Fetal MR brain radial automatic measurement and medical diagnosis method and system

Abstract

The invention provides a fetal MR brain radial automatic measurement and medical diagnosis method and system, and relates to the technical field of medical imaging, and the method comprises the steps: obtaining an original sample set; aligning the coordinate points with corresponding anatomical structures in the fetal MR brain image to obtain an enhanced sample set; performing sorting and classification processing according to the enhanced sample set to obtain a classification sample set; performing sample expansion processing according to the classification sample set to obtain an expanded sample set; performing model construction processing according to the expanded sample set to obtain an automatic radial line measurement model; and inputting the fetal MR brain image data of the patient to be diagnosed into the automatic measurement model for labeling, and generating a brain disease diagnosis conclusion. By means of processing means such as automatic radial line marking, space alignment and remapping, symmetry analysis and the like, the fetal brain radial line can be accurately extracted, automatic disease diagnosis is carried out, the measurement precision is greatly improved, and the method can adapt to changes of different fetal development stages and image noise interference.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a method and system for automated measurement and medical diagnosis of fetal MR brain diameters. Background Art

[0002] At present, fetal MR imaging is widely used in medical diagnosis, especially in the monitoring of brain development and disease diagnosis. Using fetal MR imaging to measure key diameters such as the biparietal diameter of the brain, the biparietal diameter of the skull, the frontoparietal diameter of the brain, the frontoparietal diameter of the skull, the width of the trigone of the lateral ventricle, the width of the posterior cranial fossa, the transverse diameter of the cerebellum, and the cavity of the transparent septum is an important method for evaluating fetal brain development and early detection of brain diseases. However, traditional diameter measurement methods rely on manual annotation and measurement, which is not only time-consuming and laborious, but also easily affected by operator experience and subjective factors, resulting in inaccurate and inconsistent measurement results. In addition, image analysis methods in the prior art generally rely on traditional image processing algorithms, such as image segmentation and registration, but these methods usually cannot fully consider the complexity of fetal brain development and the spatial relationship between various anatomical features. The limitations of the prior art are that the automatic recognition ability of complex anatomical structures is insufficient, the diverse anatomical features of the fetal brain at different developmental stages cannot be efficiently processed, and it is difficult to accurately measure images of different quality and noise.

[0003] In view of the above problems, the present invention proposes a method and system for automatic measurement and medical diagnosis of fetal MR brain diameters. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for automatic measurement and medical diagnosis of fetal MR brain diameters to improve the above problems. In order to achieve the above purpose, the technical solution adopted by the present invention is as follows: In a first aspect, the present application provides a method for automated measurement and medical diagnosis of fetal MR brain diameters, comprising: Acquiring an original sample set, wherein the original sample set includes historical fetal MR brain image data with manually annotated diameters and corresponding brain diagnosis results; Based on the original sample set, the manually annotated radial data are converted into coordinate points in three-dimensional space, the coordinate points are aligned with corresponding anatomical structures in the fetal MR brain image, and the radial data are embedded into the structural layer of the image by creating a virtual image layer, so as to obtain an enhanced sample set; Sorting and classifying the enhanced sample set, analyzing the developmental rules of the brain structure of fetuses of different age groups, arranging the samples in ascending order according to the fetal age, and dividing the samples into age groups of multiple developmental stages based on the changing trends of brain structure characteristics, to obtain a classified sample set; Performing sample expansion processing according to the classified sample set, generating new samples based on the change pattern of brain structure at each developmental stage, and combining random noise vectors to simulate interference in actual imaging, thereby obtaining an expanded sample set; Performing model construction processing according to the expanded sample set, using a deep convolutional neural network to transform and fuse the radial features in the image data with the geometric relationship of the adjacent anatomical structures, to obtain an automatic radial measurement model; The fetal MR brain image data of the patient to be diagnosed is input into the automatic measurement model for annotation processing to obtain the radial measurement result, and the symmetry analysis is performed based on the radial measurement result to generate a brain disease diagnosis conclusion.

[0005] In a second aspect, the present application also provides a fetal MR brain diameter automated measurement and medical diagnosis system, comprising: An acquisition module, used for acquiring an original sample set, wherein the original sample set includes historical fetal MR brain image data with manually annotated diameters and corresponding brain diagnosis results; A conversion module, based on the original sample set, converts the manually annotated radial data into coordinate points in three-dimensional space, aligns the coordinate points with corresponding anatomical structures in the fetal MR brain image, and embeds the radial data into the structural layer of the image by creating a virtual image layer, thereby obtaining an enhanced sample set; A classification module, used to sort and classify the enhanced sample set, analyze the developmental rules of the brain structure of fetuses of different age groups, arrange the samples in ascending order according to the fetal age, and divide the samples into age groups of multiple developmental stages based on the changing trend of brain structure characteristics, so as to obtain a classified sample set; An expansion module is used to perform sample expansion processing according to the classified sample set, generate new samples based on the change pattern of brain structure in each developmental stage, and combine random noise vectors to simulate interference in actual imaging to obtain an expanded sample set; A construction module is used to perform model construction processing according to the expanded sample set, and use a deep convolutional neural network to transform and fuse the radial features in the image data with the geometric relationship of the adjacent anatomical structures to obtain an automatic radial measurement model; The diagnosis module is used to input the fetal MR brain image data of the patient to be diagnosed into the automatic measurement model for annotation processing to obtain the radial measurement result, and perform symmetry analysis based on the radial measurement result to generate a brain disease diagnosis conclusion.

[0006] The beneficial effects of the present invention are: The present invention can accurately extract fetal brain diameters and perform automated disease diagnosis through processing methods such as automatic diameter labeling, spatial alignment and remapping, and symmetry analysis, greatly improving measurement accuracy and being able to adapt to changes in different fetal development stages and image noise interference.

[0007] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or be understood by implementing the embodiments of the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0009] Figure 1 A schematic diagram of a process for automated measurement and medical diagnosis of fetal MR brain diameters in an embodiment of the present invention; Figure 2 It is a structural diagram of the automatic measurement and medical diagnosis system of fetal MR brain diameter described in an embodiment of the present invention.

[0010] Markings in the figure: 1. Acquisition module; 2. Conversion module; 21. First conversion unit; 22. Second conversion unit; 23. Third conversion unit; 24. Fourth conversion unit; 3. Classification module; 31. First classification unit; 32. Second classification unit; 33. Third classification unit; 34. Fourth classification unit; 4. Expansion module; 41. First expansion unit; 42. Second expansion unit; 43. Third expansion unit; 44. Fourth expansion unit; 5. Construction module; 51. First construction unit; 52. Second construction unit; 53. Third construction unit; 54. Fourth construction unit; 6. Diagnosis module; 61. First diagnosis unit; 62. Second diagnosis unit; 63. Third diagnosis unit; 64. Fourth diagnosis unit. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0012] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0013] Embodiment 1: This embodiment provides a method for automated measurement and medical diagnosis of fetal MR brain diameters.

[0014] See also Figure 1 , the figure shows that the method includes steps S100 to S600.

