Information processing method and device for CTA image segmentation, equipment and medium

Abstract

The application discloses an information processing method and device for CTA image segmentation, equipment and medium, relates to the technical field of three-dimensional CT image segmentation, and comprises the following steps: inputting historical head and neck CTA image data as first training data into an artery and vertebra segmentation model, and obtaining an artery mask as second training data; a plaque segmentation model carrying a blood vessel guided attention mechanism is built, iterative training is carried out in combination with the two types of training data, a reasonable distribution range of plaques is demarcated according to the dilated artery mask, a spatial penalty loss is calculated for the prediction results outside the range, and the model parameters are updated until the training is completed; inputting a head and neck CTA image to be processed into the artery vertebra segmentation model to obtain a corresponding segmentation graph, inputting the artery mask in the graph and the original image into the plaque segmentation model to output a plaque segmentation graph, and fusing the two types of segmentation graphs to obtain a complete segmentation result containing artery, vertebra and plaque information. The head and neck CTA image segmentation efficiency is improved, and soft plaques are effectively segmented.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional CT image segmentation technology, and in particular to information processing methods, apparatus, equipment and media for CTA image segmentation. Background Technology

[0002] Head and neck CTA (Computed Tomography Angiography) images can clearly show the course of head and neck vessels, vascular lesions, and plaque-related morphological information. It can accurately segment key structures within the image, providing anatomical references for clinical diagnosis, surgical planning, and interventional treatment, and helping to improve the level of lesion identification and surgical safety.

[0003] Traditional manual segmentation relies on manual operation, which consumes a lot of manpower and time. The segmentation effect is easily affected by the operator's own condition and professional experience, resulting in poor consistency. At the same time, the processing speed is slow, which is difficult to meet the actual needs of rapid processing of massive clinical images, thus hindering the implementation and use of related automated diagnostic technologies.

[0004] While existing intelligent segmentation algorithms can segment some vascular plaques, they still have significant shortcomings. On the one hand, due to similar tissue density and imaging artifacts, additional processing steps are required before segmentation, making the overall process complex. On the other hand, it is difficult to distinguish soft plaques from surrounding soft tissues, making it impossible to segment soft plaques and easily missing high-risk lesions, which is not conducive to accurately assessing disease risk and formulating treatment plans.

[0005] In summary, improving the efficiency and accuracy of head and neck CTA image segmentation and effectively segmenting soft plaques are problems that need to be solved in this field. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide an information processing method, apparatus, device, and medium for CTA image segmentation, improving the efficiency and accuracy of head and neck CTA image segmentation and effectively segmenting soft plaques. The specific solution is as follows: In a first aspect, this application discloses an information processing method for CTA image segmentation, applied to a computer device, comprising: The collected historical head and neck CTA image data was selected as the first training data; The first training data is input into the first segmentation model for segmenting arteries and vertebrae in CTA image data, and the artery mask in the output artery and vertebral segmentation map is determined as the second training data. A current second segmentation model for segmenting plaques in CTA image data is constructed based on a blood vessel-guided attention mechanism. The current second segmentation model is trained using the first training data and the second training data. A penalty is applied to the predicted regions in the plaque segmentation prediction map output by the current second segmentation model that are outside the allowed distribution range of the plaque, so as to obtain the spatial penalty loss of the current second segmentation model. The allowed distribution range of the plaque is the range obtained by dilating the second training data. The parameters of the current second segmentation model are updated based on the spatial penalty loss to obtain a new current second segmentation model. Then, the process jumps back to the step of training the current second segmentation model using the first training data and the second training data until the preset training stop condition is met, and the new current second segmentation model is determined as the target second segmentation model. The current head and neck CTA image data is input into the first target segmentation model to output a target artery vertebral segmentation map. The artery mask in the target artery vertebral segmentation map and the current head and neck CTA image data are input into the second target segmentation model to output a target plaque segmentation map. The target artery vertebral segmentation map and the target plaque segmentation map are fused to generate a target segmentation map that includes the artery, vertebra, and plaque.

[0007] Optionally, the preset training stopping conditions include preset pre-training stopping conditions and preset fine-tuning training stopping conditions; training the current second segmentation model using the first training data and the second training data includes: The first training data is divided into a first sample dataset and a second sample dataset; wherein the annotation accuracy of the first sample dataset is lower than that of the second sample dataset and the number of samples in the first sample dataset is less than that in the second sample dataset; The current second segmentation model is pre-trained using the first sample dataset and the second training data, and then the process jumps to the step of penalizing the predicted regions located outside the allowed distribution range of patches in the patch segmentation prediction map output by the current second segmentation model, until the preset pre-training stop condition is met. The second sample dataset and the second training data are used to train the current second segmentation model using a five-fold cross-validation method. Then, the process jumps to the step of penalizing the predicted regions located outside the allowed distribution range of patches in the patch segmentation prediction map output by the current second segmentation model, until the preset fine-tuning training stop condition is met.

[0008] Optionally, determining the acquired historical head and neck CTA image data as the first training data includes: An adaptive data preprocessing operation is performed based on the task type corresponding to the historical raw head and neck CTA image data to obtain preprocessed image data. The image window width and window level parameters of the preprocessed image data are adaptively configured using a boundary contrast maximization method to obtain the first training data.

[0009] Optionally, the step of adaptively configuring the image window width and window level parameters of the preprocessed image data using a boundary contrast maximization method to obtain the first training data includes: Based on the initial mask label corresponding to the preprocessed image data, the original region of interest containing the target tissue is extracted from the preprocessed image data; wherein, the target tissue is a blood vessel or a vertebra. The original region of interest is morphologically dilated, and the dilated region is subtracted from the original region of interest to obtain a background neighborhood voxel set; wherein, the background neighborhood voxel set is a set of voxels that surround the target tissue and are in close contact with the boundary of the target tissue; A two-dimensional search matrix is ​​set based on a preset window width and window level range. For each combination of window width and window level in the two-dimensional search matrix, the first gray-scale distribution information after mapping the voxel set in the original region of interest and the second gray-scale distribution information after mapping the voxel set in the background neighborhood are determined respectively. The evaluation index function constructed based on Fisher's discrimination criterion, the first gray-scale distribution information, and the second gray-scale distribution information are used to quantify the class separation degree between the target organization and the target organization's neighborhood background. Iterate through each group of window width and window level combinations, and determine the window width and window level combination that maximizes the class separation as the global optimal parameter for the segmentation task. Use the global optimal parameter to perform image grayscale mapping correction on the preprocessed image data to obtain the first training data.

[0010] Optionally, the current second segmentation model includes a backbone network, an encoder, and a decoder; obtaining the patch segmentation prediction map output by the current second segmentation model includes: The backbone network is used to extract features from the first training data to obtain original image features, and the encoder is used to downsample the original image features to obtain multi-scale image features. The current second segmentation model is used to downsample the second training data to obtain multi-scale mask features. The decoder is used to normalize the multi-scale mask features to obtain spatial attention weights, and the spatial attention weights are multiplied by voxel by voxel with the multi-scale image features to output a patch segmentation prediction map. Accordingly, updating the parameters of the current second segmentation model based on the spatial penalty loss includes: The similarity coefficient loss and cross-entropy loss of the current second segmentation model are obtained based on the patch segmentation prediction map and the pre-labeled real labels in the first training data. The total loss of the current second segmentation model is obtained based on the spatial penalty loss, the similarity coefficient loss, and the cross-entropy loss, and the parameters of the current second segmentation model are updated based on the total loss.

