A lateral ventricle choroid plexus segmentation and quantification method, system, storage medium and device based on enhanced T1 weighted magnetic resonance images

Abstract

The application belongs to the field of image segmentation, and relates to a lateral ventricle choroid plexus segmentation and quantification method based on enhanced T1 weighted magnetic resonance images, S1: obtaining brain enhanced T1 weighted magnetic resonance images and performing pretreatment to obtain standardized brain images; S2: based on the standardized brain images, combining a preset enhancement intensity threshold and a lateral ventricle anatomical position constraint, constructing a choroid plexus candidate region; S3: using a trained choroid plexus segmentation model to perform feature discrimination and segmentation on the choroid plexus candidate region to generate an initial segmentation mask; S4: performing post-processing optimization on the initial segmentation mask to remove false positive regions to obtain a final lateral ventricle choroid plexus segmentation mask; and S5: based on the final lateral ventricle choroid plexus segmentation mask, performing choroid plexus quantification parameter calculation. The problems of low segmentation precision, poor integrity and insufficient stability are solved.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation, and in particular relates to a quantitative method, system, storage medium and device for segmentation of the lateral ventricle choroid plexus based on enhanced T1-weighted magnetic resonance images. Background Technology

[0002] The choroid plexus is an important structural component within the ventricular system, primarily responsible for the production of cerebrospinal fluid (CSF). It is relatively small, villous in shape, and closely adjacent to the CSF space within the ventricles. In studies of neurodegenerative diseases, neuroinflammation, and aging, the volume, morphology, and degree of vascularization of the choroid plexus are considered to have significant clinical indicative value. Therefore, precise segmentation and quantitative analysis of the lateral ventricular choroid plexus have significant medical research and application implications.

[0003] Magnetic resonance imaging (MRI) is the primary non-invasive method for observing brain anatomical structures. Currently, segmentation methods for the choroid plexus largely rely on conventional T1WI structural images. Conventional T1-weighted imaging (T1WI) is the basic sequence for structural analysis. However, in conventional T1WI images, the signal contrast between the choroid plexus tissue and cerebrospinal fluid is low, and its structural boundaries are blurred. Furthermore, the techniques used in choroid plexus segmentation methods are usually based on manually set grayscale thresholds, region growing algorithms, or general image segmentation models. These choroid plexus segmentation methods have the following drawbacks: 1) In areas where the choroid plexus is small or has a delicate structure (such as the inferior and posterior horns of the lateral ventricle), the target signal is easily submerged by the background and cannot be effectively detected, resulting in a high false negative rate; 2) Adjacent vascular structures or some ventricular edge tissues that are similar to the choroid plexus signal are easily misidentified as the choroid plexus, leading to false detection; 3) The segmentation results often show breaks, local missing parts, or unreasonable boundaries, which seriously affect the reliability and repeatability of subsequent quantitative analysis such as volume measurement, resulting in incomplete and unstable segmentation.

[0004] Contrast-enhanced T1-weighted magnetic resonance imaging (T1WI) is an imaging sequence acquired after intravenous injection of gadolinium-based contrast agents. It enables the choroid plexus to exhibit significant and stable signal enhancement, creating a distinct signal difference from the unenhanced cerebrospinal fluid, thus providing an ideal imaging basis for choroid plexus identification. However, in current clinical practice and research, contrast-enhanced T1WI is mainly used for qualitative observation and diagnosis of lesions, and the choroid plexus-specific enhancement information it contains has not been systematically developed and integrated into automated segmentation processes.

[0005] Therefore, there is an urgent need to propose a method that can overcome complex background interference and achieve high-precision and high-completeness automatic segmentation, so as to make full use of the imaging advantages of enhanced T1WI and realize robust, accurate, and automated segmentation and advanced quantitative analysis of the contralateral ventricular choroid plexus. Summary of the Invention

[0006] This invention provides a method, system, storage medium, and device for quantitative segmentation of the lateral ventricle choroid plexus based on enhanced T1-weighted magnetic resonance images, which solves the problems of low segmentation accuracy, poor integrity, and insufficient stability.

[0007] The basic solution provided by this invention is a method for segmenting and quantifying the choroid plexus of the lateral ventricle based on enhanced T1-weighted magnetic resonance imaging, the specific steps of which include: S1: Acquire enhanced T1-weighted magnetic resonance images of the subject's brain and preprocess the images according to a pre-set image preprocessing strategy to obtain standardized brain images; S2: Based on standardized brain images, and combined with preset enhancement intensity thresholds and lateral ventricle anatomical location constraints, candidate regions of the choroid plexus are constructed. S3: Use the trained venous cluster segmentation model to perform feature discrimination and segmentation on the candidate regions of venous clusters to generate an initial segmentation mask; wherein, the venous cluster segmentation model is a deep learning model trained on enhanced T1-weighted magnetic resonance image samples and their corresponding venous cluster labeled masks. The model is configured to learn the enhancement features and spatial distribution features of venous clusters in enhanced T1-weighted images. S4: Post-processing optimization of the initial segmentation mask to remove false positive regions, resulting in the final lateral ventricle choroid plexus segmentation mask; S5: Based on the final lateral ventricle choroid plexus segmentation mask, perform quantitative parameter calculation of the choroid plexus.