[0015] Step S100, obtaining an original sample set, the original sample set including historical fetal MR brain image data with manually annotated diameters and corresponding brain diagnosis results; It is understandable that in this step, a high-quality sample set is constructed by collecting historical fetal MR brain imaging data with manually annotated diameters and their corresponding brain diagnosis results. Key diameter annotations, such as the biparietal diameter of the brain, the biparietal diameter of the skull, the frontal-parietal diameter of the brain, the frontal-parietal diameter of the skull, the width of the triangular area of ​​the lateral ventricle, the width of the posterior cranial fossa, the transverse diameter of the cerebellum, the transparent compartment cavity, etc., are important indicators in the analysis of fetal brain development, and they provide an accurate measurement of the state of fetal brain development. These diameter annotations are calibrated by artificial experts based on the image data to ensure the accuracy and reliability of the data. In addition, the diagnostic results in the image data also provide a basis for judging the health status of the fetal brain, so that these image data can not only be used for the measurement of diameters, but also provide support for subsequent pathological analysis and model training. By constructing this sample set with detailed annotations, a strong training data foundation can be provided for subsequent deep learning models.

[0016] Step S200: based on the original sample set, convert the manually annotated radial data into coordinate points in three-dimensional space, align the coordinate points with the corresponding anatomical structures in the fetal MR brain image, and embed the radial data into the structural layer of the image by creating a virtual image layer, so as to obtain an enhanced sample set; It should be noted that this step greatly improves the consistency and spatial accuracy of the data by embedding the radial data into the three-dimensional image and aligning it with the anatomical structure, providing an enhanced sample set with practical application value for the subsequent training of the deep learning model.

[0017] Step S300, sorting and classifying the enhanced sample set, analyzing the developmental rules of the brain structure of fetuses of different age groups, arranging the samples in ascending order according to the fetal age, and dividing the samples into age groups of multiple developmental stages based on the changing trends of brain structure characteristics, to obtain a classified sample set; It is understandable that the structure of the fetal brain undergoes significant changes at different stages of development, especially key diameters such as the biparietal diameter of the brain, the biparietal diameter of the skull, and the frontoparietal diameter of the brain will change as the fetus grows. By quantitatively analyzing these changing patterns, the characteristic changes of each stage can be captured. Secondly, in traditional medical image analysis, sample sorting often relies on manual or random arrangement that does not follow biological laws, which can easily lead to a lack of consistency between samples in the data set. By arranging the samples in ascending order by fetal age, it is ensured that the data at each stage can truly reflect the natural development process of the fetal brain. This step ensures the efficient organization of the data through scientific sorting and classification, so that the model can fully learn the characteristics of each developmental stage.

[0018] Step S400, performing sample expansion processing according to the classified sample set, generating new samples based on the change pattern of brain structure at each developmental stage, and combining random noise vectors to simulate interference in actual imaging to obtain an expanded sample set; First, new samples are generated based on the changing patterns of brain structures at each developmental stage. This process models the changing trends of brain anatomical features at each developmental stage. This modeling method can accurately capture the changing laws of brain structures at each developmental stage and provide an accurate physiological basis for the subsequent generation of new samples. Secondly, random noise vectors are combined to simulate interference in actual imaging. This process is to enhance the robustness of the model. In actual clinical practice, fetal MR images are often affected by many factors, such as limitations of imaging equipment, motion artifacts, and noise. By introducing random noise vectors, the impact of these interference factors on image data can be simulated, thereby generating noisy samples. The addition of this noise enables the model to learn how to make accurate measurements in an environment with greater interference, improving the stability and adaptability of the model in practical applications.

[0019] Step S500, constructing a model based on the expanded sample set, using a deep convolutional neural network to transform and fuse the radial features in the image data with the geometric relationship of the adjacent anatomical structures to obtain an automatic radial measurement model; It is understandable that deep convolutional neural networks can not only process the complex features of image data, but also effectively integrate the geometric relationship between diameters and anatomical structures in the model. This feature fusion based on geometric relationships significantly improves the measurement accuracy and robustness of the model, ensuring that the automatic measurement model can still operate stably in the face of different fetal development stages or changes in image quality.

[0020] Step S600: inputting the fetal MR brain image data of the patient to be diagnosed into the automatic measurement model for annotation processing to obtain the radial measurement result, and performing symmetry analysis based on the radial measurement result to generate a brain disease diagnosis conclusion.

[0021] It should be noted that this step completely uses computer algorithms to achieve automatic measurement and symmetry analysis of fetal brain diameters, reducing human intervention and being able to efficiently and accurately extract useful information from complex fetal MR images. Automated diameter labeling and symmetry analysis not only improves the reliability of diagnosis, but also operates stably under different image qualities and developmental stages, providing strong support for clinical diagnosis.

[0022] Further, step S200 includes step S210 to step S240.

[0023] Step S210: performing encoding processing according to the original sample set, fitting the manually marked radial data by using a B-spline curve fitting algorithm to obtain a curvature code, converting the curvature code into the value of the image data pixel point, and generating a radial pixel value layer; First, the B-spline curve fitting algorithm is used to accurately fit the radial data. B-spline is a mathematical tool commonly used for data smoothing and curve fitting, and is suitable for processing high-dimensional data and complex curve structures. In fetal MR images, the radial lines are not simple straight lines or regular curves, so the use of B-spline curve fitting can efficiently capture complex curve shapes, smooth the data, and reduce noise interference. In this way, the B-spline curve can not only fit the smooth curve of each radial line, but also generate an accurate curvature code that reflects the local geometric characteristics of the radial line in the image. This code quantifies the degree of curvature of each radial point and provides important information about the shape and changes of the radial line.

[0024] Next, the generated curvature codes are converted into the values ​​of the image data pixels to form a radial pixel value layer. This conversion process maps the curvature codes to the pixel level of the image, so that the radial features can be directly used for subsequent image processing and model training. In this process, the geometric features of the radials (such as curvature) are accurately embedded in the image data, providing richer feature information for subsequent automatic measurement and anatomical structure analysis. By embedding these curvature codes into the pixel layer of the image, the geometric information related to the radials can be directly learned from the image data, and the spatial distribution of the anatomical structure can be better captured. Through B-spline curve fitting and the generation of curvature codes, the morphological changes of the radials in the image can be accurately represented, so that the radial features are not limited to simple point marks, but are converted into continuous spatial features that can be learned by computer models.