[0011] Optionally, fusing the target artery vertebral segmentation map and the target plaque segmentation map includes: The patch mask in the target patch segmentation map is binarized and subjected to three-dimensional connected component analysis. The patch in the target patch segmentation map is segmented into multiple first independent patch regions. Each first independent patch region is three-dimensionally dilated to obtain the dilated patch region. The dilated patch regions that do not spatially intersect with the arterial vessel mask in the target patch segmentation map and the target arterial vertebral segmentation map are removed to obtain the patch segmentation map after removal. The image grayscale value corresponding to the mask identifier of each voxel in the removed patch segmentation image is compared with a preset grayscale threshold range. The voxel mask with the image grayscale value exceeding the preset grayscale threshold range is determined as a prediction anomalous voxel with over-prediction. The prediction anomalous voxels in the removed patch segmentation image are then filtered to obtain the corrected patch segmentation image. Three-dimensional connected component analysis is performed on the corrected patch segmentation map to obtain multiple second independent patch regions. The second independent patch region containing the most voxels is determined as the target independent patch region. The patch size and patch volume in the target independent patch region are determined based on the image physical spacing. The corrected plaque segmentation map is fused with the target arterial vertebral segmentation map.

[0012] Optionally, after fusing the target artery vertebral segmentation map and the target plaque segmentation map to generate a target segmentation map including the artery, vertebra, and plaque, the method further includes: Based on the spatial affine matrix of the current head and neck CTA image data, the coordinate system of the target segmentation map is converted into the RAS visualization coordinate system to obtain the mask image after coordinate adaptation. The moving cube algorithm is used to extract the surface mesh from the coordinate-adapted mask image, generate vertex and face data based on the surface mesh, and construct an initial three-dimensional geometric model based on the vertex and face data. The initial three-dimensional geometric model was smoothed and denoised using the VTK window sine polynomial smoothing algorithm to obtain a smoothed mesh model. Connectivity analysis and noise removal are performed on the smoothed mesh model to obtain the target 3D visualization model.

[0013] Secondly, this application discloses an information processing apparatus for CTA image segmentation, applied to a computer device, comprising: The first data determination module is used to determine the collected historical head and neck CTA image data as the first training data; The second data determination module is used to input the first training data into the first target segmentation model for segmenting arteries and vertebrae in CTA image data, and to determine the artery mask in the output artery and vertebral segmentation map as the second training data. The loss acquisition module is used to construct a current second segmentation model for segmenting plaques in CTA image data based on a blood vessel-guided attention mechanism, and to train the current second segmentation model using the first training data and the second training data. A penalty is applied to the predicted regions located outside the allowed distribution range of the plaques in the plaque segmentation prediction map output by the current second segmentation model to obtain the spatial penalty loss of the current second segmentation model; wherein, the allowed distribution range of the plaques is the range obtained by dilating the second training data. The target model acquisition module is used to update the parameters of the current second segmentation model based on the spatial penalty loss to obtain a new current second segmentation model, and then jump back to the step of training the current second segmentation model using the first training data and the second training data until the preset training stop condition is met, and the new current second segmentation model is determined as the target second segmentation model. The image segmentation module is used to input the current head and neck CTA image data into the target first segmentation model to output the target artery vertebral segmentation map, and to input the artery mask in the target artery vertebral segmentation map and the current head and neck CTA image data into the target second segmentation model to output the target plaque segmentation map. The image fusion module is used to fuse the target artery vertebral segmentation map and the target plaque segmentation map to generate a target segmentation map including the artery, vertebra, and plaque.

[0014] Thirdly, this application discloses an electronic device, comprising: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the aforementioned disclosed information processing method for CTA image segmentation.

[0015] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed information processing method for CTA image segmentation.

[0016] The beneficial effects of this application are as follows: This application uses historical head and neck CTA image data as the first training data; it inputs the first training data into a first segmentation model for segmenting arteries and vertebrae in CTA image data, and uses the arterial mask in the output arterial-vertebral segmentation map as the second training data; it constructs a current second segmentation model for segmenting plaques in CTA image data based on a vascular-guided attention mechanism, and trains the current second segmentation model using the first and second training data. A penalty is applied to the predicted regions outside the allowed distribution range of plaques in the plaque segmentation prediction map output by the current second segmentation model to obtain the spatial penalty loss of the current second segmentation model; wherein, the allowed distribution range of plaques is obtained by dilating the second training data. The scope is determined; the parameters of the current second segmentation model are updated based on the spatial penalty loss to obtain a new current second segmentation model, and the process jumps back to the step of training the current second segmentation model using the first training data and the second training data until a preset training stop condition is met, and the new current second segmentation model is determined as the target second segmentation model; the current head and neck CTA image data is input into the target first segmentation model to output a target artery and vertebral segmentation map, and the artery mask in the target artery and vertebral segmentation map and the current head and neck CTA image data are input into the target second segmentation model to output a target plaque segmentation map; the target artery and vertebral segmentation map and the target plaque segmentation map are fused to generate a target segmentation map including arteries, vertebrae, and plaques.Therefore, this application first uses historical head and neck CTA image data as the basic training sample, and can accumulate feature learning basis based on real clinical image data. After feeding this type of data into a first segmentation model specifically for identifying arteries and vertebrae, the arterial mask is extracted as a matching training sample. The reference boundary for plaque recognition can be defined based on the actual distribution range of blood vessels. On this basis, a plaque segmentation model with a blood vessel-guided attention structure is built. The model is trained simultaneously using two types of training data. During the training process, the region after the arterial mask is expanded is used to define the reasonable range of plaque existence. For prediction results that exceed this range, the loss value is calculated and the model output is constrained. This can reduce the situation where the model identifies plaques without basis. The internal parameters of the model are continuously corrected based on the calculated loss, and the model is iteratively optimized repeatedly. The model's recognition capability can be gradually improved until the end of training, gradually increasing the accuracy of the model in distinguishing the location and morphology of plaques. After training, a mature and usable plaque segmentation model is obtained. In practical applications, the first segmentation model is used to divide the arterial and vertebral regions. Then, the arterial mask obtained from the segmentation is combined with the original image and input into the plaque segmentation model to complete plaque detection. The segmentation results of different tissues are obtained by relying on the two segmentation steps. Finally, the segmentation images of multiple tissues are integrated to form a complete segmentation pattern. It can accurately distinguish the three types of target tissues in head and neck CTA images: arteries, vertebrae, and plaques. This reduces the recognition bias caused by a single model recognizing multiple types of tissues at the same time, conforms to the actual distribution pattern of plaques in head and neck blood vessels, and effectively improves the accuracy and reliability of the segmentation results of various lesions and tissues in head and neck CTA images. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart of an information processing method for CTA image segmentation disclosed in this application; Figure 2 This is a schematic diagram of a specific two-stage training method disclosed in this application; Figure 3 This is a schematic diagram of a specific patch segmentation model disclosed in this application; Figure 4 This is a specific example of the false positive rejection effect disclosed in this application; Figure 5 This is a schematic diagram illustrating a specific visualization model construction disclosed in this application; Figure 6 This is a specific clinical assessment diagram disclosed in this application; Figure 7 This is a schematic diagram of an information processing device for CTA image segmentation disclosed in this application; Figure 8 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] CTA can more intuitively present the course of blood vessels in the head and neck region, the degree of luminal stenosis, and the location, size, and morphological characteristics of plaques. Precise segmentation of key CTA structures can provide three-dimensional anatomical evidence for clinical diagnosis, surgical planning, and interventional treatment, improving lesion identification accuracy and surgical safety. However, current head and neck CTA segmentation techniques still face two major bottlenecks, as follows:

[0021] Traditional manual modeling and segmentation methods rely on manual operation by clinicians or imaging technicians. This not only consumes a significant amount of time and energy but is also easily affected by factors such as the operator's subjective experience and fatigue level, leading to substantial individual differences in segmentation results and making it difficult to guarantee the consistency and stability of segmentation accuracy. Furthermore, the manual segmentation process is inefficient and cannot meet the needs of rapid processing of large amounts of image data in clinical settings, severely hindering the development and application of automated diagnostic technologies for head and neck vascular diseases.