[0008] Preferably, in step S1, the image preprocessing steps are as follows: Bias field correction was performed on the acquired enhanced T1-weighted magnetic resonance images of the brain; Identify the foreground region of brain parenchyma, extract the intensity features of the region, and use the intensity features to perform intensity normalization processing on the image; The brain parenchyma region was extracted, the skull and its external structures were removed to form a brain parenchyma image, and a brain parenchyma mask was constructed based on the brain parenchyma image; The brain parenchyma image and brain parenchyma mask are aligned to the standard template space to form a standardized brain image.

[0009] Preferably, in step S2, the steps for constructing the candidate region of the venule cluster are as follows: S2-1) Obtain a mask of the lateral ventricle region in a standardized brain image; S2-2) In standardized brain images, voxels with signal strength higher than a preset enhancement threshold are identified to obtain a set of enhanced voxels; S2-3) Perform spatial logic and operations on the enhanced voxel set and the lateral ventricle region mask to obtain a preliminary candidate voxel set; S2-4) Perform three-dimensional connected component analysis on the preliminary candidate voxel set, remove connected components with a volume smaller than the second preset volume threshold, and obtain the venule cluster candidate region.

[0010] More preferably, the enhanced voxel recognition strategy includes any of the following: a. Based on the statistical analysis of cerebrospinal fluid region signals, a preset enhancement threshold is set. The preset enhancement threshold is greater than or equal to the sum of the mean signal of the cerebrospinal fluid region and twice the standard deviation of the cerebrospinal fluid region signal. Voxels with signal strength greater than the preset enhancement threshold are identified as enhanced voxels. b. The percentage threshold based on whole-brain signal statistics is set as a preset enhancement threshold, wherein the percentage threshold is 10% to 15%. When identifying enhancement voxels, voxels whose signal intensity distribution is before the percentage threshold are identified as enhancement voxels.

[0011] Preferably, the convolutional neural network segmentation model is a three-dimensional fully convolutional neural network, including a multi-scale feature extraction module and an attention mechanism module; the multi-scale feature extraction module is used to fuse image features under different receptive fields, and the attention mechanism module is used to adaptively enhance feature channels or spatial regions related to convolutional cluster discrimination.

[0012] Preferably, in step S4, the post-processing optimization strategy is as follows: S4-1) Perform three-dimensional connected component analysis on the initial segmentation mask to remove isolated regions with a volume smaller than the third preset volume threshold; S4-2) Perform morphological opening and / or closing operations on the segmentation mask processed in step S4-1 to remove small burrs and fill local voids, and obtain a structurally continuous final segmentation mask as the final lateral ventricle choroid plexus segmentation mask.

[0013] Preferably, in step S5, the quantitative parameter calculation includes: S5-1) Calculate the total volume of the final lateral ventricle choroid plexus segmentation mask, and the lateral volumes of the left and right choroid plexuses; S5-2) Within the final lateral ventricle choroid plexus segmentation mask, the choroid plexus is distinguished into vascular-related components and non-vascular components based on the signal intensity of voxels in enhanced T1-weighted images. S5-3) Calculate the volume of vascular-related components and the volume of non-vascular components respectively, and calculate the proportion of the two to the total volume.

[0014] A lateral ventricular choroid plexus segmentation and quantification system based on enhanced T1-weighted magnetic resonance imaging, used to run the above method, includes: The image preprocessing module is used to acquire enhanced T1-weighted magnetic resonance images of the subject's brain and preprocess the images according to a preset image preprocessing strategy to obtain standardized brain images. The candidate region construction module constructs choroid plexus candidate regions based on standardized brain images and by combining preset enhancement intensity thresholds and lateral ventricle anatomical location constraints. The initial segmentation mask recognition module is used to call the trained vesicle cluster segmentation model to perform feature discrimination and segmentation on the vesicle cluster candidate region and generate an initial segmentation mask; wherein, the vesicle cluster segmentation model is a deep learning model trained based on enhanced T1 weighted magnetic resonance image samples and their corresponding vesicle cluster labeled masks. The model is configured to learn the enhancement features and spatial distribution features of vesicle clusters in enhanced T1 weighted images. The post-processing optimization module is used to perform post-processing optimization on the initial segmentation mask, remove false positive regions, and obtain the final lateral ventricle choroid plexus segmentation mask. The quantitative analysis module is used to calculate quantitative parameters of the choroid plexus based on the final lateral ventricle choroid plexus segmentation mask.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for quantifying the choroid plexus segmentation of the lateral ventricle based on enhanced T1-weighted magnetic resonance images.

