Coronary artery segmentation method based on DSA image continuous frame sequence, computer device and readable storage medium

Abstract

The present application relates to a coronary artery segmentation method based on a continuous frame sequence of DSA images, a computer device, and a readable storage medium. The method includes: obtaining a training data set, the training data set including an input data set and an output data set, the input data set being the first continuous frame with a sequence length of N in the continuous frame sequence of coronary DSA images, and the output data set being an intermediate frame pre-labeled with a target vessel and located in the middle of the first continuous frame; using the training data set to train a deep learning network model: the deep learning network model includes a feature fusion device, an encoder, and a decoder, the feature fusion device performs feature fusion on the first continuous frame to obtain a three-dimensional image, the three-dimensional image passes through the encoder and decoder in sequence, and the output of the decoder is connected to the intermediate frame; continuously training until a trained deep learning network model is obtained, using the trained deep learning network model to detect the input continuous frame sequence of coronary DSA images, and outputting a segmentation prediction map of the target vessel.

Description

Technical Field

[0001] The present application relates to the technical field of coronary artery physiology, and in particular to a coronary vessel segmentation method based on a continuous frame sequence of DSA images, a computer device, and a readable storage medium. Background Art

[0002] Coronary heart disease, also known as coronary atherosclerotic heart disease (CAD), and sometimes called ischemic heart disease, refers to heart disease caused by myocardial ischemia and hypoxia due to atherosclerosis of the coronary arteries. The coronary arteries are the only blood vessels supplying blood to the heart, and their shape resembles that of a crown, hence the name. Like other blood vessels throughout the body, the coronary arteries can also develop sclerosis and atherosclerotic changes, disrupting blood circulation to the heart and causing myocardial ischemia and hypoxia, a condition known as coronary heart disease. CAD is a common and frequently occurring disease among middle-aged and elderly people, and a serious threat to their lives.

[0003] Conventional methods for diagnosing coronary artery disease include simple, noninvasive electrocardiograms (ECGs), coronary CTA (which can only produce static images), and invasive methods such as intravascular ultrasound (IVUS) and dynamic coronary angiography. Of these, coronary angiography is considered the "gold standard" for diagnosing coronary artery disease. As the primary imaging technique for diagnosing coronary artery disease, coronary artery morphology is obtained through real-time imaging in the catheterization laboratory during coronary angiography. Quantitative coronary angiography (QCA) can also provide objective quantitative morphological measurements.

[0004] Because coronary DSA is a projection of three-dimensional coronary arteries onto a two-dimensional plane, QCA is prone to projection errors, and accurate identification of overlapping vessels requires a deep understanding of the coronary artery tree structure and considerable training. Despite the use of computer-assisted tools such as edge detection methods, manual correction is necessary for accurate segmentation of the coronary arteries. Although new image processing methods have been proposed to automatically detect entire vascular regions, the processing time required to apply multiple filters is impractical, and the identification of specific vessels is not considered. Recently, deep learning models have been introduced for DSA segmentation. However, deep learning methods for major vessel segmentation have not yet achieved the predictive accuracy required for clinical application. Summary of the Invention

[0005] Based on this, it is necessary to provide a coronary artery segmentation method based on continuous frame sequences of DSA images to address the above technical problems.

[0006] The present application provides a coronary artery segmentation method based on a continuous frame sequence of DSA images, including:

[0007] Obtaining a training data set, the training data set comprising an input data set and an output data set, the input data set being the first continuous frame of a sequence length N in a sequence of continuous frames of coronary DSA images, and the output data set being an intermediate frame pre-labeled with a target vessel and located in the middle of the first continuous frame;

[0008] Using the training data set to train a deep learning network model, the deep learning network model includes a feature fusion unit, an encoder, and a decoder, the feature fusion unit performs feature fusion on the first continuous frames to obtain a three-dimensional image, the three-dimensional image sequentially passes through the encoder and the decoder, and the output of the decoder is connected to the intermediate frame;

[0009] The training is continued until a trained deep learning network model is obtained. The trained deep learning network model is used to detect the input coronary DSA image continuous frame sequence and output the segmentation prediction map of the target blood vessel.