[0022] Existing AI diagnostic software utilizes 3D convolutional neural networks for artery and plaque segmentation and modeling. However, this approach still has two limitations: First, the processing is cumbersome. Skeleton and arteries with contrast agents have similar densities, making them prone to occlusion and artifacts. Skeleton must be segmented and removed separately before accurate extraction of vascular structures can be achieved, significantly increasing the complexity of the segmentation process. Second, it lacks segmentation capabilities for soft plaques. Soft plaques are mainly composed of a lipid core, fibrous cap, and inflammatory components, with CT values ​​typically between 20-80 HU. Their grayscale distribution highly overlaps with surrounding tissues such as muscle (40-60 HU), soft tissue (20-60 HU), and fat (-120 to -40 HU). Furthermore, soft plaque boundaries are often indistinct, lacking clear morphological features, making it difficult for the model to accurately distinguish the target from the background. Due to differences in pathological structure, soft plaques pose a significantly higher clinical risk than the relatively stable hard plaques. Current methods can only segment hard plaques but not soft plaques, which will directly miss high-risk vulnerable plaques and underestimate the risk of stroke, seriously affecting clinical risk assessment, treatment decisions and follow-up efficacy judgment.

[0023] Therefore, this application provides a CTA image segmentation scheme to improve the efficiency and accuracy of head and neck CTA image segmentation and effectively segment soft plaques.

[0024] See Figure 1 As shown in the embodiment of this application, an information processing method for CTA image segmentation is disclosed, applied to a computer device, including: Step S11: The collected historical head and neck CTA image data is determined as the first training data.

[0025] In this embodiment, determining the collected historical head and neck CTA image data as the first training data includes: performing an adaptive data preprocessing operation according to the task type corresponding to the historical raw head and neck CTA image data to obtain preprocessed image data; and adaptively configuring the image window width and window level parameters of the preprocessed image data using a boundary contrast maximization method to obtain the first training data.

[0026] First, historical raw head and neck CTA image data were collected, specifically including: 1) Collecting head and neck CTA image data from patients with clinical history of visits, whose clinical diagnoses included carotid artery stenosis or occlusion, atherosclerosis, or carotid plaque. The image data came from a 3.0T CT device, with the scanning range covering the entire head and neck (from the top of the skull to the upper edge of the clavicle). The scanning parameters were uniformly set to slice thicknesses of 0.5mm, 0.625mm, and 1.0mm; 2) All CTA image data were stored in DICOM format, containing basic patient information (anonymized to remove privacy information such as name and ID number), scanning parameters, CT values ​​(Henry's units, HU), and other metadata. In subsequent processing, the DICOM format was converted to NifTI format to facilitate model reading and processing, while preserving the spatial location information and CT value accuracy of the original images during the conversion process; 3) The collected image data were screened to remove images with motion artifacts, metal artifacts, incomplete scans, and insufficient contrast. Finally, 387 qualified head and neck CTA image data were retained as subsequent training data.

[0027] Secondly, adaptive data preprocessing was performed based on the four segmentation task types corresponding to the historical raw head and neck CTA image data: arteries, vertebrae, and soft and hard plaques, to obtain preprocessed image data. Specifically, the converted NifTI format CTA images were normalized, mapping the CT value (HU) to the [0, 255] interval to eliminate grayscale differences caused by different devices and scanning parameters; subsequently, the images were resampled, and all images were uniformly adjusted to a voxel spacing of 1×1×1mm to ensure consistency in subsequent model training and inference.

[0028] Next, the image window width and window level parameters of the preprocessed image data are adaptively configured using a boundary contrast maximization method to obtain the first training data. First, the original CTA image data to be processed and its corresponding mask labels are obtained. The region of interest (ROI) of the target is extracted based on the mask labels, and the original mask region is subtracted after morphological dilation of the target ROI to construct a background neighborhood voxel set. Second, a two-dimensional search matrix is ​​set to map the Heinz units (HU) in the original image to a unified grayscale scale space. Subsequently, an adaptive traversal search is performed in the search matrix. For each combination of window width and window level to be tested, the gray average and gray variance of the target voxel set after mapping, as well as the gray average and gray variance of the background neighborhood voxel set after mapping, are calculated respectively. Based on this, the Fisher discrimination criterion is used to construct an evaluation index. Finally, by comparing the Fisher discrimination ratio values ​​under different parameter combinations, the window width and window level corresponding to maximizing the ratio are selected as the globally optimal parameters.

[0029] In this embodiment, the adaptive configuration of the image window width and window level parameters of the preprocessed image data using the boundary contrast maximization method to obtain the first training data includes: extracting the original region of interest (ROI) containing the target tissue from the preprocessed image data based on the initial mask label corresponding to the preprocessed image data; wherein the target tissue is a blood vessel or vertebra; performing morphological dilation processing on the original ROI, and subtracting the original ROI from the dilated region to obtain a background neighborhood voxel set; wherein the background neighborhood voxel set is a set of voxels that wrap around the target tissue and are in close contact with the boundary of the target tissue; and setting a two-dimensional search matrix based on a preset window width and window level range. For each group of window width and window level combinations in the two-dimensional search matrix, the first gray-level distribution information after mapping the voxel set in the original region of interest and the second gray-level distribution information after mapping the voxel set in the background neighborhood are determined respectively. The evaluation index function constructed based on Fisher's discrimination criterion, the first gray-level distribution information, and the second gray-level distribution information are used to quantify the class separation degree between the target tissue and the target tissue's neighborhood background. Each group of window width and window level combinations is traversed, and the window width and window level combination that maximizes the class separation degree is determined as the global optimal parameter for the segmentation task. The global optimal parameter is used to perform image gray-level mapping correction on the preprocessed image data to obtain the first training data.

[0030] First, obtain the initial mask label corresponding to the preprocessed image data, and extract the region of interest (ROI) of the target based on the initial mask label, that is, extract the original region of interest containing the target tissue (blood vessel, vertebra) in the preprocessed image data.

[0031] By performing morphological dilation on the target region of interest and then subtracting the original mask region (i.e., performing morphological dilation on the original region of interest with a 3×3×3 cubic structuring element as the dilation kernel and dilution times of 2), and then subtracting the original region of interest from the dilated region, a set of background neighborhood voxels that wraps around the target and is in close contact with its boundary is constructed.

[0032] Secondly, a two-dimensional search matrix is ​​set based on preset window width and window level ranges (vascular window: window width 600-800HU, window level 120-160HU; bone window: window width 1000-1500HU, window level 250-350HU; soft tissue window: window width 300-500HU, window level 40-60HU). The search range for window width (WW) is 200-1500HU with a step size of 50HU; the search range for window level (WL) is 0-500HU with a step size of 20HU, ensuring coverage of all possible optimal parameter ranges. This involves setting a two-dimensional search matrix composed of different window width and window level values, and using a linear piecewise function for mapping to a unified grayscale scale space. For voxel values ​​higher or lower than the display threshold during the mapping process, saturation truncation is performed. The mapping function is defined as follows: ; in, The grayscale value of the voxel obtained after mapping; The original input CT value; The window width represents the dynamic range of grayscale transformation; The window level indicates the center position of the grayscale transformation.