[0025] Step S220, coordinate conversion processing is performed according to the radial pixel value layer, and the manually marked radial data is converted into three-dimensional coordinate points using a cubic interpolation algorithm to obtain a radial annotation point set in a three-dimensional space; It should be noted that cubic interpolation can effectively smooth and interpolate given discrete data points, and is suitable for processing continuity and smoothness problems in image data. Compared with other interpolation methods (such as linear interpolation), cubic interpolation can provide higher smoothness and more accurate spatial mapping. In this process, cubic interpolation infers the exact position of each pixel point in three-dimensional space based on the radial pixel value layer in the image, starting from the original annotation point. This conversion process ensures that the data transitions smoothly from the two-dimensional pixel coordinate system to the three-dimensional space coordinate system, so that the radial data not only has the accuracy of spatial position, but also can be accurately represented in three-dimensional space. Through this conversion, the radial annotation data in the two-dimensional image is matched with its position in three-dimensional space, which can better reflect the relative positions of various anatomical structures in the fetal brain. Step S230, performing non-rigid image registration based on the mutual information method, aligning the three-dimensional coordinate point set with the brain gray matter, white matter and ventricle anatomical structure in the fetal MR image to obtain an aligned data set; It can be understood that the mutual information method is a technique commonly used in image registration and is suitable for alignment between multimodal images (such as images of different image types or acquired by different devices). The mutual information method measures the similarity between two images by calculating the statistical dependence between them. The commonly used information measure is mutual information, which measures the relationship between the pixel intensities in the two images. In this step, by comparing the three-dimensional coordinate point set with the anatomical structure in the fetal MR image, the mutual information method can find the best spatial transformation so that the spatial positions of the radial annotation points and the anatomical structures (such as gray matter, white matter and ventricles) are optimally aligned. Unlike rigid registration (such as rigid rotation and displacement), non-rigid image registration can handle deformation, distortion and local differences in images. The anatomical structure of the fetal brain may have morphological differences at different stages of development. Through non-rigid registration, the algorithm can flexibly adjust the registration transformation according to the local morphological changes between images, so that the anatomical structure in the image data can be accurately aligned with the annotated radial data. This processing method is particularly suitable for fetal MR images. Since the fetal brain anatomical structure changes greatly at different ages, non-rigid registration can better adapt to these changes and ensure the accuracy of data alignment. Through this registration step, the spatial consistency between the radial annotation data and the anatomical structure is ensured, providing reliable basic data for subsequent analysis and automatic measurement.

[0026] Step S240, performing virtual image layer creation processing according to the aligned data set, segmenting the structure in the fetal brain image by using threshold segmentation and region growing method, and embedding the aligned radial data into the gray matter, white matter or ventricle layer to obtain an enhanced sample set.

[0027] It should be noted that the threshold segmentation method divides the image data into different areas by setting a pixel value threshold. In fetal brain MR images, the pixel values ​​of gray matter, white matter, and ventricles have obvious intensity differences from the background and other brain areas. The threshold segmentation method can automatically identify and extract these areas. By selecting a suitable threshold, the various anatomical structures in the fetal brain image (such as gray matter, white matter, ventricles, etc.) can be accurately segmented from the entire image. This process can effectively distinguish the different structures of the fetal brain and provide clear structural boundaries for subsequent data processing and analysis.

[0028] Next, the region growing method is used to further refine the segmentation results. The region growing method is an image segmentation algorithm based on seed point expansion. It starts from an initial seed point and gradually expands the seed point area to the surrounding area by calculating the similarity between the neighboring pixels and the seed point until a certain stopping condition is reached. In the anatomical structure segmentation of fetal brain images, the region growing method can capture the boundaries of structures such as gray matter, white matter, and ventricles in more detail, especially when irregular shapes or blurred boundaries appear in the image. The region growing method can improve the accuracy of segmentation and ensure that different brain regions are accurately calibrated.

[0029] Finally, the aligned radial data are embedded into the gray matter, white matter, or ventricle layers. This process ensures accurate spatial alignment of the radial data by embedding the calculated three-dimensional coordinate point set (i.e., radial annotation points) into these segmented structural layers. This embedding not only ensures the spatial consistency of the radial data, but also further clarifies the relationship between the radial annotation and the anatomical structure.

[0030] Further, step S300 includes step S310 to step S340.

[0031] Step S310, performing age sorting processing according to the enhanced sample set, quantitatively analyzing the changes of the diameters in the fetal MR images with age, and arranging the samples in ascending order according to the fetal age based on the growth law of the anatomical structure in the image data, to obtain a sorted sample set; First, the changes in diameters in fetal MR images with age are quantitatively analyzed. This step extracts key diameter data from fetal images (such as biparietal brain diameter, biparietal skull diameter, etc.) and combines them with fetal age for quantitative analysis to reveal the laws of fetal brain development at different ages. The structure and diameter of the fetal brain show a certain growth pattern over time, and these changes are usually nonlinear. Through this quantitative analysis, the growth trend of the fetal brain structure and the change pattern of anatomical characteristics can be accurately captured. For example, diameters such as ventricular width or cerebellar transverse diameter will gradually increase with fetal age, and this process can be quantified and tracked through regression models or other mathematical methods. This analysis provides a basis for the sorting of samples, so that each sample can correspond to an accurate developmental stage.

[0032] Next, the samples are arranged in ascending order according to the fetal age based on the growth law of the anatomical structure in the image data. In this process, the arrangement of the samples depends not only on their age labels, but also on the growth law of the anatomical structure. By analyzing the measurement data of the fetal brain anatomical structure (such as gray matter, white matter, etc.), it can be further ensured that the samples are arranged in the true developmental order. For example, by analyzing the growth patterns of structures such as the ventricles and cerebral cortex at different ages, the age group corresponding to each sample can be more accurately determined and sorted in ascending order. This sorting not only ensures the consistency of the age labels, but also ensures that the model takes into account the natural development trajectory of the anatomical structure during training.

[0033] Step S320, performing developmental stage division processing according to the sorted sample set, performing linear regression fitting on the measured data of the brain anatomical structure in the fetal MR image, obtaining a relationship between the age and the anatomical structure corresponding to each sample, and arranging the samples in ascending order of age according to the calculated fitting values, to obtain a sorted sample set; It should be noted that the fetal brain anatomical structure shows certain linear or near-linear changes during development. By using the linear regression method, the relationship between the measured data of these anatomical structures and the fetal age can be modeled to obtain a mathematical expression that can reflect the changing trend of the anatomical structure with age. The advantage of linear regression is that it can generate an accurate fitting curve for each structural feature, describing its regularity with age. The specific expression is: ; in, Measurements representing anatomical features of the fetal brain; Indicates the age of the fetus in gestational weeks (weeks); are the coefficients in the polynomial regression, representing the nonlinear relationship between anatomical characteristics and fetal age; is the coefficient of the differential term; is the error term; Represents the order of each term in polynomial regression; Indicates the maximum order of polynomial regression; It indicates the rate or acceleration of change of fetal age over time; Represents the effect of fetal age on anatomical features.

[0034] Next, the samples are sorted in ascending order by age based on the calculated fitting values. After obtaining the fitting relationship for each anatomical structure, these fitting values ​​can be used to further confirm the developmental stage of the sample and ensure that the age sorting of the data is more refined. By sorting the samples in ascending order according to the age corresponding to the fitting results, it can be ensured that each sample accurately corresponds to its developmental stage. This processing method not only helps to ensure the accuracy of the age sorting, but also reflects the growth pattern of the fetal brain structure at different developmental stages, thereby providing more biologically reasonable data input for subsequent model training.