[0010] Optionally, the feature fusion unit includes a first feature fusion unit and a second feature fusion unit, and the feature fusion unit performs feature fusion on the first continuous frames to obtain a three-dimensional image, specifically including:

[0011] Dividing the first continuous frames into (N-M+1) second continuous frames with a sequence length of M in time sequence;

[0012] The first feature fusion unit performs feature fusion on each second continuous frame to obtain (N-M+1) fused frames;

[0013] The second feature fusion unit performs feature fusion on each of the fused frames to obtain the three-dimensional image.

[0014] Optionally, N is an odd number, and M is an odd number smaller than N.

[0015] Optionally, the first feature fuser uses a bidirectional long short-term memory network, and the second feature fuser uses a deep residual neural network;

[0016] The value of N is five, and the value of M is three.

[0017] Optionally, the encoder includes a shallow feature extractor at a relatively shallow layer and a deep feature extractor at a relatively deep layer, the shallow feature extractor uses a deep residual neural network, and the deep feature extractor uses a Transformer model.

[0018] Optionally, the three-dimensional image is sequentially down-sampled by the encoder and up-sampled by the decoder, and the encoder includes, from shallow to deep layers:

[0019] The first feature extractor uses a deep residual neural network;

[0020] The second feature extractor uses a deep residual neural network;

[0021] The third feature extractor uses the Transformer model;

[0022] The fourth feature extractor uses the Transformer model.

[0023] Optionally, the target blood vessel is any one of the following: anterior descending artery, circumflex artery, right coronary artery, posterior descending artery belonging to the right coronary artery, posterolateral artery belonging to the right coronary artery;

[0024] The segmentation prediction image is a binary image of the target blood vessel.

[0025] Optionally, a trained deep learning network model is used to detect a continuous frame sequence of input coronary DSA images and output a segmentation prediction map of the target vessel, specifically including:

[0026] Obtaining a test data set, wherein the test data set is the first continuous frame of a sequence length N in a continuous frame sequence of coronary DSA images;

[0027] The first continuous frames are detected using the trained deep learning network model, and a segmentation prediction map of the intermediate frame located in the middle of the first continuous frames is output.

[0028] The present application also provides a computer device comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the coronary vessel segmentation method based on a continuous frame sequence of DSA images described in the present application.

[0029] The present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the coronary vessel segmentation method based on a continuous frame sequence of DSA images described in the present application are implemented.

[0030] The coronary artery segmentation method based on continuous frame sequences of DSA images in this application has at least the following effects:

[0031] This application inputs multiple consecutive image frames, namely the first consecutive frames, from the DSA image into a deep learning model for training and use, and outputs information about the intermediate frames when used. Compared to the technical solution of using a single image frame as input when training a semantic segmentation model, this improves the problem of large differences in segmentation results between frames due to not considering information between image frames, avoids affecting the application of segmentation results in subsequent processes, improves the segmentation consistency of consecutive image frames, and can meet clinical needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of the flow of a coronary artery segmentation method based on a continuous frame sequence of DSA images in one embodiment of the present application;

[0033] Figure 2 This is a flowchart of a coronary artery segmentation method based on a continuous frame sequence of DSA images in one embodiment of the present application;

[0034] Figure 3 This is a schematic diagram of the framework of the deep learning network model in one embodiment of the present application;

[0035] Figure 4 This is a schematic diagram of an original image of a DSA image provided in one embodiment of the present application;

[0036] Figure 5 A DSA image pre-labeled with a target blood vessel provided in an embodiment of the present application;

[0037] Figure 6 This is a schematic diagram of verifying the output results of the deep learning network model in one embodiment of the present application;

[0038] Figure 7 Inputting an intermediate frame of the first continuous frame into the deep learning network model trained in one embodiment of the present application;

[0039] Figure 8 For Figure 7 Output segmentation prediction map;

[0040] Figure 9 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0042] See also Figure 1 and Figure 2 In one embodiment of the present application, a coronary artery segmentation method based on a continuous frame sequence of DSA images is provided, comprising steps S100 to S300. In particular:

[0043] Step S100, obtaining a training data set, the training data set includes an input data set and an output data set, the input data set is the first continuous frame with a sequence length of N in the continuous frame sequence of coronary DSA images, and the output data set is an intermediate frame pre-marked with a target blood vessel and located in the middle of the first continuous frame.