[0033] Subsequently, an adaptive traversal search is performed within the search matrix, and for each combination of window width and window level to be tested, the average grayscale value after mapping the target voxel set is calculated. ) and gray variance ( ), and the background neighborhood voxel set after mapping ( ) and gray variance ( This refers to the first gray-level distribution information after mapping the voxel set in the original region of interest and the second gray-level distribution information after mapping the voxel set in the background neighborhood. Based on this, an evaluation index is constructed using the Fisher discriminant criterion, which quantifies the class separation degree between the target and its neighborhood background in the current window by calculating the ratio of the square of the difference between the average gray values ​​of the two sets to the sum of the variances of the gray values ​​of the two sets. ; in, Fisher's discrimination score, i.e., class separation degree; and The parameters are as follows: The mean and variance of the mapped gray levels; and These represent the mean and variance of the grayscale values ​​after mapping the boundary background region, respectively. It is a very small constant used to prevent the denominator from being zero.

[0034] Finally, the window width and window level combinations are traversed. By comparing the Fisher discrimination ratio values ​​under different parameter combinations, the window width and window level that maximize the ratio are selected as the global optimal parameters. The global optimal parameters are then used to perform image grayscale mapping correction on the preprocessed image data to obtain the first training data. This enables the adaptive enhancement of the edge contrast between the target and surrounding tissues by utilizing the data's own distribution characteristics, thus assisting in subsequent segmentation.

[0035] Step S12: Input the first training data into the first segmentation model for segmenting arteries and vertebrae in CTA image data, and determine the artery mask in the output artery and vertebral segmentation map as the second training data.

[0036] The first segmentation model uses the nnU-Net v2 model as its basic segmentation framework. This framework has advantages such as adaptive network structure adjustment, automatic data preprocessing, and multi-scale feature fusion, making it suitable for medical image segmentation tasks. The model input is preprocessed CTA images with optimal window width and level (i.e., the first training data), and the output is a binary segmentation mask for arteries and vertebrae (background is 0, target is 1).

[0037] A primary segmentation model for segmenting arteries and vertebrae in CTA image data is constructed. This model is trained using primary training data to obtain the target primary segmentation model. The primary training data is divided into a primary sample dataset and a secondary sample dataset. The annotation accuracy of the primary sample dataset is lower than that of the secondary sample dataset, and the number of primary sample datasets is less than that of the secondary sample dataset. Next, the first stage of basic training is carried out. The primary sample dataset is input into the primary segmentation model, and the loss is calculated and the model parameters are iteratively updated with the ground truth annotations of arteries and vertebrae as references, so that the model initially has the ability to identify the two types of tissues. Then, the second stage of fine-tuning training is carried out. The primary segmentation model after the first stage of training is fine-tuned using the secondary sample dataset. Considering that blood vessels and bones have similar densities and are prone to artifact interference, the model's feature extraction ability is continuously optimized by combining samples containing occlusion and artifacts. The process is iterated repeatedly until the model's segmentation accuracy and stability on the validation set reach the preset standard. Finally, the target primary segmentation model that can accurately segment arteries and vertebrae in CTA images is obtained.

[0038] Step S13: Construct a current second segmentation model for segmenting plaques in CTA image data based on the blood vessel-guided attention mechanism, and train the current second segmentation model using the first training data and the second training data. Apply a penalty to the predicted regions located outside the allowed distribution range of plaques in the plaque segmentation prediction map output by the current second segmentation model to obtain the spatial penalty loss of the current second segmentation model; wherein, the allowed distribution range of plaques is the range obtained by dilution processing of the second training data.

[0039] In this embodiment, the preset training stop condition includes a preset pre-training stop condition and a preset fine-tuning training stop condition. Training the current second segmentation model using the first training data and the second training data includes: dividing the first training data into a first sample dataset and a second sample dataset; wherein the annotation accuracy of the first sample dataset is lower than that of the second sample dataset, and the number of elements in the first sample dataset is less than that in the second sample dataset; pre-training the current second segmentation model using the first sample dataset and the second training data, and then proceeding to the step of penalizing predicted regions located outside the allowed distribution range of patches in the patch segmentation prediction map output by the current second segmentation model, until the preset pre-training stop condition is met; training the current second segmentation model using the second sample dataset and the second training data according to a five-fold cross-validation method, and then proceeding to the step of penalizing predicted regions located outside the allowed distribution range of patches in the patch segmentation prediction map output by the current second segmentation model, until the preset fine-tuning training stop condition is met.

[0040] like Figure 2 As shown, the historical head and neck CTA image data, which serves as the first training data, is first split into a first sample dataset and a second sample dataset. The first sample dataset has a smaller total number of samples and relatively lower annotation accuracy, while the second sample dataset has a larger number of samples and higher annotation accuracy. These two datasets are used in conjunction with the second training data for different stages of training the second segmentation model. Specifically, 387 qualified CTA images are randomly divided into two batches: batch 1 contains 100 images, and batch 2 contains 287 images. Initial data annotation is performed by three radiologists with three or more years of experience, and cross-checking and correction are conducted. For difficult cases, annotation is completed by one neurosurgeon and one radiologist (both with 15 or more years of experience) through consensus discussion. All annotations are ultimately reviewed and confirmed by these two senior experts. The annotation tool used is 3D Slicer (v5.8.0) software.

[0041] After data partitioning, the pre-training phase began. Training was performed on an Ubuntu 20.04.4 system, using an Intel Xeon Platinum 8358 CPU and an NVIDIA RTX 4090 graphics card. The PyTorch framework was used, with a 3d_fullres network configuration, the Adam optimizer, an initial learning rate of 0.01, weight decay of 3e-5, and 500 training epochs. Five-fold cross-validation was employed to ensure robustness.

[0042] In the pre-training phase, the first sample dataset and the second training data are input into the current second segmentation model for training. After the model outputs the patch segmentation prediction map, a penalty is applied to the predicted region that exceeds the allowed distribution range of the patch and the spatial penalty loss is calculated. The model parameters are updated according to the loss result and the above training process is executed repeatedly until all indicators reach the preset pre-training stopping condition, thus completing the model pre-training process.

[0043] After pre-training, the model enters the fine-tuning phase, employing five-fold cross-validation. A second dataset with higher annotation precision and a larger sample size is used in conjunction with this second training dataset to further train the model. For example... Figure 2 As shown, the second sample dataset is obtained by inferring all CTA images using a pre-trained model. The inference output is then combined with the lower-precision annotations from the first sample dataset, and manually reviewed and corrected. After each training iteration, spatial penalty loss is calculated for regions exceeding the reasonable distribution range in the patch segmentation prediction map, and the model parameters are updated. This process is iterated continuously until the model performance meets the preset fine-tuning training stopping criteria, completing the entire training process. The fine-tuning phase employs multi-GPU parallel inference combined with an ensemble voting strategy to improve inference efficiency and the robustness of the segmentation results. The specific implementation is as follows: 1) Parallel inference configuration: A parallel inference environment is built using 5 NVIDIA RTX 4090s, and the 5 independent models obtained from the 5-fold cross-validation are deployed in 5 inference processes respectively.

[0044] 2) Integrated voting decision: The predicted masks of the same target region output by the 5 models are integrated and voted on. The voting rules are as follows: the predicted category of each voxel is counted, the prediction results with ≥3 votes are retained, and the prediction results with <3 votes are discarded; if the highest number of votes for a voxel is less than 3, it is judged as background (to avoid false detections); through integrated voting, the prediction bias of the single model is effectively avoided, the phenomenon of missed detections and false detections is reduced, and the final segmentation results of various tasks are finally output.

