. If all sub-branches of a certain blood vessel supply blood to the tumor, it is not necessary to embolize each sub-branchose of that blood vessel separately; the embolization point can be directly set on that blood vessel. Alternatively, based on the complete arterial tree structure, starting from the leaf node, the child nodes corresponding to the previous level blood vessel are determined to see if all child nodes of that node need to be embolized. If so, the marking of the vascular branches corresponding to all child nodes of that node is canceled, and the vascular branch corresponding to that node is marked. After traversing all child nodes, summarize all marked vascular branches that need to be embolized, and output the vascular branches that need to be embolized for the target tumor.

15. The apparatus according to claim 14, characterized in that, The acquisition module acquires input data in the following manner: Receive raw data input by the user through an input device; The original data is preprocessed to obtain the input data.

16. The apparatus according to claim 15, characterized in that, The raw data includes, but is not limited to: CT images, MRI images, digital subtraction angiography images, and ultrasound images; The input devices include, but are not limited to, computers, mobile phones, tablets, CT scanners, MRI scanners, and X-ray machines.

17. The apparatus according to claim 15, characterized in that, The acquisition module preprocesses the raw data in the following manner: The data type of the original data is determined based on the header information of the original data, and preprocessing is performed according to the type of the original data. The preprocessing includes, but is not limited to: pixel size normalization, pixel value normalization, image size normalization, header processing, window width and window level changes, image dimension adjustment, and image orientation adjustment.

18. The apparatus according to claim 14, characterized in that, The segmentation module segments the input data in the following manner to obtain segmentation results, and then classifies the segmentation results: The required organs and tissues are segmented from the input data using image processing methods, and the segmented lesions are divided into high-risk lesions and low-risk lesions; among them, in multi-phase image types, multi-phase image fusion is used to determine tumors.

19. The apparatus according to claim 18, characterized in that, The image processing methods include, but are not limited to, machine learning and traditional image processing algorithms.

20. The apparatus according to claim 19, characterized in that, The machine learning methods include, but are not limited to, semantic segmentation, semi-supervised learning, and unsupervised learning. The traditional image processing algorithms include, but are not limited to, active contour models, graph cut algorithms, and region growing algorithms.

21. The apparatus according to claim 14, characterized in that, The calculation module detects arteriovenous fistulas in the following manner: Voxels were extracted from intrahepatic arterial vessels and a certain area surrounding the lesion; The extracted voxels were clustered and classified using the K-means algorithm. The presence of an arteriovenous fistula is determined based on the empirical threshold for arteriovenous fistula.

22. The apparatus according to claim 14, characterized in that, The calculation module detects arteriovenous fistulas in the following ways: The characteristics of the liver region are detected using machine learning or traditional algorithms to determine arteriovenous fistulas.

23. The apparatus according to claim 14, characterized in that, The recommended protocol includes: recommended medication, recommended dosage, the vascular branch requiring embolization, and marking the tail end of the vascular branch as the recommended embolization placement point according to the blood flow direction.

24. The apparatus according to claim 14, characterized in that, The recommended approach also includes: if an arteriovenous fistula is detected, marking the area where the voxel of the class with high CT value is located to indicate the presence of an arteriovenous fistula.

25. The apparatus according to claim 14, characterized in that, The output module outputs the recommended solution in the following manner: The recommended solution is displayed on a display device in a preset output format; wherein: The display devices include, but are not limited to, mobile phones, tablets, and computers; The output formats include, but are not limited to, 2D images, 3D images, and time-series images.

26. The apparatus according to claim 14, characterized in that, All intermediate results obtained by the segmentation module and the calculation module can be replaced by the user.

Images