[0044] The first continuous frame refers to several consecutive image frames in time sequence. A sequence length of N refers to N consecutive image frames. The target vessel is any of the following: LAD, LCX, RCA, RPDA, or RPLV.

[0045] Step S200, using the training data set to train the deep learning network model: the deep learning network model includes a feature fuser, an encoder and a decoder. The feature fuser performs feature fusion on the first continuous frame to obtain a three-dimensional image. The three-dimensional image passes through the encoder and decoder in sequence, and the output of the decoder is connected to the intermediate frame.

[0046] Step S300: Continuously train until a trained deep learning network model is obtained, use the trained deep learning network model to detect a continuous frame sequence of input coronary DSA images, and output a segmentation prediction map of the target blood vessel.

[0047] The segmentation prediction map is specifically a binary map of the target blood vessel. During training, the decoder output is connected to the intermediate frame. After training is completed, the decoder outputs the segmentation prediction map of the intermediate frame.

[0048] In the prior art, there is a method of inputting a single image frame for training and using a deep learning network model. In this method, if the image quality of the single image is not high, it will eventually cause the binary image output by the trained deep learning network model to be distorted. In this embodiment, multiple continuous image frames in the DSA image, namely the first continuous frame, are input into the deep learning model. Compared with the technical solution of using a single image frame as input for model training, this improves the problem of large differences in segmentation results between frames due to failure to consider the information between image frames, avoids affecting the application of segmentation results in subsequent processes, improves the segmentation consistency of continuous image frames, and can adapt to clinical needs. The coronary artery segmentation method based on the continuous frame sequence of DSA images in this embodiment is an automated image segmentation algorithm that can efficiently and accurately complete coronary artery segmentation for coronary DSA images.

[0049] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0050] In step S100, firstly, a DSA image is acquired, and then the DICOM image is annotated, that is, the target blood vessels ( Figure 2 coronary DSA images and data annotation in the process to obtain the output dataset.

[0051] Acquiring DSA images involves performing angiography using a standard catheter through the femoral or radial catheter and digitally recording the coronary angiogram to generate DICOM images. Personal patient information in DICOM format images is desensitized using an anonymization tool to remove personal information from DICOM images.

[0052] Pre-annotation (labeling of DICOM images) was performed as follows: a panel of clinician experts annotated a major vessel region on the angiogram to generate a label mask. Specifically, a semi-automatic edge detection tool was used to generate an initial mask of the major vessel boundary on the end-diastolic image frame and the image frames within the first five frames when the major vessel region was filled, and then manual correction was performed. For each major vessel, the segmentation region was set from the ostium to the distal end, and the target vessels to be labeled were the left anterior descending artery (LAD), the circumflex artery (LCX), and the right coronary artery (RCA). For the right coronary artery (RCA), the distal end of the segmental region was distinguished from the bifurcation between the posterior descending artery (PDA) and the posterolateral artery (PL). The continued vessel ends were labeled separately and distinguished using the RPDA and RPLV labels, respectively. Customized annotation software was used to capture and extract pixel information on the boundary and create label masks for each.

[0053] In step S100, the feature fusion device performs feature fusion on the first continuous frame to obtain a three-dimensional image, which can be understood as the channel reconstruction of the input image frame image ( Figure 2Image frame image channel reconstruction in). Taking N as 5 as an example, the input data (DICOM image) size is n×5×512×512, where n is the batch size of training data at one time and 5 is the number of channels). Therefore, it is necessary to perform channel reconstruction on the original input image frame. Channel reconstruction uses pre-labeled (labeled DICOM images) image frame images as intermediate frames. Assuming that the number of intermediate frames is m, then m-2, m-1, m+1, and m+2 are taken as data for the other four channels, forming the "first continuous frame with a sequence length of N" as a whole, which is stacked in order from left to right. Before channel reconstruction, there are 5 frames of continuous grayscale images, and after channel reconstruction, it is a 5×512×512 nii format numpy array data, that is, the feature fusion device performs feature fusion on the first continuous frame to obtain a three-dimensional image.