[0045] In this embodiment, the current second segmentation model includes a backbone network, an encoder, and a decoder. Obtaining the patch segmentation prediction map output by the current second segmentation model includes: extracting features from the first training data using the backbone network to obtain original image features; downsampling the original image features using the encoder to obtain multi-scale image features; downsampling the second training data using the current second segmentation model to obtain multi-scale mask features; normalizing the multi-scale mask features using the decoder to obtain spatial attention weights; and performing voxel-by-voxel multiplication of the spatial attention weights with the multi-scale image features to output the patch segmentation prediction map.

[0046] A second segmentation model (Vessel-GuidedAttention U-Net) is constructed for segmenting plaques in CTA image data. This model is based on the U-Net architecture and introduces a vessel mask-guided attention mechanism, such as... Figure 3 As shown, the current second segmentation model includes a backbone network, an encoder, and a decoder.

[0047] The first and second training data from different modalities are input into the current second segmentation model. To reduce memory usage, a sliding window strategy is used to simultaneously segment the first and second training data into 128×128×128 pixel blocks, input into the network with a batch size of 2, and input tensor sizes of (2, 1, 128, 128, 128). The input images from the two modalities must be spatially aligned to avoid the attention mechanism from failing due to misalignment.

[0048] The backbone network extracts features only from the first training data to obtain the original image features. The encoder uses a five-layer convolutional structure with 32, 64, 128, 256, and 320 channels for each layer. The first convolutional layer has a stride of 1×1×1 to maintain the same size, while the next four layers use a stride of 2×2×2 to downsample, resulting in an overall downsampling rate of 16 times to obtain multi-scale mask features. The network uses 3×3×3 three-dimensional convolutions, instance normalization, and the LeakyReLU activation function throughout. Simultaneously, the second segmentation model is used to downsample the second training data (using the same downsampling rhythm as the encoder) to obtain multi-scale mask features. The feature scale of the multi-scale mask features perfectly matches the scale of the CTA feature maps extracted by the backbone network.

[0049] In the decoder section, the multi-scale mask features are normalized using the decoder to obtain spatial attention weights. The attention weights are normalized using the Sigmoid activation function to ensure that the weight range is between [0,1]. The spatial attention weights are multiplied voxel-by-voxel with the multi-scale mask features of the corresponding layer to enhance the feature response of the region near the blood vessels (the feature value is amplified after multiplication) and weaken the feature signal of the non-vascular region (the feature value approaches 0 after multiplication). This guides the model to focus on the plaque region near the blood vessels and suppress false positive segmentation of the non-vascular region to output a plaque segmentation prediction map.

[0050] This embodiment employs a spatial penalty loss, imposing a significant penalty if the plaque predicted by the network is too far from the blood vessel. Since plaques can cause outward expansion of the blood vessel (positive reconstruction), a moderate dilation of the blood vessel mask is used. This involves first dilating the arterial mask obtained from the second training data to define the permissible distribution range of the plaque, and then penalizing predictions outside the dilated region. Let the plaque probability map output by the model be... The true plaque label is The input blood vessel mask is First of all, Morphological dilation was performed to obtain (Represents the maximum anatomical extent within which plaques are allowed to exist). Define spatial penalty loss. : ; When the model is outside the inflated blood vessel mask ( Predicting plaque probability When the blood vessel is near the vessel, a loss is generated directly; in the vicinity of the blood vessel, the loss is 0.

[0051] Step S14: Update the parameters of the current second segmentation model based on the spatial penalty loss to obtain a new current second segmentation model, and jump back to the step of training the current second segmentation model using the first training data and the second training data until the preset stop training condition is met, and determine the new current second segmentation model as the target second segmentation model.

[0052] In this embodiment, updating the parameters of the current second segmentation model based on the spatial penalty loss includes: obtaining the similarity coefficient loss and cross-entropy loss of the current second segmentation model according to the patch segmentation prediction map and the pre-annotated real labels in the first training data; obtaining the total loss of the current second segmentation model based on the spatial penalty loss, the similarity coefficient loss and the cross-entropy loss; and updating the parameters of the current second segmentation model based on the total loss.

[0053] The patch segmentation prediction map output by the current second segmentation model is compared with the pre-labeled true patch labels in the first training data, and the similarity coefficient loss corresponding to the current second segmentation model is calculated. With cross-entropy loss The similarity coefficient loss is used to measure the predicted segmentation map. With real labels The overlap between regions; the cross-entropy loss provides class-by-class probability supervision for each voxel, measuring the difference between the predicted class distribution and the true label distribution. Integrating the aforementioned spatial penalty loss, similarity coefficient loss, and cross-entropy loss, the total loss of the current second segmentation model is calculated, as shown in the following formula: ; in, It is the balance coefficient.

[0054] Then, based on the total loss, the error is backpropagated, and the internal parameters of the current second segmentation model are iteratively updated to gradually correct the prediction bias of the model until the preset stopping training condition is met, such as the convergence of the new current second segmentation model reaching the preset convergence threshold. In this way, the new current second segmentation model is determined as the target second segmentation model.

[0055] Step S15: Input the current head and neck CTA image data into the first target segmentation model to output the target artery vertebral segmentation map. Input the artery mask in the target artery vertebral segmentation map and the current head and neck CTA image data into the second target segmentation model to output the target plaque segmentation map.

[0056] The current head and neck CTA image data to be detected is input into the trained target first segmentation model. The model identifies and segments the arterial and vertebral regions in the image, and then outputs the corresponding target arterial and vertebral segmentation map. The arterial mask representing the vascular region is extracted from the segmentation map. The extracted arterial mask and the original current head and neck CTA image data are then input into the trained target second segmentation model. Based on the built-in vascular-guided attention structure of the model and the reference range defined by the arterial mask, plaque identification and segmentation are completed. Finally, the target second segmentation model outputs the corresponding target plaque segmentation map.

[0057] Step S16: Fuse the target artery vertebral segmentation map and the target plaque segmentation map to generate a target segmentation map including the artery, vertebra, and plaque.

[0058] In this embodiment, fusing the target arterial vertebral segmentation map and the target plaque segmentation map includes: binarizing and performing three-dimensional connected component analysis on the plaque mask in the target plaque segmentation map, segmenting the plaque in the target plaque segmentation map into multiple first independent plaque regions, performing three-dimensional dilation on each first independent plaque region to obtain a dilated plaque region, removing the dilated plaque regions in the target plaque segmentation map that do not spatially intersect with the arterial vessel mask in the target arterial vertebral segmentation map to obtain a removed plaque segmentation map; and then, combining the image grayscale values ​​corresponding to the mask identifiers of each voxel in the removed plaque segmentation map with... A preset grayscale threshold range is compared, and voxel masks with grayscale values ​​exceeding the preset grayscale threshold range are identified as over-predicted abnormal voxels. These abnormal voxels in the removed patch segmentation map are then filtered to obtain a corrected patch segmentation map. Three-dimensional connected component analysis is performed on the corrected patch segmentation map to obtain multiple second independent patch regions. The second independent patch region containing the most voxels is identified as the target independent patch region. The patch size and volume in the target independent patch region are determined based on the image physical spacing. Finally, the corrected patch segmentation map is fused with the target arterial vertebral segmentation map.

[0059] Robust optimization of the inference results and statistical quantification of plaques are performed based on computer vision algorithms. Firstly, false positive plaque elimination is based on blood vessel guidance. Specifically, the plaque mask within the target plaque segmentation map is binarized and three-dimensional connected component analysis is conducted. Based on this, the plaque in the image is divided into multiple first independent plaque regions. Then, each first independent plaque region is subjected to three-dimensional dilation using a cube structuring element of a specified size to obtain the dilated plaque region, simulating the spatial neighborhood of the plaque. ; in, This represents the first element extracted through connected component analysis. Individual patch areas, For a three-dimensional cube structural element of a preset size, For a three-dimensional dilation operator, This is the extended region of the generated spatial neighborhood.