[0035] Step S330, performing developmental stage division processing according to the sorted sample set, analyzing the trend of the anatomical characteristics of each sample changing with age, and dividing the sample into multiple developmental stages based on the change amplitude of the anatomical characteristics, to obtain a developmental stage classification sample set; Specifically, this step first analyzes the trend of the anatomical features of each sample with age, which involves quantitative analysis of the key anatomical features of each fetal sample. Different structures of the fetal brain show different patterns of change as the development progresses. Some anatomical features (such as ventricle width, cortical thickness, etc.) change significantly in the early development stage, but tend to be stable in the later stage. By analyzing the trend of these features with age, the change pattern of each sample at its specific developmental stage can be revealed. At this time, the analysis method may include regression analysis, trend detection and other techniques to help understand the growth rate and change pattern of these features. Next, the samples are divided into multiple developmental stages based on the magnitude of the change in anatomical features. This step determines the changes in the features of each anatomical feature at different developmental stages by calculating the magnitude of the change in each anatomical feature with age. By comparing the magnitude of the change in the features of each developmental stage, the samples can be divided into multiple developmental stages with similar anatomical features. This division method is not only based on age, but also takes into account the actual developmental changes of the fetal brain to ensure that the samples at each stage are physiologically reasonable.

[0036] Step S340: Perform feature extraction processing on the developmental stage classification sample set, extract the anatomical features of each developmental stage and map the features to standard spatial positions to obtain a classification sample set.

[0037] It can be understood that the classification sample set obtained by the feature extraction and standard space mapping processing in this step not only retains the spatial consistency of the anatomical features of the fetal brain at each developmental stage, but also makes these features comparable in the standard space.

[0038] Further, step S400 includes step S410 to step S440.

[0039] Step S410, extracting and processing the change pattern of the developmental stage characteristics according to the classified sample set, performing differential calculation on the change trend of the brain anatomical characteristics of the samples at each developmental stage, and modeling the changes of the brain structure at different developmental stages in combination with a nonlinear regression model, to obtain the change pattern of the anatomical characteristics at each developmental stage; First, it is necessary to perform differential calculations on the anatomical features of the samples in each developmental stage. Differential calculation is a basic mathematical tool for analyzing data change trends, which can capture the rate and trend of change of anatomical features in continuous age intervals. Specifically, for each anatomical feature in the classified sample set, differential calculation can be used to obtain its discrete rate of change with age, thereby quantifying the subtle changes in anatomical features within the developmental stage.

[0040] After completing the differential analysis, in order to more comprehensively describe the change pattern of anatomical features, it is necessary to combine the differential results with nonlinear regression models. Nonlinear regression models are suitable for dealing with complex physiological changes. For example, brain anatomical features are usually not simple linear growth during fetal development, but may show exponential growth, logarithmic change or quadratic parabolic trend. By fitting nonlinear regression models (such as polynomial regression, logistic model or exponential model), mathematical expressions of anatomical features changing with age can be generated. These expressions can not only describe the change law of anatomical features in a specific developmental stage, but also predict the possible changes of anatomical features at points that are not observed in this stage. For significant differences in anatomical features between developmental stages, segmented regression models can also be used to accurately capture these stage changes. The anatomical feature change pattern generated at each developmental stage specifically describes the law of growth, acceleration or slowdown of anatomical features during the developmental stage.

[0041] Step S420, performing sample generation processing according to the feature change pattern of the development stage, matching the anatomical features of the development stage with the shape context template by applying a deformation method based on shape context, calculating the transformation matrix of the morphological deformation, and generating a new sample set; It can be understood that in this step, the shape context is used to describe the shape of the fetal brain anatomical features, and the morphological deformation of the sample is calculated by this method. The core idea of ​​the shape context is to define a local coordinate system for each key point of the anatomical structure, and to capture the geometric features of the entire shape by calculating the relative relationship between these key points. The shape context template is used as a standard template. The anatomical features to be matched are aligned with the template through the shape descriptor to build a spatial framework, thereby ensuring the consistency of shape changes between different developmental stages. Next, the anatomical features of the developmental stage are matched with the shape context template by the deformation method. At this time, the relative positions between the key points will be compared with the positions in the standard template, and a transformation matrix will be obtained by optimizing the matching process. The transformation matrix describes the morphological changes of anatomical features at different developmental stages, thereby converting the original anatomical features into a morphology that conforms to the target developmental stage. This transformation matrix includes not only basic transformations such as translation, rotation, and scaling, but also more complex morphological deformations, such as bending or twisting, to ensure that the new sample can reasonably reflect the morphological changes of the fetal brain anatomical structure at different developmental stages. Finally, a new set of samples is generated, which not only conforms to the changing trend of anatomical features, but also covers the differences in anatomical structures at different developmental stages.

[0042] Step S430: Perform noise simulation processing on the new sample set, introduce a random noise model based on Poisson distribution to perturb the newly generated samples, and simulate the noise influence in actual imaging to obtain a new sample set with noise; It should be noted that Poisson distribution is a probability distribution used to describe the number of times a counting event occurs per unit time or per unit space, and is suitable for simulating photon counting in images, noise during scanning, etc. Fetal MR images may be affected by a variety of noise sources during the imaging process, such as equipment noise, motion artifacts, random errors during scanning, etc. In order to simulate these noise sources, the noise model based on Poisson distribution can generate noise effects similar to the actual imaging environment. Specifically, the Poisson noise model introduces random perturbations to the values ​​of each pixel so that the generated image data presents noise characteristics that conform to the actual imaging characteristics. Then, the process of noise simulation is to perturb the newly generated samples, that is, to add random noise based on Poisson distribution to each pixel value of the new sample. This perturbation will make the change of each pixel value have a certain randomness according to the characteristics of Poisson distribution, but it conforms to the statistical characteristics of noise during imaging. In this way, an image data set containing noise can be generated. These noisy samples can more realistically simulate the complex environment of fetal MR images in actual clinical practice and enhance the model's adaptability to noise interference.

[0043] Step S440: weightedly combine the original samples and the new sample set with noise to obtain an expanded sample set.

[0044] It can be understood that the expanded sample set includes both high-quality original samples and potential noise interference in noisy samples, ensuring the comprehensiveness of the data set.

[0045] Further, step S500 includes step S510 to step S540.

[0046] Step S510: extracting radial features and anatomical structure features according to the expanded sample set, decoupling and extracting radial features from the fetal MR images and the brain anatomical structures through a multi-channel convolutional neural network, and obtaining a radial-anatomical feature map; Specifically, in this step, each channel of the multi-channel convolutional neural network is responsible for extracting different types of features in the image (such as diameter, gray matter, white matter, ventricle and other anatomical structures). By using multiple convolution channels, the network can learn and extract multi-level information in the image data at different levels, thereby realizing the joint extraction of different types of anatomical structures and diameter features. The key to this process is decoupled extraction, that is, extracting diameter features and anatomical structure features separately through multiple convolution layers to avoid mixing or information interference between them. Specifically, the network will perform multiple convolution operations on the image data, first extracting low-level features (such as edges, textures, etc.) from the original data, and then gradually learning more complex features (such as the spatial relationship between diameter and brain anatomical structure) through deeper convolution layers. Each channel can focus on different features, for example, one channel focuses on extracting diameter features, while another channel focuses on extracting the geometric morphology of brain anatomical structures. Through decoupled extraction, the network can more accurately understand and learn the relationship between anatomical features and diameters. The radial-anatomical feature map not only contains the structural information of the fetal brain, but also can clearly reflect the spatial correspondence between the anatomical structure and the radial features.