[0016] An electronic device, comprising: Memory, which stores computer programs; A processor, connected to the memory, is configured to execute the computer program to implement the steps of a method for quantifying the lateral ventricular choroid plexus segmentation based on enhanced T1-weighted magnetic resonance images.

[0017] The principles and advantages of this invention are as follows: 1. By using enhanced T1-weighted magnetic resonance images as the core input, and making full use of the specific enhancement features of the choroid plexus, the bottleneck problem of insufficient contrast between the target and the background (cerebrospinal fluid) in conventional T1 images is fundamentally solved. By combining a two-level segmentation strategy of coarse localization of candidate regions based on enhancement features and fine discrimination based on deep learning models, especially in the anatomically complex inferior horn and posterior horn regions of the lateral ventricle, the false negative and false positive rates can be effectively reduced, and high-quality segmentation results with clear boundaries and continuous structures can be obtained.

[0018] 2. From image preprocessing, candidate region construction, deep learning segmentation to post-processing optimization and quantitative parameter calculation, a fully automated solution is formed, eliminating the need for human experience intervention, significantly reducing manual operation costs and subjective errors, and improving analysis efficiency in clinical and scientific research. At the same time, quantitative parameter calculation is performed on the dedicated mask after segmentation to achieve accurate matching between segmentation results and quantitative analysis, supporting advanced quantitative research needs.

[0019] 3. This invention systematically develops and integrates previously underutilized choroidal bundle enhancement information in enhanced T1WI, transforming it into key discriminative features for segmentation tasks. This breaks through the technical limitations of existing methods that mainly rely on conventional structural images, and opens up a new approach for quantitative analysis of choroidal bundle images. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the structure of the vesicle cluster segmentation model of the present invention; Figure 3 This is a three-dimensional spatial diagram of nnUNet segmenting the vesicle bundle based on enhanced T1 weighted imaging in an embodiment of the present invention; Figure 4 This is a diagram showing the vascular cluster segmentation mask and its vascular / non-vascular component identification results of the present invention. Detailed Implementation

[0021] The following detailed description illustrates the specific implementation method: The specific implementation process is as follows: (See details) Figures 1 to 4 A method for segmenting and quantifying the choroid plexus of the lateral ventricle based on enhanced T1-weighted magnetic resonance imaging, the specific steps of which include: S1: Acquire enhanced T1-weighted magnetic resonance images of the subject's brain and preprocess the images according to a pre-set image preprocessing strategy to obtain standardized brain images.

[0022] In step S1, the image preprocessing steps are as follows: Bias field correction is performed on the acquired enhanced T1-weighted MRI images of the brain. Specifically, due to the potential for magnetic field inhomogeneity in MRI scans, local brightness unevenness in the image (bias field effect) can occur, affecting the accuracy of subsequent segmentation. In this embodiment, the N4 Bias Field Correction algorithm is used to perform bias field correction on the original enhanced T1-weighted images to correct low-frequency intensity inhomogeneity, making the signal intensity of the same tissue more uniform across the entire image. The N4 Bias Field Correction algorithm is used to eliminate the influence of low-frequency intensity inhomogeneity on enhancement discrimination. Its iterative strategy can be set to four-level multi-resolution iteration (e.g., [50,50,50,50]), and the convergence threshold can be set to 1×10⁻⁶. -6 The B-spline grid spacing can be set in the range of 200-300 mm, and preferably about 250 mm; The brain parenchyma foreground region is identified, and the intensity features of the region are extracted. The intensity features are then used to normalize the image intensity. Specifically, after correcting the bias field, the brain parenchyma foreground region is identified (usually by binarizing the image and taking the largest connected component). The intensity features of the region are extracted, and the image intensity is normalized based on these features. Specifically, a normalization standard based on the percentile of the foreground region is adopted. That is, the 1st percentile (P1) and the 99th percentile (P99) in the brain parenchyma foreground are calculated as the lower and upper limits of the intensity, respectively. P1 is mapped to 0 and P99 is mapped to 1, and the remaining voxels are linearly scaled. At the same time, outliers are truncated to the [0,1] interval for easy processing by the subsequent model. Brain parenchyma regions are extracted, and the skull and its external structures are removed to form a brain parenchyma image. A brain parenchyma mask is then constructed based on the brain parenchyma image. Specifically, to eliminate differences in brain shape and size between different individuals and to achieve cross-individual comparison at the voxel level, in this embodiment, the brain parenchyma image I_brain and the brain parenchyma mask M_brain are aligned to a standard template space (such as the MNI152 standard space) using a nonlinear registration algorithm (such as the SyN algorithm) to generate a standardized brain image I_std and a corresponding standardized brain parenchyma mask M_std. In other embodiments, a deep learning brain extraction model (such as HD-BET or an equivalent model) is used to obtain a brain parenchyma probability map, and a brain parenchyma mask is formed with a probability threshold ≥0.5. Subsequently, three-dimensional morphological closing operations (such as 3×3×3 structuring elements, 1–2 iterations) can be used to fill local holes and suppress background interference. Brain parenchyma images and brain parenchyma masks are aligned to a standard template space to form standardized brain images. Specifically, to reduce the impact of differences in imaging voxel size and coordinate system on segmentation consistency, enhanced T1 images can be registered or resampled to a unified template space (e.g., MNI152 or an equivalent standard space). The size of the resampled voxels can be explicitly defined as 1.0 × 1.0 × 1.0 mm³. The images are interpolated using trilinear interpolation, and the labels are interpolated using nearest neighbor interpolation to ensure the spatial consistency and reproducibility of subsequent candidate region construction, segmentation, and quantitative indicators.