[0060] Next, it is determined whether the expanded plaque region has spatial intersection with the arterial vascular mask in the target arterial vertebral segmentation map. If it does, it is retained; otherwise, it is considered a false positive and the portion of the expanded plaque region that does not spatially intersect with the arterial vascular mask is removed. The retained effective regions are combined into a new mask, thus obtaining the plaque segmentation map after removal. The result is shown in the image below. Figure 4 As shown, Figure 4 In the image, (a) represents the original CTA transverse image. Figure 4In the diagram, (b) and (c) represent the segmentation mask and 3D modeling image before false positive removal of plaques. Figure 4 In this context, (d) represents the actual mask label. Figure 4 In this diagram, (e) and (f) represent the segmentation mask and 3D model image after false positive removal of the plaque, as shown in the following formula: ; in, This is the effective patch mask after reconstruction, i.e., the patch segmentation map after removal. This is the corresponding arterial mask.

[0061] Subsequently, for overprediction, a threshold filtering method is used. Specifically, the image grayscale value corresponding to the mask identifier of each voxel in the patch segmentation image after removal is extracted. This grayscale value is compared with a preset grayscale threshold range one by one. Voxel masks with grayscale values ​​exceeding the corresponding threshold range are identified as abnormal voxels with overprediction problems. These abnormal voxels are then removed from the patch segmentation image after removal, thus obtaining the corrected patch segmentation image. In other words, invalid voxels whose mask values ​​are for the target area but whose corresponding CT grayscale values ​​are outside the set threshold range are selected, as shown below: ; in, In three-dimensional space coordinates, The input is the initial prediction mask. The image CT value at the corresponding coordinates. A reasonable grayscale threshold range is set for the target organization. Voxels that meet the above conditions are judged as invalid target regions (i.e., predicted abnormal voxels), and their mask values ​​are forcibly modified to the background to eliminate abnormal predictions. Finally, the corrected mask (i.e., the corrected patch segmentation map) is output. : ; Subsequently, a three-dimensional connected component analysis was performed on the corrected patch segmentation map using a six-connected structuring element, which identified multiple second independent patch regions. The calculation is as follows: ; in, This is the patch mask to be processed (i.e., the corrected patch segmentation image). The operator represents the connected component labeling operator. It is a six-connected structural element.

[0062] The distribution pattern along the Z-axis was statistically analyzed, and the region with the highest number of voxels was selected from all the second independent patch regions as the target independent patch region. That is, the axial layer with the highest number of voxels was determined as the core analysis layer, and the physical size of the patch was quantified based on this layer. ; in, For the first The core layer index with the largest patch cross-sectional area; and These represent the actual physical width and height of the patch, respectively. and The horizontal and vertical physical spacing of the image.

[0063] The size and volume of the target independent plaque region are calculated by combining the physical spacing corresponding to the image. Finally, the corrected plaque segmentation map and the target arterial vertebra segmentation map after all the above processing are fused together.

[0064] In this embodiment, after fusing the target artery and vertebral segmentation map and the target plaque segmentation map to generate a target segmentation map including the artery, vertebra, and plaque, the method further includes: converting the coordinate system of the target segmentation map into the RAS visualization coordinate system based on the spatial affine matrix of the current head and neck CTA image data to obtain a coordinate-adapted mask image; extracting surface meshes from the coordinate-adapted mask image using the moving cube algorithm, generating vertex and face data based on the surface meshes, and constructing an initial three-dimensional geometric model based on the vertex and face data; performing smoothing and denoising processing on the initial three-dimensional geometric model using the VTK window sine polynomial smoothing algorithm to obtain a smoothed mesh model; and performing connected component analysis and noise region removal on the smoothed mesh model to obtain the target three-dimensional visualization model.

[0065] The spatial affine matrix corresponding to the current head and neck CTA image data is retrieved. Based on this matrix, a coordinate system transformation operation is performed on the target segmentation image to convert the original coordinate system into the RAS visualization coordinate system, resulting in a coordinate-adapted mask image. Subsequently, the moving cube algorithm is used to extract the surface mesh from the coordinate-adapted mask image. Based on the extracted surface mesh, the corresponding vertex and face data are generated. Using the moving cube algorithm with a threshold of 0.5 as the isosurface threshold, the three-dimensional surface mesh is extracted from the 3D blood vessel segmentation image. The model vertex and face data are calculated and generated, and finally, a complete three-dimensional geometric model structure of the blood vessel is constructed.

[0066] The VTK window sinusoidal polynomial smoothing algorithm was used to smooth and denoise the initial 3D geometric model, resulting in a smoothed mesh model. Specifically, the VTK window sinusoidal polynomial smoothing algorithm was used to smooth and denoise the extracted 3D mesh model. The smoothing iterations were 50 times and the smoothing coefficient was 0.05, eliminating jagged edges and burrs on the mesh surface, making the model surface smoother and closer to the real anatomical structure.

[0067] Next, connected component analysis is performed on the smoothed mesh model to remove noisy regions according to set rules. Specifically, connected component analysis is used to remove small noise fragments. A minimum voxel value of 50 is set, and small regions with fewer than this value are deleted. Figure 5 As shown, the final result is a target 3D visualization model that can be used for display.

[0068] The beneficial effects of this application are as follows: This application uses historical head and neck CTA image data as the first training data; it inputs the first training data into a first segmentation model for segmenting arteries and vertebrae in CTA image data, and uses the arterial mask in the output arterial-vertebral segmentation map as the second training data; it constructs a current second segmentation model for segmenting plaques in CTA image data based on a vascular-guided attention mechanism, and trains the current second segmentation model using the first and second training data. A penalty is applied to the predicted regions outside the allowed distribution range of plaques in the plaque segmentation prediction map output by the current second segmentation model to obtain the spatial penalty loss of the current second segmentation model; wherein, the allowed distribution range of plaques is obtained by dilating the second training data. The scope is determined; the parameters of the current second segmentation model are updated based on the spatial penalty loss to obtain a new current second segmentation model, and the process jumps back to the step of training the current second segmentation model using the first training data and the second training data until a preset training stop condition is met, and the new current second segmentation model is determined as the target second segmentation model; the current head and neck CTA image data is input into the target first segmentation model to output a target artery and vertebral segmentation map, and the artery mask in the target artery and vertebral segmentation map and the current head and neck CTA image data are input into the target second segmentation model to output a target plaque segmentation map; the target artery and vertebral segmentation map and the target plaque segmentation map are fused to generate a target segmentation map including arteries, vertebrae, and plaques.Therefore, this application first uses historical head and neck CTA image data as the basic training sample, and can accumulate feature learning basis based on real clinical image data. After feeding this type of data into a first segmentation model specifically for identifying arteries and vertebrae, the arterial mask is extracted as a matching training sample. The reference boundary for plaque recognition can be defined based on the actual distribution range of blood vessels. On this basis, a plaque segmentation model with a blood vessel-guided attention structure is built. The model is trained simultaneously using two types of training data. During the training process, the region after the arterial mask is expanded is used to define the reasonable range of plaque existence. For prediction results that exceed this range, the loss value is calculated and the model output is constrained. This can reduce the situation where the model identifies plaques without basis. The internal parameters of the model are continuously corrected based on the calculated loss, and the model is iteratively optimized repeatedly. The model's recognition capability can be gradually improved until the end of training, gradually increasing the accuracy of the model in distinguishing the location and morphology of plaques. After training, a mature and usable plaque segmentation model is obtained. In practical applications, the first segmentation model is used to divide the arterial and vertebral regions. Then, the arterial mask obtained from the segmentation is combined with the original image and input into the plaque segmentation model to complete plaque detection. The segmentation results of different tissues are obtained by relying on the two segmentation steps. Finally, the segmentation images of multiple tissues are integrated to form a complete segmentation pattern. It can accurately distinguish the three types of target tissues in head and neck CTA images: arteries, vertebrae, and plaques. This reduces the recognition bias caused by a single model recognizing multiple types of tissues at the same time, conforms to the actual distribution pattern of plaques in head and neck blood vessels, and effectively improves the accuracy and reliability of the segmentation results of various lesions and tissues in head and neck CTA images.