[0047] Step S520: performing spatial dependency modeling processing according to the radial-anatomical feature map, and modeling the spatial relationship between the radial features and the anatomical structure features by using a graph convolutional network based on anatomical constraints to obtain a spatial dependency feature map; In this step, the role of the graph convolution network is to represent the anatomical features and radial features as nodes in the graph structure through graph convolution operations. Each node represents an anatomical structure or radial feature, and the edges between nodes represent the spatial relationship or mutual dependence between them. The graph convolution network uses the information of these graph edges to perform convolution operations to model the spatial dependence between anatomical structures and radial features. This method can capture the adjacency of anatomical structures and the changing trend of radial features. Especially in the complex fetal brain anatomical structure, the graph convolution network can effectively transmit information through the graph convolution layer to ensure that the mutual connection between features is fully utilized. In order to ensure that the spatial modeling process conforms to the anatomical laws, anatomical constraints are introduced into the graph convolution network. Anatomical constraints are based on known anatomical knowledge and the physiological characteristics of the fetal brain to ensure that the spatial relationship learned by the model is consistent with the relative position and morphological changes of the actual anatomical structure. For example, the ventricles should be near the midline of the brain, the relationship between gray matter and white matter should conform to the known developmental laws, and the radial features should have a clear spatial correspondence with these anatomical structures. By incorporating these constraints into the design of the graph convolutional network, the network can ensure that the learned anatomical structures and radial features are physiologically reasonable during the spatial modeling process, avoiding the generation of spatial relationships that are inconsistent with anatomical laws.

[0048] Step S530, performing spatial remapping processing according to the spatial dependency feature map, by adjusting the proportion, morphology and spatial position of the anatomical structure and optimizing the geometric relationship between the radial feature and the adjacent anatomical structure, ensuring the spatial alignment of the anatomical structure and the radial at different developmental stages, and obtaining a geometrically optimized feature map; It should be noted that the core goal of spatial remapping is to accurately align the anatomical structures and radial features of different developmental stages in space. Since the fetal brain anatomical structures show different growth patterns at different developmental stages, the positions, proportions and morphologies of these structures in space will change. Therefore, how to unify the anatomical structures of these different stages into a standard space for comparison and ensure that the geometric relationship between the radial features and the anatomical structures is consistent is the key task of this step. First, for the anatomical structures of each developmental stage, the proportions, morphologies and spatial positions need to be optimized and adjusted. In order to unify the imaging data of different stages into a standard space, the proportions of the anatomical structures may need to be scaled according to the different developmental stages (for example, the ventricles may be smaller in early development and need to be enlarged or reduced according to their growth laws). Morphological adjustment involves the shape changes of anatomical structures. For example, the morphology of the ventricles may change with the development of the fetus. These morphologies are adjusted through deformation technology to make them conform to physiological laws. The adjustment of spatial position is mainly achieved through registration technology to align the anatomical structures of different developmental stages to ensure that they are in the same reference frame in the standard space. Secondly, another important task of this step is to optimize the geometric relationship between the diameter features and the anatomical structure. The diameter features (such as the biparietal diameter of the brain, the biparietal diameter of the skull, etc.) and the anatomical structures (such as gray matter, white matter, ventricles, etc.) are interrelated in space, and the measured values ​​of the diameter features usually depend on the position and morphology of the anatomical structure in which they are located. At different stages of development, the geometric relationship of these anatomical structures may change, so it is necessary to optimize the geometric relationship between the diameter features and their adjacent anatomical structures to ensure their spatial alignment. For example, when adjusting the measurement value of the biparietal diameter of the brain, it is necessary to take into account the changes in the skull at the same time to ensure that the relative position of the diameter features and the corresponding anatomical structures remains consistent.

[0049] Step S540, deep learning training is performed according to the geometric optimization feature map, the automatic diameter measurement task and the anatomical structure segmentation task are trained simultaneously by adopting a multi-task learning framework, and the parameters of the network are optimized using a joint loss function to construct an automatic diameter measurement model.

[0050] The joint loss function is: in, represents the joint loss function; Indicates the sample serial number; represents the total number of samples; Indicates the actual measured value With the predicted value The error between Represents model output For input data The gradient of Indicates the actual label With the model prediction label The difference between is a regularization term used to constrain the learnable parameters of the model ; Represents the weight coefficient of the gradient penalty term; Represents the weight coefficient of the regularization term.

[0051] Multi-task learning enables the model to share knowledge, improves learning efficiency, and improves the accuracy of diameter measurement while processing anatomical structures. The use of a joint loss function ensures the balance of the two tasks, allowing the model to achieve good convergence and optimization between different tasks. The final constructed automatic diameter measurement model can not only provide high-precision diameter measurement results, but also effectively process complex fetal brain imaging data. It has strong robustness and generalization capabilities, and can cope with diverse imaging situations in actual clinical practice.

[0052] Further, step S600 includes step S610 to step S640.

[0053] Step S610, inputting the fetal MR brain image data of the patient to be diagnosed into the automatic measurement model, using the convolutional neural network to automatically mark the radial lines in the image, extracting the key radial line features and predicting the radial line position through the regression network to obtain the radial line marking result; It can be understood that after the convolutional neural network processing, the key diameter features in the image are extracted. The key diameter features include the various important diameters of the fetal brain, which are important indicators for brain health assessment. During the extraction process, the network not only identifies the location of the diameter, but also determines important information such as their direction and size. The quality of feature extraction directly determines the accuracy of subsequent diameter prediction. After feature extraction, the actual position of the diameter is predicted using a regression network. In this step, the regression network receives the extracted diameter features and outputs the specific position of each diameter through regression calculation. Unlike traditional classification problems, the goal of regression tasks is to accurately predict continuous values, such as the actual length, angle, or position coordinates of the diameter.

[0054] Step S620, performing symmetry analysis based on the diameter marking result, by comparing the diameter difference between the left and right cerebral hemispheres and combining the geometric constraints of anatomical features to quantify the symmetry deviation, and obtaining the symmetry analysis result; Specifically, this step first uses the diameter annotation results to automatically extract the diameter features of the left and right cerebral hemispheres and calculate their differences in size, shape, and position. Then, by combining the geometric constraints of anatomical features, ensure that the analysis conforms to normal biological laws, such as the ventricles and other structures should maintain a certain degree of symmetry. By calculating the absolute value or standardized difference between the diameters of the left and right cerebral hemispheres, the symmetry deviation is quantified. If the deviation exceeds the preset threshold, it may indicate pathological asymmetry. Ultimately, the symmetry analysis results obtained provide a reliable basis for subsequent disease diagnosis.