[0023] Enhanced T1-weighted magnetic resonance images of the brain used to construct the dataset and for subsequent practical application of this method for quantitative analysis of lateral ventricular choroid plexus segmentation are processed using this image preprocessing step, thereby providing a unified data foundation for model training and subsequent automated processing.

[0024] In this embodiment, for dataset construction, enhanced T1-weighted magnetic resonance imaging (MRI) images of subjects' brains collected clinically at our hospital are selected as the research object. The above-mentioned image preprocessing steps ensure the consistency of spatial resolution and data format of image data from different subjects. In the dataset construction stage, based on the image manifestation of the choroid plexus having obvious enhancement features in the enhanced T1-weighted images, a choroid plexus reference annotation dataset is constructed. The reference annotation data is used to indicate the spatial distribution location of the lateral ventricle choroid plexus in the enhanced T1 images and serves as a reference for model training and validation. The above annotation data can be obtained through manual annotation, semi-automatic annotation, or existing annotation datasets. Through the above steps, the brain magnetic resonance imaging dataset used for training and verification of the method of the present invention is constructed, laying the foundation for subsequent image preprocessing, feature extraction and segmentation model training.

[0025] Preferably, considering that the proportion of voxels in brain images is relatively low, a foreground sampling strategy can be introduced during the construction of training samples in the dataset. During the model training phase, the probability of image patches containing choroidal plexus regions being sampled is increased, thereby increasing the frequency of the model encountering foreground voxels of the choroidal plexus during training and improving the learning ability for small-volume targets.

[0026] S2: Based on standardized brain images, and combined with preset enhancement intensity thresholds and lateral ventricle anatomical location constraints, candidate regions of the choroid plexus are constructed. Among them, the anatomical position constraint of the lateral ventricle can be determined based on the lateral ventricle region mask or based on the equivalent anatomical prior. In this embodiment, the anatomical position constraint of the lateral ventricle is determined based on the lateral ventricle region mask. In step S2, the construction steps of the candidate region of the venule cluster are as follows: S2-1) Obtain a mask of the lateral ventricle region in a standardized brain image; S2-2) In standardized brain images, voxels with signal strength higher than a preset enhancement threshold are identified to obtain a set of enhanced voxels; S2-3) Perform spatial logic and operations on the enhanced voxel set and the lateral ventricle region mask to obtain a preliminary candidate voxel set; Preferably, when performing spatial logic and operations, it is necessary to perform the operations according to the pre-set anatomical distance constraints and anatomical axial distribution constraints; in, The anatomical distance constraint is to select areas where the lateral ventricle region mask extends into the parenchyma no more than a first preset distance (preferably 3 mm to 5 mm), and to exclude areas where the distance from the boundary of the lateral ventricle region mask exceeds a second preset distance (preferably 8 mm). Isolated enhanced regions (mm~12mm) are identified. Specifically, based on the anatomical growth characteristics of the choroid plexus tissue, which is attached to the lateral ventricle wall and extends inward, a dual distance constraint is applied to the spatial relationship between the enhancing voxels and the lateral ventricle structure. First, to fully cover the choroid plexus tissue that may be located near the choroidal fissure or slightly extended outward, a three-dimensional expansion operation with a radius of 5mm is performed on the lateral ventricle region mask, limiting the search range to the expanded area to ensure that no choroid plexus voxels located at the edge of the ventricle boundary are missed. Second, for possible artifacts or non-specific enhancements, the obtained enhanced regions are traversed, and isolated enhanced connected regions that are spatially more than 10mm away from the boundary of the original lateral ventricle region mask M_lv_std are automatically identified and excluded. Such isolated regions are anatomically unlikely to be structures attached to the choroid plexus and are mostly enhancing blood vessels or noise far from the ventricle. This strategy can effectively eliminate them.

[0027] The anatomical axial distribution constraint is as follows: along the Z-axis, a region overlapping with the typical distribution range related to the trigone and inferior horn of the lateral ventricle is selected. The typical distribution range is determined by experts in the field based on vascular or meningeal structures to avoid including enhanced vascular or meningeal structures. Specifically, based on the specific anatomical distribution pattern of the choroid plexus within the lateral ventricle (mainly concentrated in the trigone, body, and areas extending downward and backward to the inferior and posterior horns), anatomical prior constraints are introduced in the Z-axis (cephalothorax) and sagittal (left-right) directions. In a standard template space (such as the MNI space), the typical axial distribution range of the choroid plexus of the lateral ventricle is pre-defined based on extensive anatomical statistics or expert knowledge. In practice, enhanced voxels located outside this typical distribution range can be removed by consulting a table or calling a predefined axial range mask. This strategy aims to exclude enhanced structures (such as some perforating vessels or superior anastomotic veins) located in atypical positions such as the anterior horn of the lateral ventricle or above the corpus callosum, avoiding these structures from being mistakenly included in the choroid plexus candidate region.