[0069] An additional 100 head and neck CTAs were clinically evaluated by four radiologists. The annotation time of the radiologists during the pre-training phase and the execution time of the model inference in the 100 cases were statistically analyzed. The review time of the radiologists before and after using the AI ​​model was statistically analyzed. For each CTA, two different radiologists reviewed the images, with and without the AI ​​reconstruction model reference. The radiologists took turns to eliminate subjective differences. Finally, the radiologists scored the reconstruction quality of the four segmentation tasks of arteries, hard plaques, soft plaques and vertebrae. The scoring criteria were as follows: (1) 9-10: The target area was completely segmented, with no significant FP or FN. The segmentation boundary was highly consistent with the real boundary and could be directly used for clinical analysis and decision-making. (2) 8-9: There were slight FP and FN. The segmentation boundary had slight deviations in a few places. It met the clinical application requirements and required a small amount of review. (3) 7-8: There were some FP and FN. The segmentation boundary had visible deviations in multiple places. It can only meet the clinical reference needs and requires a lot of manual correction (4) 6-7: There are a lot of FP and FN, and the segmentation boundary does not match the true boundary seriously, which cannot meet the clinical needs. However, the clinical evaluation results of this embodiment are as follows Figure 6 As shown, Figure 6 (a) in the figure represents a statistical comparison of the time taken for manual reconstruction versus AI reconstruction, and for traditional image reading versus AI-assisted image reading. Figure 6(b) in the figure represents a statistical comparison chart of the radiologists' scores for the 3D reconstruction quality of the four tasks, and this embodiment has the following advantages: 1) Excellent multi-target segmentation accuracy, breaking through the bottleneck of soft plaque segmentation: After fine training, the average Dice coefficient of artery segmentation with five-fold cross-validation is 0.9757, the average Dice coefficient of vertebral segmentation with five-fold cross-validation is 0.9773, the average Dice coefficient of hard plaque segmentation with five-fold cross-validation is 0.8971, and the average Dice coefficient of soft plaque segmentation with five-fold cross-validation is 0.7894.

[0070] 2) Post-processing algorithms further improve segmentation specificity: The vascular-guided false positive removal algorithm and threshold filtering algorithm increase the average Dice coefficient of hard plaque segmentation to 0.9050 and the accuracy from 0.9594 to 0.9812; the average Dice coefficient of soft plaque segmentation increases to 0.7983 and the accuracy from 0.9284 to 0.9590, effectively suppressing false positives.

[0071] 3) Improved annotation efficiency and data quality, reducing research and clinical costs: The two-stage strategy of "pre-training-fine training" reduces the amount of annotation while ensuring data quality, providing a replicable model for the tedious medical data annotation work, and providing a reliable data foundation for related segmentation studies.

[0072] 4) Significantly improves clinical work efficiency and shortens the diagnosis and treatment cycle: AI-automated reasoning workflow reduces image reconstruction time from an average of 56.54 minutes to 7.1 minutes, a reduction of 87.44%; and reduces doctors' image reading time from an average of 23.35 minutes to 8.71 minutes, a reduction of 62.70%.

[0073] 5) The 3D reconstruction results meet clinical application needs and are suitable for clinical diagnosis and treatment scenarios: The median reconstruction quality scores of radiologists for the four segmentation tasks were: artery 9.4, vertebral 9.4, hard plaque 9.2, and soft plaque 8.9. With minimal manual review, it meets clinical reference requirements. It has already been deployed and applied in neurosurgery.

[0074] See Figure 7 As shown in the figure, this application discloses an information processing apparatus for CTA image segmentation, applied to a computer device, comprising: The first data determination module 11 is used to determine the collected historical head and neck CTA image data as the first training data; The second data determination module 12 is used to input the first training data into the first target segmentation model for segmenting arteries and vertebrae in CTA image data, and to determine the artery mask in the output artery and vertebral segmentation map as the second training data. The loss acquisition module 13 is used to construct a current second segmentation model for segmenting plaques in CTA image data based on a blood vessel-guided attention mechanism, and to train the current second segmentation model using the first training data and the second training data. A penalty is applied to the predicted regions located outside the allowed distribution range of the plaques in the plaque segmentation prediction map output by the current second segmentation model to obtain the spatial penalty loss of the current second segmentation model; wherein, the allowed distribution range of the plaques is the range obtained by dilating the second training data. The target model acquisition module 14 is used to update the parameters of the current second segmentation model based on the spatial penalty loss to obtain a new current second segmentation model, and then jump back to the step of training the current second segmentation model using the first training data and the second training data until the preset training stop condition is met, and the new current second segmentation model is determined as the target second segmentation model. The image segmentation module 15 is used to input the current head and neck CTA image data into the target first segmentation model to output the target artery vertebral segmentation map, and to input the artery mask in the target artery vertebral segmentation map and the current head and neck CTA image data into the target second segmentation model to output the target plaque segmentation map. The image fusion module 16 is used to fuse the target artery vertebral segmentation map and the target plaque segmentation map to generate a target segmentation map including the artery, vertebra, and plaque.

[0075] Furthermore, embodiments of this application also provide an electronic device. Figure 8 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0076] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Specifically, it may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the information processing method for CTA image segmentation performed by the electronic device disclosed in any of the foregoing embodiments.

[0077] In this embodiment, the power supply 23 is used to provide operating voltage for various hardware devices on the electronic device; the communication interface 24 can create a data transmission channel between the electronic device and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0078] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0079] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0080] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. The operating system can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the CTA image segmentation information processing method executed by the electronic device as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input / output interface 25.

[0081] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned information processing method for CTA image segmentation. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0083] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly in hardware, software modules executed by a processor, or a combination of both. The software module may be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), register, hard disk, removable disk, CD-ROM (Compact Disc Read-Only Memory), or any other form of storage medium known in the art.

[0084] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0085] The above provides a detailed description of an information processing method, apparatus, device, and medium for CTA image segmentation provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An information processing method for CTA image segmentation, characterized in that, Applied to computer devices, including: The collected historical head and neck CTA image data was selected as the first training data; The first training data is input into the first segmentation model for segmenting arteries and vertebrae in CTA image data, and the artery mask in the output artery and vertebral segmentation map is determined as the second training data. A current second segmentation model for segmenting plaques in CTA image data is constructed based on a blood vessel-guided attention mechanism. The current second segmentation model is trained using the first training data and the second training data. A penalty is applied to the predicted regions in the plaque segmentation prediction map output by the current second segmentation model that are outside the allowed distribution range of the plaque, so as to obtain the spatial penalty loss of the current second segmentation model. The allowed distribution range of the plaque is the range obtained by dilating the second training data. The parameters of the current second segmentation model are updated based on the spatial penalty loss to obtain a new current second segmentation model. Then, the process jumps back to the step of training the current second segmentation model using the first training data and the second training data until the preset training stop condition is met, and the new current second segmentation model is determined as the target second segmentation model. The current head and neck CTA image data is input into the first target segmentation model to output a target artery vertebral segmentation map. The artery mask in the target artery vertebral segmentation map and the current head and neck CTA image data are input into the second target segmentation model to output a target plaque segmentation map. The target artery vertebral segmentation map and the target plaque segmentation map are fused to generate a target segmentation map that includes the artery, vertebra, and plaque.