[0055] Step S630, performing pathological feature derivation processing according to the symmetry analysis result, by comparing the symmetry analysis result with the known pathological anatomical change pattern, using a support vector machine to automatically determine whether there is pathological asymmetry, and obtaining a pathological feature derivation result; Specifically, the symmetry analysis results are first compared with known pathological anatomical change patterns. Pathological change patterns are based on typical anatomical features summarized in historical cases, clinical data, and medical literature, and can reflect common patterns in brain diseases (such as brain dysplasia, ventricular expansion, asymmetric hemispheric development, etc.). By comparing the symmetry deviation of the current fetal brain with known pathological patterns, it is possible to infer whether there are possible pathological changes. Then, in order to accurately determine whether there is pathological asymmetry, a supervised learning algorithm, support vector machine, is introduced. Support vector machine is a classification algorithm that can classify by learning features in sample data. Specifically, support vector machine is trained to distinguish between normal brain symmetry and pathological asymmetry. By taking the symmetry analysis results as input, it is possible to automatically determine whether the fetal brain shows pathological asymmetry, and by finding the optimal hyperplane, the data is divided into normal and abnormal categories, and the classification results are output. The final pathological feature derivation result indicates whether there is pathological asymmetry. If the support vector machine determines that the symmetry deviation exceeds the normal range, it will output the result of "pathological asymmetry exists", indicating that there may be abnormal brain development or disease. Conversely, if the asymmetry deviation is within the normal range, it will output the conclusion of "no pathological abnormality found".

[0056] Step S640: Input the pathological feature derivation result into a preset brain disease classification model, and based on the pattern of the pathological features and the disease samples annotated in the training data, deduce whether there is a lesion in the fetal brain and generate a brain disease diagnosis conclusion.

[0057] It should be noted that the brain disease classification model has been trained on a large amount of historical case data and can understand the relationship between different pathological features and specific diseases. By comparing the input pathological features with the disease patterns in the training data, the model automatically infers whether there is a lesion in the fetal brain.

[0058] Embodiment 2: like Figure 2 As shown, this embodiment provides a fetal MR brain diameter automatic measurement and medical diagnosis system, the system includes: An acquisition module 1 is used to acquire an original sample set, wherein the original sample set includes historical fetal MR brain image data with manually annotated diameters and corresponding brain diagnosis results; The conversion module 2 converts the manually annotated radial data into coordinate points in three-dimensional space based on the original sample set, aligns the coordinate points with the corresponding anatomical structures in the fetal MR brain image, and embeds the radial data into the structural layer of the image by creating a virtual image layer, thereby obtaining an enhanced sample set; Classification module 3 is used to sort and classify the enhanced sample set, analyze the development rules of the brain structure of fetuses at different ages, arrange the samples in ascending order according to the fetal age, and divide the samples into multiple age groups of developmental stages based on the change trend of brain structure characteristics to obtain a classified sample set; An expansion module 4 is used to perform sample expansion processing according to the classified sample set, generate new samples based on the change pattern of brain structure in each developmental stage, and combine random noise vectors to simulate interference in actual imaging to obtain an expanded sample set; Construction module 5 is used to perform model construction processing according to the expanded sample set, and use a deep convolutional neural network to transform and fuse the radial features in the image data with the geometric relationship of the adjacent anatomical structures to obtain an automatic radial measurement model; The diagnosis module 6 is used to input the fetal MR brain image data of the patient to be diagnosed into the automatic measurement model for annotation processing to obtain the radial measurement results, and perform symmetry analysis based on the radial measurement results to generate a brain disease diagnosis conclusion.

[0059] In a specific implementation disclosed in the present application, the conversion module 2 includes: The first conversion unit 21 is used to perform encoding processing according to the original sample set, fit the manually marked radial data by using a B-spline curve fitting algorithm to obtain a curvature code, convert the curvature code into the value of the image data pixel point, and generate a radial pixel value layer; The second conversion unit 22 is used to perform coordinate conversion processing according to the radial line pixel value layer, and convert the manually marked radial line data into three-dimensional coordinate points using a cubic interpolation algorithm to obtain a radial line marked point set in a three-dimensional space; The third conversion unit 23 performs non-rigid image registration based on the mutual information method, aligns the three-dimensional coordinate point set with the brain gray matter, white matter and ventricle anatomical structure in the fetal MR image, and obtains an aligned data set; The fourth conversion unit 24 is used to create a virtual image layer according to the aligned data set, segment the structure in the fetal brain image by using threshold segmentation and region growing method, and embed the aligned radial data into the gray matter, white matter or ventricle layer to obtain an enhanced sample set.

[0060] In a specific implementation disclosed in the present application, the classification module 3 includes: The first classification unit 31 is used to perform age sorting processing according to the enhanced sample set, by quantitatively analyzing the change of the diameter in the fetal MR image with age, and arranging the samples in ascending order according to the fetal age based on the growth law of the anatomical structure in the image data, to obtain a sorted sample set; The second classification unit 32 is used to perform developmental stage classification processing according to the sorted sample set, obtain the relationship between the age and the anatomical structure corresponding to each sample by performing linear regression fitting on the measurement data of the brain anatomical structure in the fetal MR image, and arrange the samples in ascending order of age according to the calculated fitting values ​​to obtain the sorted sample set; The third classification unit 33 is used to perform developmental stage classification processing according to the sorted sample set, by analyzing the trend of the anatomical characteristics of each sample changing with age, and dividing the sample into multiple developmental stages based on the change amplitude of the anatomical characteristics, so as to obtain a developmental stage classification sample set; The fourth classification unit 34 is used to perform feature extraction processing according to the developmental stage classification sample set, and obtain the classification sample set by extracting the anatomical features of each developmental stage and mapping the features to the standard spatial position.

[0061] In a specific implementation disclosed in the present application, the expansion module 4 includes: The first expansion unit 41 is used to extract and process the change pattern of the developmental stage characteristics according to the classification sample set, and to obtain the change pattern of the anatomical characteristics of each developmental stage by performing differential calculation on the change trend of the brain anatomical characteristics of the samples at each developmental stage and combining the nonlinear regression model to model the changes of the brain structure at different developmental stages; The second expansion unit 42 is used to perform sample generation processing according to the change pattern of the developmental stage characteristics, match the anatomical characteristics of the developmental stage with the shape context template by applying a deformation method based on the shape context, calculate the transformation matrix of the morphological deformation, and generate a new sample set; The third expansion unit 43 is used to perform noise simulation processing according to the new sample set, and to introduce a random noise model based on Poisson distribution to perturb the newly generated samples, so as to simulate the noise influence in actual imaging and obtain a new sample set with noise; The fourth expansion unit 44 is used to perform weighted combination of the original samples and the new sample set with noise to obtain an expanded sample set.