[0054] In step S100, for the detailed process of channel reconstruction, the feature fusion unit includes a first feature fusion unit and a second feature fusion unit, and the feature fusion unit performs feature fusion on the first continuous frame to obtain a three-dimensional image, which specifically includes steps S110 to S130. Among them:

[0055] Step S110, dividing the first continuous frames into (N-M+1) second continuous frames with a sequence length of M according to time sequence;

[0056] Step S120: The first feature fusion unit performs feature fusion on each of the second consecutive frames to obtain (N-M+1) fused frames;

[0057] In step S130 , the second feature fusion unit performs feature fusion on each fused frame to obtain a three-dimensional image.

[0058] In step S100, N is an odd number, and M is an odd number smaller than N. For example, the value of N can be 5, and the value of M can be 3. The first feature fusion unit uses a bidirectional long short-term memory network BiLSTM, and the second feature fusion unit uses a deep residual neural network ResNet50.

[0059] See also Figure 3 Specifically, the first continuous frame with a sequence length of 5 is input and divided into three second continuous frames with a sequence length of 3, with the number of intermediate frames being m. The three second continuous frames are the first three frames (m-2, m-1, m), the middle three frames (m-1, m, m+1), and the last three frames (m, m+1, m+2). During the feature fusion process, the first feature fusion device, a bidirectional long short-term memory network (BiLSTM), extracts semantic information between consecutive frames, and obtains (N-M+1) fused frames corresponding to the (N-M+1) second continuous frames, such as 3 fused frames. The second feature fusion device, a deep residual neural network (ResNet50), performs stacking and feature fusion to obtain a 3D image for input into the encoder.

[0060] In conjunction with the above, the first continuous frame is used as input, and the continuous information of different frames in the first continuous frame is used to optimize the input and output training process of the model, thereby avoiding the distortion of the output binary image caused by using a single image frame as the input set. However, if the first continuous frame contains too much continuous information, it will lead to the introduction of image information with poor continuity (compared to the continuity of the intermediate frames) during the input and output process of the model, ultimately causing distortion of the output single image.

[0061] Steps S110 to S130 of this embodiment not only obtain the technical advantage of using the first continuous frame as input, but also avoid the problem of poor continuity by dividing the first continuous frame and fusing the first feature fuser and the second feature fuser in steps.

[0062] In step S200, the encoder includes a shallow feature extractor at a relatively shallow layer and a deep feature extractor at a relatively deep layer. The shallow feature extractor uses a deep residual neural network, and the deep feature extractor uses a Transformer model.

[0063] In this step, the data in the relatively deep layer is larger, while the data in the relatively shallow layer is smaller. The shallow feature extractor uses the deep residual neural network, and the deep feature extractor uses the Transformer model, taking into account both the accuracy of the Transformer model and the efficiency of the deep residual neural network ResNet.

[0064] In step S200, the three-dimensional image is down-sampled by the encoder and up-sampled by the decoder in sequence. The encoder includes, from the shallow layer to the deep layer: a first feature extractor using a deep residual neural network, a second feature extractor using a deep residual neural network, a third feature extractor using a Transformer model, and a fourth feature extractor using a Transformer model.

[0065] See also Figure 3 The deep learning network model trained and used in each embodiment of the present application is the BiLSTM (bidirectional long short-term memory network)-Res (deep residual neural network ResNet)-Trans (Transformer model)-UNet semantic segmentation network model.

[0066] Based on the U-Net semantic segmentation architecture, the BiLSTM-Res-Trans-UNet semantic segmentation network model modified the encoding strategy of the encoder module to innovate the model and optimize the results. It constructed a binary classification semantic segmentation model with an encoding and decoding structure, and completed the automated segmentation of coronary DSA images, ensuring that the accuracy and consistency of segmentation between frames meet clinical needs.

[0067] The model uses the three-dimensional image output by the feature fusion device as the original input. The encoder stage completes downsampling through four 3×3 convolutions with a step size of 2 to achieve the purpose of deeply extracting image features (from the relatively shallow layer to the relatively deep layer, the first feature extractor in the first layer, the second feature extractor in the second layer, the third feature extractor in the third layer, and the fourth feature extractor in the fourth layer).

[0068] The decoder stage uses deconvolution to perform upsampling, gradually restoring the feature map size to 512×512, ultimately outputting a 512×512 segmentation prediction map (a binary image of the target vessel). The encoder adds long-range skip connections with a Transformer structure (Transformer model) to the last two layers of downsampling (the third and fourth layers) to extract deep semantic features. These features are then fused in the subsequent decoder upsampling stage. While restoring feature map details, global information is further extracted to improve the model's overall performance.