[0028] S2-4) Perform three-dimensional connected component analysis on the preliminary candidate voxel set, remove connected components with a volume smaller than the second preset volume threshold, and obtain the venule cluster candidate region.

[0029] Preferably, the enhanced voxel recognition strategy includes any of the following: a. Based on the statistical analysis of cerebrospinal fluid region signals, a preset enhancement threshold is set. The preset enhancement threshold is greater than or equal to the sum of the mean signal of the cerebrospinal fluid region and twice the standard deviation of the cerebrospinal fluid region signal. Voxels with signal strength greater than the preset enhancement threshold are identified as enhanced voxels. b. The percentage threshold based on whole-brain signal statistics is set as a preset enhancement threshold, wherein the percentage threshold is 10% to 15%. When identifying enhancement voxels, voxels whose signal intensity distribution is before the percentage threshold are identified as enhancement voxels.

[0030] Step S2 leverages the unique advantages of enhancing the T1 image to generate a candidate region containing vein clusters but excluding most background regions through a simple and fast image processing method. This provides a highly focused search space for subsequent refined deep learning segmentation models, reducing computational burden and false detection rate.

[0031] S3: Use the trained venous cluster segmentation model to perform feature discrimination and segmentation on the candidate regions of venous clusters to generate an initial segmentation mask; wherein, the venous cluster segmentation model is a deep learning model trained on enhanced T1-weighted magnetic resonance image samples and their corresponding venous cluster labeled masks. The model is configured to learn the enhancement features and spatial distribution features of venous clusters in enhanced T1-weighted images. The vesicle cluster segmentation model is a three-dimensional fully convolutional neural network, including a multi-scale feature extraction module and an attention mechanism module. The multi-scale feature extraction module is used to fuse image features under different receptive fields, and the attention mechanism module is used to adaptively enhance feature channels or spatial regions related to vesicle cluster discrimination.

[0032] Specifically, in this embodiment, a three-dimensional fully convolutional segmentation network based on nnU-Net is used as the basic framework. Its encoder and decoder can adopt a symmetrical five-level hierarchical structure. The number of feature channels can be increased progressively in the order of 32, 64, 128, 256, and 320. The convolutional kernel can be 3×3×3. Downsampling can use convolution with a stride of 2 or an equivalent downsampling structure. The normalization method can use Instance... Normalization can be achieved by using LeakyReLU as the activation function (e.g., α=0.01) and adding Dropout (e.g., 0.2) at the encoding end to reduce the risk of overfitting. To enhance the representation ability of small-volume vesicle structures, multi-scale feature extraction and attention mechanisms can be introduced. The scale levels of multi-scale features can include different receptive fields or equivalent scales such as 1×1×1, 2×2×2, and 4×4×4. The fusion method can be to perform channel compression by concatenating scales one by one and then performing channel compression through 1×1×1 convolution. The attention mechanism can be channel attention (such as the SE module) or spatial attention or a combination of both. The compression coefficient of the SE module can be set to r=16, and the attention weight can be calculated by a nonlinear function composed of ReLU and Sigmoid to achieve adaptive emphasis on the enhanced features of the vesicle cluster.

[0033] Regarding training parameters, this invention clearly defines data augmentation, optimizer, learning rate, batch size, number of epochs, loss function, and key thresholds for ensemble and semi-supervised training to ensure reproducibility and ease of review and understanding. Data augmentation may include random rotation (e.g., ±15°), random scaling (e.g., 0.9–1.1), random translation (e.g., ±10 voxels), and intensity perturbation (e.g., ±10%) to improve the model's robustness to individual differences and variations in imaging conditions. The training optimizer may use Adam, with an initial learning rate set to 1×10⁻⁶. -4 The batch size can be set to 2, and the number of training epochs can be set to 500 epochs or an equivalent number of iterations. The loss function can be a combination of Dice Loss and Focal Loss to balance the overlapping optimization of small-volume targets and the learning of difficult examples. The α of Focal Loss can be set to 0.75 or adjusted within a reasonable range. Model ensemble can obtain multiple sub-models through cross-validation training and average their output probability maps. The number of ensemble models can be 5 or an equivalent number to improve the stability of the final segmentation. If a semi-supervised learning strategy is adopted, pseudo-labels can be generated for unlabeled samples after the initial model training. The pseudo-label generation threshold can be set to 0.9, and about 50 rounds of fine-tuning training or equivalent iterations can be performed to enhance the model's adaptability to complex anatomical regions.