[0062] In a specific embodiment disclosed in the present application, the building block 5 includes: The first construction unit is used to extract and process the radial features and anatomical structure features according to the expanded sample set, decouple and extract the radial features and brain anatomical structures in the fetal MR images through a multi-channel convolutional neural network, and obtain a radial-anatomical feature map; The second construction unit 52 is used to perform spatial dependency modeling processing according to the radial-anatomical feature map, and to obtain a spatial dependency feature map by modeling the spatial relationship between the radial feature and the anatomical structure feature using a graph convolutional network based on anatomical constraints; The third construction unit 53 is used to perform spatial remapping processing according to the spatial dependency feature map, and obtain a geometric optimization feature map by adjusting the proportion, morphology and spatial position of the anatomical structure and optimizing the geometric relationship between the radial feature and the adjacent anatomical structure to ensure the spatial alignment of the anatomical structure and the radial at different developmental stages; The fourth construction unit 54 is used to perform deep learning training processing based on the geometric optimization feature map. The automatic diameter measurement task and the anatomical structure segmentation task are trained simultaneously by adopting a multi-task learning framework, and the parameters of the network are optimized using a joint loss function to construct an automatic diameter measurement model.

[0063] In a specific embodiment disclosed in the present application, the diagnosis module 6 includes: The first diagnosis unit 61 is used to input the fetal MR brain image data of the patient to be diagnosed into the automatic measurement model, automatically mark the radial lines in the image using a convolutional neural network, extract key radial line features, and predict the radial line position through a regression network to obtain the radial line marking result; The second diagnosis unit 62 is used to perform symmetry analysis according to the diameter marking result, and obtain the symmetry analysis result by comparing the diameter difference between the left and right cerebral hemispheres and quantifying the symmetry deviation in combination with the geometric constraints of the anatomical features; The third diagnosis unit 63 is used to perform pathological feature derivation processing according to the symmetry analysis result, by comparing the symmetry analysis result with the known pathological anatomical change pattern, applying the support vector machine to automatically determine whether there is pathological asymmetry, and obtain the pathological feature derivation result; The fourth diagnostic unit 64 is used to input the derivation results of the pathological characteristics into a preset brain disease classification model, and deduce whether there are lesions in the fetal brain and generate a brain disease diagnosis conclusion based on the pattern of the pathological characteristics and the disease samples marked in the training data. The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited to this. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for automated measurement and medical diagnosis of fetal MR brain diameters, characterized in that: include: Acquiring an original sample set, wherein the original sample set includes historical fetal MR brain image data with manually annotated diameters and corresponding brain diagnosis results; Based on the original sample set, the manually annotated radial data are converted into coordinate points in three-dimensional space, the coordinate points are aligned with corresponding anatomical structures in the fetal MR brain image, and the radial data are embedded into the structural layer of the image by creating a virtual image layer, so as to obtain an enhanced sample set; Sorting and classifying the enhanced sample set, analyzing the developmental rules of the brain structure of fetuses of different age groups, arranging the samples in ascending order according to the fetal age, and dividing the samples into age groups of multiple developmental stages based on the changing trends of brain structure characteristics, to obtain a classified sample set; Performing sample expansion processing according to the classified sample set, generating new samples based on the change pattern of brain structure at each developmental stage, and combining random noise vectors to simulate interference in actual imaging, thereby obtaining an expanded sample set; Performing model construction processing according to the expanded sample set, using a deep convolutional neural network to transform and fuse the radial features in the image data with the geometric relationship of the adjacent anatomical structures, to obtain an automatic radial measurement model; The fetal MR brain image data of the patient to be diagnosed is input into the automatic measurement model for annotation processing to obtain the radial measurement result, and the symmetry analysis is performed based on the radial measurement result to generate a brain disease diagnosis conclusion.

2. The method for automated measurement and medical diagnosis of fetal MR brain diameters according to claim 1, characterized in that: Based on the original sample set, the manually annotated radial data is converted into coordinate points in three-dimensional space, the coordinate points are aligned with corresponding anatomical structures in the fetal MR brain image, and the radial data is embedded into the structural layer of the image by creating a virtual image layer, thereby obtaining an enhanced sample set, including: Performing encoding processing according to the original sample set, fitting the manually marked radial data by using a B-spline curve fitting algorithm to obtain a curvature code, converting the curvature code into the value of the pixel point of the image data, and generating a radial pixel value layer; Coordinate conversion processing is performed according to the radial pixel value layer, and the manually marked radial data is converted into three-dimensional coordinate points using a cubic interpolation algorithm to obtain a radial annotation point set in a three-dimensional space; Performing non-rigid image registration based on mutual information method, aligning the three-dimensional coordinate point set with the brain gray matter, white matter and ventricle anatomical structure in the fetal MR image to obtain an aligned data set; A virtual image layer creation process is performed according to the aligned data set, the structure in the fetal brain image is segmented by using threshold segmentation and region growing method, and the aligned radial data is embedded into the brain gray matter, white matter or ventricle layer to obtain an enhanced sample set.

3. The method for automated measurement and medical diagnosis of fetal MR brain diameters according to claim 1, characterized in that: Sorting and classification processing is performed according to the enhanced sample set, and by analyzing the developmental rules of fetal brain structures at different age groups, the samples are arranged in ascending order according to fetal age, and the samples are divided into age groups at multiple developmental stages based on the changing trends of brain structural characteristics, to obtain a classified sample set, including: Performing age sorting processing according to the enhanced sample set, quantitatively analyzing the change of the diameter in the fetal MR image with age, and arranging the samples in ascending order of fetal age based on the growth law of the anatomical structure in the image data, to obtain a sorted sample set; Performing developmental stage division processing according to the sorted sample set, performing linear regression fitting on the measurement data of the brain anatomical structure in the fetal MR image, obtaining a relationship between the age and the anatomical structure corresponding to each sample, and arranging the samples in ascending order of age according to the calculated fitting values ​​to obtain a sorted sample set; Performing developmental stage division processing according to the sorted sample set, analyzing the trend of the anatomical characteristics of each sample changing with age, and dividing the sample into multiple developmental stages based on the magnitude of the change in the anatomical characteristics, to obtain a developmental stage classification sample set; Feature extraction processing is performed on the developmental stage classification sample set, and the classification sample set is obtained by extracting anatomical features of each developmental stage and mapping the features to standard spatial positions.

4. The method for automated measurement and medical diagnosis of fetal MR brain diameters according to claim 1, characterized in that: Sample expansion processing is performed according to the classified sample set, new samples are generated based on the change pattern of brain structure in each developmental stage, and random noise vectors are combined to simulate interference in actual imaging to obtain an expanded sample set, including: Extracting and processing the change pattern of the developmental stage characteristics according to the classified sample set, performing differential calculation on the change trend of the brain anatomical characteristics of the samples at each developmental stage, and modeling the changes of the brain structure at different developmental stages in combination with a nonlinear regression model, thereby obtaining the change pattern of the anatomical characteristics at each developmental stage; Performing sample generation processing according to the feature change pattern of the developmental stage, matching the anatomical features of the developmental stage with the shape context template by applying a deformation method based on shape context, calculating the transformation matrix of the morphological deformation, and generating a new sample set; Performing noise simulation processing on the new sample set, introducing a random noise model based on Poisson distribution to perturb the newly generated samples, simulating the noise influence in actual imaging to obtain a new sample set with noise; The original samples are weightedly combined with the new sample set with noise to obtain an expanded sample set.