[0069] The above-mentioned BiLSTM-Res-Trans-UNet semantic segmentation network model is improved based on the traditional UNet framework, and the model is described in detail as follows. The BiLSTM-Res-Trans-UNet model includes an encoder, a decoder, a feature fusion device (a BiLSTM that extracts semantic information between frames), a long-range skip connection with a Transformer structure, and a ResNet feature extractor with a skip connection. The selected model of the deep residual neural network in each embodiment can be ResNet50.

[0070] The encoder consists of four encoding blocks and four downsampling operations. Each encoding block consists of two 3×3 convolutional blocks, each containing a 3×3 convolutional layer followed by a batch normalization layer and a ReLU activation layer. Downsampling is achieved using convolutions with a stride of 2.

[0071] The decoder consists of four decoding blocks and four upsampling operations. Each decoding block consists of two 3×3 convolutional blocks, each containing a 3×3 convolutional layer followed by a batch normalization layer and a ReLU activation layer. Upsampling is achieved using deconvolution with a stride of 2.

[0072] The long-range skip connection consists of two paths (the third and fourth layers), with a Transformer structure inserted into each path with the same number of repetitions as a feature extractor for deep semantics. The skip connection also consists of two paths (the first and second layers), with a ResNet inserted into each path as a feature extractor for low-level semantics. The Transformer takes the feature encoding and position encoding as its overall input and passes through two core algorithm blocks in sequence. The first algorithm block consists of a multi-head attention mechanism and a fusion normalization layer, while the second algorithm block consists of a multi-layer perceptron and a fusion normalization layer. Both blocks introduce short-range skip connections between the input and output of the modules.

[0073] Step S300 specifically includes steps S310 to S320, wherein:

[0074] Step S310, obtaining a test data set, the test data set being the first continuous frame with a sequence length of N in a continuous frame sequence of coronary DSA images;

[0075] Step S320: Detect the first continuous frames using the trained deep learning network model, and output a segmentation prediction map of the middle frame located in the middle of the first continuous frames.

[0076] In step S300, the trained deep learning network model is used to automatically segment the first continuous frame of the DSA image, and the vascular area is marked in white. The accurate and consistent segmentation results can be used for subsequent three-dimensional reconstruction to facilitate hemodynamic analysis.

[0077] The DSA coronary artery region extraction method based on a continuous image frame sequence in each embodiment of the present application has the following advantages: (1) When the deep learning network model is used, the output of the segmentation prediction map of the target blood vessel is highly automated, and no human intervention is required to add the starting point position or modify the boundary position. (2) The first continuous frame input method is used, so that the segmentation result consistency of the deep learning network model is higher than that of the general deep learning algorithm. (3) It can take into account the accuracy of the Transformer model and the efficiency of the deep residual neural network (resnet), improve the diagnosis speed of patients, and has high clinical application value.

[0078] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 9As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a coronary vessel segmentation method based on a continuous frame sequence of DSA images is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0079] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0080] Step S100, obtaining a training data set, the training data set comprising an input data set and an output data set, the input data set being the first continuous frame of a sequence length N in a continuous frame sequence of coronary DSA images, and the output data set being an intermediate frame pre-labeled with a target vessel and located in the middle of the first continuous frame;

[0081] Step S200: training a deep learning network model using a training data set. The deep learning network model includes a feature fusion unit, an encoder, and a decoder. The feature fusion unit fuses features of the first continuous frame to obtain a three-dimensional image. The three-dimensional image passes through the encoder and decoder in sequence. The output of the decoder is connected to the intermediate frame.

[0082] Step S300: Continuously train until a trained deep learning network model is obtained, use the trained deep learning network model to detect a continuous frame sequence of input coronary DSA images, and output a segmentation prediction map of the target blood vessel.