[0034] For a new standardized brain image to be segmented and its corresponding choroid plexus candidate region, the image patch within the candidate region is used as input to the trained choroid plexus segmentation model. After forward propagation, the model outputs a probability map of the same size as the input image patch, where each voxel has a value between 0 and 1, representing the probability that the voxel belongs to the choroid plexus. Subsequently, the probability map is binarized using a preset probability threshold (e.g., 0.5) to obtain an initial segmentation mask. Preferably, if a model ensemble strategy (such as 5-fold cross-validation) is used during the training phase, the output probability maps of multiple models are fused on a voxel-by-voxel average basis during the inference phase, and then the fused probability map is thresholded to obtain a more stable and robust segmentation result.

[0035] S4: Post-processing optimization of the initial segmentation mask to remove false positive regions, resulting in the final lateral ventricle choroid plexus segmentation mask.

[0036] False positive regions can be defined as regions that meet any of the following conditions: enhanced connected regions located outside the ventricle and more than 10 mm from the ventricle boundary, small isolated connected regions, or regions that are shaped like long and thin blood vessels and lack spatial continuity. These regions are automatically removed in the post-processing stage to obtain a final choroid plexus segmentation mask with continuous structure, reasonable boundaries, and usable for quantitative analysis.

[0037] In step S4, the post-processing optimization strategy is as follows: S4-1) Perform three-dimensional connected component analysis on the initial segmentation mask to remove isolated regions with a volume smaller than a third preset volume threshold; wherein, the third preset volume threshold is 50-100mm. 3 In this embodiment, 75mm is preferred. 3 ; S4-2) Perform morphological opening and / or closing operations on the segmentation mask processed in step S4-1 to remove small burrs and fill local voids, and obtain a structurally continuous final segmentation mask as the final lateral ventricle choroid plexus segmentation mask.

[0038] Specifically, in this embodiment, during post-optimization, three-dimensional opening and closing operations are used. The opening operation structuring element can be set to 2×2×2 to remove small burrs, and the closing operation structuring element can be set to 3×3×3 to fill local holes. The number of iterations can be set to 1–2 times to achieve a balance between noise reduction and structure preservation.

[0039] S5: Based on the final lateral ventricle choroid plexus segmentation mask, perform quantitative parameter calculation of the choroid plexus.

[0040] In step S5, the quantitative parameter calculation includes: S5-1) Calculate the total volume of the final lateral ventricle choroid plexus segmentation mask, and the lateral volumes of the left and right choroid plexuses; S5-2) Within the final lateral ventricle choroid plexus segmentation mask, the choroid plexus is distinguished into vascular-related components and non-vascular components based on the signal intensity of voxels in enhanced T1-weighted images. In this invention, in the final result of the choroid plexus segmentation mask, we further subdivide the components within the choroid plexus region by enhancing the signal intensity in the T1-weighted image. To more accurately identify vascular and non-vascular components, we employ a clustering analysis-based method to further distinguish the segmented choroid plexus region into tissue components with different intensity characteristics.

[0041] Specifically, vascular components typically appear as regions of high signal intensity, a feature caused by blood flow within the vessels or the accumulation of enhancers. Because blood vessels exhibit significantly higher contrast in enhanced T1-weighted images, clustering algorithms can automatically identify and extract these high-signal regions. Vascular-related components are usually located within choroid plexuses and exhibit diverse spatial distributions and morphological characteristics.

[0042] To achieve this goal, we first performed cluster analysis on the segmented vascular plexus regions, dividing them into two categories based on signal intensity: one category for vascular components with higher signal intensity and the other for non-vascular components with lower signal intensity. Specifically, K-means clustering was used to analyze the signal intensity of each voxel in the image, assigning voxels to the two different categories based on the analysis results. Cluster analysis effectively distinguishes between vascular and non-vascular components based on differences in signal intensity.

[0043] Furthermore, after cluster analysis, vascular and non-vascular components were identified and labeled separately. This not only enhanced the structural details of the choroid plexus but also provided more accurate segmentation results in subsequent analyses, especially in quantitative analysis, providing accurate volume and spatial distribution data for vascular and non-vascular regions. Figure 4 The images show the choroid plexus segmentation mask and the results of vascular / non-vascular component identification. Based on the choroid plexus segmentation model, the images are automatically distinguished by clustering algorithms, clearly separating vascular-related components from non-vascular components within the choroid plexus. Specifically, the T1 plain scan image on the left and the overall choroid plexus mask CP (represented by outlines or semi-transparent color blocks) are used to show the anatomical location of the choroid plexus in the whole brain and the overall segmentation results obtained by this invention, providing a spatial reference for subsequent refined analysis. The enhanced T1 image on the right shows the vascular-related component VC and the non-vascular component NVC identified by clustering analysis. As can be seen from the images, the highly enhanced vascular components are mainly clustered in the core area of ​​the choroid plexus, consistent with the expected vascular course, while the low-enhanced non-vascular components are distributed in the surrounding matrix. The boundaries between the two are clear, and the spatial distribution is reasonable, providing a precise spatial localization and classification basis for calculating the volume and proportion of the two types of components in the subsequent steps S5-3.