5. The method for automated measurement and medical diagnosis of fetal MR brain diameters according to claim 1, characterized in that: The model is constructed based on the expanded sample set, and a deep convolutional neural network is used to transform and fuse the radial features in the image data with the geometric relationship of the adjacent anatomical structures to obtain an automatic radial measurement model, including: Extracting radial features and anatomical structure features according to the expanded sample set, decoupling and extracting radial features from the fetal MR images and the brain anatomical structures through a multi-channel convolutional neural network, and obtaining a radial-anatomical feature map; Performing spatial dependency modeling processing according to the diameter-anatomical feature map, modeling the spatial relationship between the diameter feature and the anatomical structure feature by using a graph convolutional network based on anatomical constraints, and obtaining a spatial dependency feature map; Performing spatial remapping processing according to the spatial dependency feature map, adjusting the proportion, morphology and spatial position of the anatomical structure, and optimizing the geometric relationship between the radial feature and the adjacent anatomical structure, thereby ensuring the spatial alignment of the anatomical structure and the radial line at different developmental stages, and obtaining a geometrically optimized feature map; Deep learning training is performed according to the geometric optimization feature map. The automatic diameter measurement task and the anatomical structure segmentation task are trained simultaneously by adopting a multi-task learning framework, and the parameters of the network are optimized using a joint loss function to construct an automatic diameter measurement model.

6. A fetal MR brain diameter automatic measurement and medical diagnosis system, characterized in that: include: An acquisition module, used for acquiring an original sample set, wherein the original sample set includes historical fetal MR brain image data with manually annotated diameters and corresponding brain diagnosis results; A conversion module, based on the original sample set, converts the manually annotated radial data into coordinate points in three-dimensional space, aligns the coordinate points with corresponding anatomical structures in the fetal MR brain image, and embeds the radial data into the structural layer of the image by creating a virtual image layer, thereby obtaining an enhanced sample set; A classification module, used to sort and classify the enhanced sample set, analyze the developmental rules of the brain structure of fetuses of different age groups, arrange the samples in ascending order according to the fetal age, and divide the samples into multiple age groups of developmental stages based on the change trend of brain structure characteristics to obtain a classified sample set; An expansion module is used to perform sample expansion processing according to the classified sample set, generate new samples based on the change pattern of brain structure in each developmental stage, and combine random noise vectors to simulate interference in actual imaging to obtain an expanded sample set; A construction module is used to perform model construction processing according to the expanded sample set, and use a deep convolutional neural network to transform and fuse the radial features in the image data with the geometric relationship of the adjacent anatomical structures to obtain an automatic radial measurement model; The diagnosis module is used to input the fetal MR brain image data of the patient to be diagnosed into the automatic measurement model for annotation processing to obtain the radial measurement result, and perform symmetry analysis based on the radial measurement result to generate a brain disease diagnosis conclusion.

7. The automatic fetal MR brain diameter measurement and medical diagnosis system according to claim 6, characterized in that: The conversion module comprises: A first conversion unit is used to perform encoding processing according to the original sample set, obtain curvature coding by fitting the manually marked radial data using a B-spline curve fitting algorithm, convert the curvature coding into the value of the pixel point of the image data, and generate a radial pixel value layer; A second conversion unit is used to perform coordinate conversion processing according to the radial pixel value layer, and convert the manually marked radial data into three-dimensional coordinate points using a cubic interpolation algorithm to obtain a radial annotation point set in a three-dimensional space; A third conversion unit performs non-rigid image registration based on a mutual information method to align the three-dimensional coordinate point set with the brain gray matter, white matter and ventricle anatomical structure in the fetal MR image to obtain an aligned data set; The fourth conversion unit is used to create a virtual image layer according to the aligned data set, segment the structure in the fetal brain image by using threshold segmentation and region growing method, and embed the aligned radial data into the brain gray matter, white matter or ventricle layer to obtain an enhanced sample set.

8. The automatic fetal MR brain diameter measurement and medical diagnosis system according to claim 6, characterized in that: The classification module comprises: A first classification unit is used to perform age sorting processing according to the enhanced sample set, by quantitatively analyzing the change of the diameter in the fetal MR image with age, and arranging the samples in ascending order according to the fetal age based on the growth law of the anatomical structure in the image data, to obtain a sorted sample set; a second classification unit, configured to perform developmental stage classification processing according to the sorted sample set, obtain a relationship between age and anatomical structure corresponding to each sample by performing linear regression fitting on the measurement data of brain anatomical structure in fetal MR images, and arrange the samples in ascending order of age according to the calculated fitting values ​​to obtain a sorted sample set; A third classification unit is used to perform developmental stage classification processing according to the sorted sample set, by analyzing the trend of the anatomical characteristics of each sample changing with age, and dividing the sample into multiple developmental stages based on the change amplitude of the anatomical characteristics, so as to obtain a developmental stage classification sample set; The fourth classification unit is used to perform feature extraction processing according to the developmental stage classification sample set, and obtain the classification sample set by extracting the anatomical features of each developmental stage and mapping the features to a standard spatial position.

9. The automatic fetal MR brain diameter measurement and medical diagnosis system according to claim 6, characterized in that: The expansion module comprises: A first expansion unit is used to extract and process the change pattern of the developmental stage characteristics according to the classification sample set, perform differential calculation on the change trend of the brain anatomical characteristics of the samples at each developmental stage, and model the changes of the brain structure at different developmental stages in combination with a nonlinear regression model to obtain the change pattern of the anatomical characteristics at each developmental stage; A second expansion unit is used to perform sample generation processing according to the feature change pattern of the development stage, match the anatomical features of the development stage with the shape context template by applying a deformation method based on shape context, calculate the transformation matrix of the morphological deformation, and generate a new sample set; A third expansion unit is used to perform noise simulation processing according to the new sample set, and to introduce a random noise model based on Poisson distribution to perturb the newly generated samples, so as to simulate the noise influence in actual imaging and obtain a new sample set with noise; The fourth expansion unit is used to perform weighted combination of the original sample and the new sample set with noise to obtain an expanded sample set.

10. The automatic fetal MR brain diameter measurement and medical diagnosis system according to claim 6, characterized in that: The building blocks include: The first construction unit is used to extract radial features and anatomical structure features according to the expanded sample set, decouple and extract radial features from brain anatomical structures in fetal MR images through a multi-channel convolutional neural network, and obtain a radial-anatomical feature map; A second construction unit is used to perform spatial dependency modeling processing according to the diameter-anatomical feature map, and to obtain a spatial dependency feature map by modeling the spatial relationship between the diameter feature and the anatomical structure feature using a graph convolutional network based on anatomical constraints; The third construction unit is used to perform spatial remapping processing according to the spatial dependency feature map, and obtain a geometrically optimized feature map by adjusting the proportion, morphology and spatial position of the anatomical structure and optimizing the geometric relationship between the diameter feature and the adjacent anatomical structure to ensure the spatial alignment of the anatomical structure and the diameter at different developmental stages; the fourth construction unit is used to perform deep learning training processing according to the geometrically optimized feature map, and simultaneously train the automatic diameter measurement task and the anatomical structure segmentation task by adopting a multi-task learning framework, and optimize the parameters of the network using a joint loss function, so as to construct an automatic diameter measurement model.

Images