[0083] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0084] Step S100, obtaining a training data set, the training data set comprising an input data set and an output data set, the input data set being the first continuous frame of a sequence length N in a continuous frame sequence of coronary DSA images, and the output data set being an intermediate frame pre-labeled with a target vessel and located in the middle of the first continuous frame;

[0085] Step S200: training a deep learning network model using a training data set. The deep learning network model includes a feature fusion unit, an encoder, and a decoder. The feature fusion unit fuses features of the first continuous frame to obtain a three-dimensional image. The three-dimensional image passes through the encoder and decoder in sequence. The output of the decoder is connected to the intermediate frame.

[0086] Step S300: Continuously train until a trained deep learning network model is obtained, use the trained deep learning network model to detect a continuous frame sequence of input coronary DSA images, and output a segmentation prediction map of the target blood vessel.

[0087] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0088] The technical features of the above embodiments may be combined in any manner. To simplify the description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there are no conflicts in the combination of these technical features, they should be considered to be within the scope of this specification. When technical features in different embodiments are reflected in the same figure, it can be regarded as that figure also discloses the combination examples of the various embodiments involved.

[0089] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A coronary artery segmentation method based on a continuous frame sequence of DSA images, characterized in that: include: Obtaining a training data set, the training data set comprising an input data set and an output data set, the input data set being the first continuous frame of a sequence length N in a continuous frame sequence of coronary DSA images, and the output data set being an intermediate frame in the middle of the first continuous frame, the intermediate frame being pre-annotated with a target vessel on the DICOM image; The deep learning network model is trained using the training data set: the deep learning network model includes a first feature fusion device and a second feature fusion device, an encoder, and a decoder, and divides the first continuous frame into (N-M+1) second continuous frames with a sequence length of M in time sequence; the first feature fusion device performs feature fusion on each second continuous frame to obtain (N-M+1) fused frames; the second feature fusion device performs feature fusion on each fused frame to obtain a three-dimensional image; the three-dimensional image passes through the encoder and decoder in sequence, and the output of the decoder is connected to the intermediate frame; The training is continued until a trained deep learning network model is obtained. The trained deep learning network model is used to detect the input coronary DSA image continuous frame sequence and output the segmentation prediction map of the target blood vessel.

2. The coronary artery segmentation method based on continuous frame sequences of DSA images according to claim 1, characterized in that: N is an odd number, and M is an odd number smaller than N.

3. The coronary artery segmentation method based on continuous frame sequences of DSA images according to claim 2, characterized in that: The first feature fusion device uses a bidirectional long short-term memory network, and the second feature fusion device uses a deep residual neural network; The value of N is five, and the value of M is three.

4. The coronary artery segmentation method based on continuous frame sequences of DSA images according to claim 1, characterized in that: The encoder includes a shallow feature extractor at a relatively shallow layer and a deep feature extractor at a relatively deep layer. The shallow feature extractor uses a deep residual neural network, and the deep feature extractor uses a Transformer model.

5. The method for coronary vessel segmentation based on continuous frame sequences of DSA images according to claim 4, characterized in that: The three-dimensional image is sequentially down-sampled by the encoder and up-sampled by the decoder, and the encoder includes, from shallow to deep layers: The first feature extractor uses a deep residual neural network; The second feature extractor uses a deep residual neural network; The third feature extractor uses the Transformer model; The fourth feature extractor uses the Transformer model.

6. The method for coronary artery segmentation based on continuous frame sequences of DSA images according to claim 1, characterized in that: The target blood vessel is any one of the following: the anterior descending artery, the circumflex artery, the right coronary artery, the posterior descending artery belonging to the right coronary artery, and the posterolateral artery belonging to the right coronary artery; The segmentation prediction image is a binary image of the target blood vessel.

7. The method for segmenting coronary arteries based on a continuous frame sequence of DSA images according to any one of claims 1 to 6, characterized in that: The trained deep learning network model is used to detect the continuous frame sequence of the input coronary DSA image and output the segmentation prediction map of the target blood vessel, including: Obtaining a test data set, wherein the test data set is the first continuous frame of a sequence length N in a continuous frame sequence of coronary DSA images; The first continuous frames are detected using the trained deep learning network model, and a segmentation prediction map of the intermediate frame located in the middle of the first continuous frames is output.

8. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that The processor executes the computer program to implement the steps of the coronary vessel segmentation method based on a continuous frame sequence of DSA images according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the coronary vessel segmentation method based on a continuous frame sequence of DSA images according to any one of claims 1 to 6 are implemented.

Images