[0044] This method combines cluster analysis with signal intensity thresholding, avoiding the challenges faced by traditional single-threshold methods. It can automatically adjust and optimize the segmentation process to address potential differences between samples. This approach enables more precise differentiation between vascular and non-vascular components in the choroid plexus, providing a solid foundation for further quantitative analysis, lesion assessment, and clinical research.

[0045] S5-3) Calculate the volume of vascular-related components and the volume of non-vascular components respectively, and calculate the proportion of the two to the total volume.

[0046] Specifically, in this embodiment, the components within the choroid plexus are classified based on the difference in enhancement intensity in enhanced T1 images. Voxels with enhancement intensity reaching P90 or higher within the choroid plexus region are defined as vascular-related components, while voxels below this threshold are defined as non-vascular components. Their volume and proportion are then statistically analyzed. This distinction can be further verified through consistency assessment with radiologists' manual interpretation of enhanced vascular pathways or with consistency assessment of repeated scans of the same subject at different time points. To verify the reproducibility of quantitative parameters, the same batch of samples can be processed three times independently (including preprocessing, inference, and postprocessing). The coefficient of variation (CV) of the output volume parameter is required to be no more than 5%, and the ICC index (e.g., ICC ≥ 0.90) of left-right volume consistency can be reported as evidence of reproducibility, thereby demonstrating that the method of this invention has stable quantitative output capability and reproducibility in practical applications.

[0047] This invention also discloses a lateral ventricular choroid plexus segmentation and quantification system based on enhanced T1-weighted magnetic resonance imaging, comprising: The image preprocessing module is used to acquire enhanced T1-weighted magnetic resonance images of the subject's brain and preprocess the images according to a preset image preprocessing strategy to obtain standardized brain images. The candidate region construction module constructs choroid plexus candidate regions based on standardized brain images and by combining preset enhancement intensity thresholds and lateral ventricle anatomical location constraints. The initial segmentation mask recognition module is used to call the trained vesicle cluster segmentation model to perform feature discrimination and segmentation on the vesicle cluster candidate region and generate an initial segmentation mask; wherein, the vesicle cluster segmentation model is a deep learning model trained based on enhanced T1 weighted magnetic resonance image samples and their corresponding vesicle cluster labeled masks. The model is configured to learn the enhancement features and spatial distribution features of vesicle clusters in enhanced T1 weighted images. The post-processing optimization module is used to perform post-processing optimization on the initial segmentation mask, remove false positive regions, and obtain the final lateral ventricle choroid plexus segmentation mask. The quantitative analysis module is used to calculate quantitative parameters of the choroid plexus based on the final lateral ventricle choroid plexus segmentation mask.

[0048] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for quantifying the choroid plexus segmentation of the lateral ventricle based on enhanced T1-weighted magnetic resonance images.

[0049] The present invention also discloses an electronic device, comprising: Memory, which stores computer programs; A processor, connected to the memory, is configured to execute the computer program to implement the steps of a method for quantifying the lateral ventricular choroid plexus segmentation based on enhanced T1-weighted magnetic resonance images.

[0050] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for quantitative segmentation of the choroid plexus of the lateral ventricle based on enhanced T1-weighted magnetic resonance imaging, characterized in that, The specific steps include: S1: Acquire enhanced T1-weighted magnetic resonance images of the subject's brain and preprocess the images according to a pre-set image preprocessing strategy to obtain standardized brain images; S2: Based on standardized brain images, and combined with preset enhancement intensity thresholds and lateral ventricle anatomical location constraints, candidate regions of the choroid plexus are constructed. S3: Use the trained venous cluster segmentation model to perform feature discrimination and segmentation on the candidate regions of venous clusters to generate an initial segmentation mask; wherein, the venous cluster segmentation model is a deep learning model trained on enhanced T1-weighted magnetic resonance image samples and their corresponding venous cluster labeled masks. The model is configured to learn the enhancement features and spatial distribution features of venous clusters in enhanced T1-weighted images. S4: Post-processing optimization of the initial segmentation mask to remove false positive regions, resulting in the final lateral ventricle choroid plexus segmentation mask; S5: Based on the final lateral ventricle choroid plexus segmentation mask, perform quantitative parameter calculation of the choroid plexus.

2. The method for segmenting and quantifying the lateral ventricle choroid plexus based on enhanced T1-weighted magnetic resonance images according to claim 1, characterized in that: In step S1, the image preprocessing steps are as follows: Bias field correction was performed on the acquired enhanced T1-weighted magnetic resonance images of the brain; Identify the foreground region of brain parenchyma, extract the intensity features of the region, and use the intensity features to perform intensity normalization processing on the image; The brain parenchyma region was extracted, the skull and its external structures were removed to form a brain parenchyma image, and a brain parenchyma mask was constructed based on the brain parenchyma image; The brain parenchyma image and brain parenchyma mask are aligned to the standard template space to form a standardized brain image.