2. The information processing method for CTA image segmentation according to claim 1, characterized in that, The preset training stop conditions include preset pre-training stop conditions and preset fine-tuning stop conditions; the step of training the current second segmentation model using the first training data and the second training data includes: The first training data is divided into a first sample dataset and a second sample dataset; wherein the annotation accuracy of the first sample dataset is lower than that of the second sample dataset and the number of samples in the first sample dataset is less than that in the second sample dataset; The current second segmentation model is pre-trained using the first sample dataset and the second training data, and then the process jumps to the step of penalizing the predicted regions located outside the allowed distribution range of patches in the patch segmentation prediction map output by the current second segmentation model, until the preset pre-training stop condition is met. The second sample dataset and the second training data are used to train the current second segmentation model using a five-fold cross-validation method. Then, the process jumps to the step of penalizing the predicted regions located outside the allowed distribution range of patches in the patch segmentation prediction map output by the current second segmentation model, until the preset fine-tuning training stop condition is met.

3. The information processing method for CTA image segmentation according to claim 1, characterized in that, The step of determining the collected historical head and neck CTA image data as the first training data includes: An adaptive data preprocessing operation is performed based on the task type corresponding to the historical raw head and neck CTA image data to obtain preprocessed image data. The image window width and window level parameters of the preprocessed image data are adaptively configured using a boundary contrast maximization method to obtain the first training data.

4. The information processing method for CTA image segmentation according to claim 3, characterized in that, The method of adaptively configuring the image window width and window level parameters of the preprocessed image data using a boundary contrast maximization approach to obtain the first training data includes: Based on the initial mask label corresponding to the preprocessed image data, the original region of interest containing the target tissue is extracted from the preprocessed image data; wherein, the target tissue is a blood vessel or a vertebra. The original region of interest is morphologically dilated, and the dilated region is subtracted from the original region of interest to obtain a background neighborhood voxel set; wherein, the background neighborhood voxel set is a set of voxels that surround the target tissue and are in close contact with the boundary of the target tissue; A two-dimensional search matrix is ​​set based on a preset window width and window level range. For each combination of window width and window level in the two-dimensional search matrix, the first gray-level distribution information after mapping the voxel set in the original region of interest and the second gray-level distribution information after mapping the voxel set in the background neighborhood are determined respectively. The evaluation index function constructed based on Fisher's discrimination criterion, the first gray-scale distribution information, and the second gray-scale distribution information are used to quantify the class separation degree between the target organization and the target organization's neighborhood background. Iterate through each group of window width and window level combinations, and determine the window width and window level combination that maximizes the class separation as the global optimal parameter for the segmentation task. Use the global optimal parameter to perform image grayscale mapping correction on the preprocessed image data to obtain the first training data.

5. The information processing method for CTA image segmentation according to claim 1, characterized in that, The current second segmentation model includes a backbone network, an encoder, and a decoder; obtaining the patch segmentation prediction map output by the current second segmentation model includes: The backbone network is used to extract features from the first training data to obtain original image features, and the encoder is used to downsample the original image features to obtain multi-scale image features. The current second segmentation model is used to downsample the second training data to obtain multi-scale mask features. The decoder is used to normalize the multi-scale mask features to obtain spatial attention weights, and the spatial attention weights are multiplied by voxel-by-voxel with the multi-scale image features to output a patch segmentation prediction map. Accordingly, updating the parameters of the current second segmentation model based on the spatial penalty loss includes: The similarity coefficient loss and cross-entropy loss of the current second segmentation model are obtained based on the patch segmentation prediction map and the pre-labeled real labels in the first training data. The total loss of the current second segmentation model is obtained based on the spatial penalty loss, the similarity coefficient loss, and the cross-entropy loss, and the parameters of the current second segmentation model are updated based on the total loss.

6. The information processing method for CTA image segmentation according to claim 1, characterized in that, The step of fusing the segmentation map of the target artery vertebra and the segmentation map of the target plaque includes: The patch mask in the target patch segmentation map is binarized and subjected to three-dimensional connected component analysis. The patch in the target patch segmentation map is segmented into multiple first independent patch regions. Each first independent patch region is three-dimensionally dilated to obtain the dilated patch region. The dilated patch regions that do not spatially intersect with the arterial vessel mask in the target patch segmentation map and the target arterial vertebral segmentation map are removed to obtain the removed patch segmentation map. The image grayscale value corresponding to the mask identifier of each voxel in the removed patch segmentation image is compared with a preset grayscale threshold range. The voxel mask with the image grayscale value exceeding the preset grayscale threshold range is determined as a prediction anomalous voxel with over-prediction. The prediction anomalous voxels in the removed patch segmentation image are then filtered to obtain the corrected patch segmentation image. Three-dimensional connected component analysis is performed on the corrected patch segmentation map to obtain multiple second independent patch regions. The second independent patch region containing the most voxels is determined as the target independent patch region. The patch size and patch volume in the target independent patch region are determined based on the image physical spacing. The corrected plaque segmentation map is fused with the target arterial vertebral segmentation map.

7. The information processing method for CTA image segmentation according to any one of claims 1 to 6, characterized in that, After fusing the target artery vertebral segmentation map and the target plaque segmentation map to generate a target segmentation map including the artery, vertebra, and plaque, the method further includes: Based on the spatial affine matrix of the current head and neck CTA image data, the coordinate system of the target segmentation map is converted into the RAS visualization coordinate system to obtain the mask image after coordinate adaptation. The moving cube algorithm is used to extract the surface mesh from the coordinate-adapted mask image, generate vertex and face data based on the surface mesh, and construct an initial three-dimensional geometric model based on the vertex and face data. The initial three-dimensional geometric model was smoothed and denoised using the VTK window sine polynomial smoothing algorithm to obtain a smoothed mesh model. The smoothed mesh model is subjected to connected component analysis and noise region removal to obtain the target 3D visualization model.

8. An information processing apparatus for CTA image segmentation, characterized in that, Applied to computer devices, including: The first data determination module is used to determine the collected historical head and neck CTA image data as the first training data; The second data determination module is used to input the first training data into the first target segmentation model for segmenting arteries and vertebrae in CTA image data, and to determine the artery mask in the output artery and vertebral segmentation map as the second training data. The loss acquisition module is used to construct a current second segmentation model for segmenting plaques in CTA image data based on a blood vessel-guided attention mechanism, and to train the current second segmentation model using the first training data and the second training data. A penalty is applied to the predicted regions located outside the allowed distribution range of the plaques in the plaque segmentation prediction map output by the current second segmentation model to obtain the spatial penalty loss of the current second segmentation model; wherein, the allowed distribution range of the plaques is the range obtained by dilating the second training data. The target model acquisition module is used to update the parameters of the current second segmentation model based on the spatial penalty loss to obtain a new current second segmentation model, and then jump back to the step of training the current second segmentation model using the first training data and the second training data until the preset training stop condition is met, and the new current second segmentation model is determined as the target second segmentation model. The image segmentation module is used to input the current head and neck CTA image data into the target first segmentation model to output the target artery vertebral segmentation map, and to input the artery mask in the target artery vertebral segmentation map and the current head and neck CTA image data into the target second segmentation model to output the target plaque segmentation map. The image fusion module is used to fuse the target artery vertebral segmentation map and the target plaque segmentation map to generate a target segmentation map including the artery, vertebra, and plaque.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the information processing method for CTA image segmentation as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the information processing method for CTA image segmentation as described in any one of claims 1 to 7.

Images