3. The method for segmenting and quantifying the lateral ventricle choroid plexus based on enhanced T1-weighted magnetic resonance images according to claim 1, characterized in that: In step S2, the construction steps of the candidate region of the venule cluster are as follows: S2-1) Obtain a mask of the lateral ventricle region in a standardized brain image; S2-2) In standardized brain images, voxels with signal strength higher than a preset enhancement threshold are identified to obtain a set of enhanced voxels; S2-3) Perform spatial logic and operations on the enhanced voxel set and the lateral ventricle region mask to obtain a preliminary candidate voxel set; S2-4) Perform three-dimensional connected component analysis on the preliminary candidate voxel set, remove connected components with a volume smaller than the second preset volume threshold, and obtain the venule cluster candidate region.

4. The method for segmenting and quantifying the choroid plexus of the lateral ventricle based on enhanced T1-weighted magnetic resonance images according to claim 3, characterized in that: The enhanced voxel recognition strategy includes any of the following: a. Based on the statistical analysis of cerebrospinal fluid region signals, a preset enhancement threshold is set. The preset enhancement threshold is greater than or equal to the sum of the mean signal of the cerebrospinal fluid region and twice the standard deviation of the cerebrospinal fluid region signal. Voxels with signal strength greater than the preset enhancement threshold are identified as enhanced voxels. b. The percentage threshold based on whole-brain signal statistics is set as a preset enhancement threshold, wherein the percentage threshold is 10% to 15%. When identifying enhancement voxels, voxels whose signal intensity distribution is before the percentage threshold are identified as enhancement voxels.

5. The method for segmenting and quantifying the choroid plexus of the lateral ventricle based on enhanced T1-weighted magnetic resonance images according to claim 1, characterized in that: The vesicle cluster segmentation model is a three-dimensional fully convolutional neural network, including a multi-scale feature extraction module and an attention mechanism module. The multi-scale feature extraction module is used to fuse image features under different receptive fields, and the attention mechanism module is used to adaptively enhance feature channels or spatial regions related to vesicle cluster discrimination.

6. The method for segmenting and quantifying the choroid plexus of the lateral ventricle based on enhanced T1-weighted magnetic resonance images according to claim 1, characterized in that: In step S4, the post-processing optimization strategy is as follows: S4-1) Perform three-dimensional connected component analysis on the initial segmentation mask to remove isolated regions with a volume smaller than the third preset volume threshold; S4-2) Perform morphological opening and / or closing operations on the segmentation mask processed in step S4-1 to remove small burrs and fill local voids, and obtain a structurally continuous final segmentation mask as the final lateral ventricle choroid plexus segmentation mask.

7. The method for segmenting and quantifying the choroid plexus of the lateral ventricle based on enhanced T1-weighted magnetic resonance images according to claim 1, characterized in that: In step S5, the quantitative parameter calculation includes: S5-1) Calculate the total volume of the final lateral ventricle choroid plexus segmentation mask, and the lateral volumes of the left and right choroid plexuses; S5-2) Within the final lateral ventricle choroid plexus segmentation mask, the choroid plexus is distinguished into vascular-related components and non-vascular components based on the signal intensity of voxels in enhanced T1-weighted images. S5-3) Calculate the volume of vascular-related components and the volume of non-vascular components respectively, and calculate the proportion of the two to the total volume.

8. The lateral ventricle choroid plexus segmentation and quantification system based on enhanced T1-weighted magnetic resonance images according to any one of claims 1-7, characterized in that, include: The image preprocessing module is used to acquire enhanced T1-weighted magnetic resonance images of the subject's brain and preprocess the images according to a preset image preprocessing strategy to obtain standardized brain images. The candidate region construction module constructs choroid plexus candidate regions based on standardized brain images and by combining preset enhancement intensity thresholds and lateral ventricle anatomical location constraints. The initial segmentation mask recognition module is used to call the trained vesicle cluster segmentation model to perform feature discrimination and segmentation on the vesicle cluster candidate region and generate an initial segmentation mask; wherein, the vesicle cluster segmentation model is a deep learning model trained based on enhanced T1 weighted magnetic resonance image samples and their corresponding vesicle cluster labeled masks. The model is configured to learn the enhancement features and spatial distribution features of vesicle clusters in enhanced T1 weighted images. The post-processing optimization module is used to perform post-processing optimization on the initial segmentation mask, remove false positive regions, and obtain the final lateral ventricle choroid plexus segmentation mask. The quantitative analysis module is used to calculate quantitative parameters of the choroid plexus based on the final lateral ventricle choroid plexus segmentation mask.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: Memory, which stores computer programs; A processor, connected to the memory, is configured to perform the steps of the method as described in any one of claims 1 to 7 by executing the computer program.

Images