Processing method based on heart three-dimensional image

Abstract

The invention discloses a processing method and device based on a heart three-dimensional image, equipment and a medium. The method comprises the following steps: acquiring a heart three-dimensional image and a chamber segmentation result to obtain a left ventricular heart cavity and a cardiac muscle; according to the inner diameter and the outer diameter of the myocardial, constructing a three-dimensional inner diameter curved surface and generating myocardial wall thickness mapping; determining a pre-trained classification model and constructing two-channel input of the classification model; inputting the curved surface and the mapping into a first channel of a model, and combining whole heart segmentation or original image data of a second channel; and processing dual-channel input through the model, and outputting a cardiomyopathy classification result. According to the method, a complete spatial form is reserved through the three-dimensional inner diameter curved surface and wall thickness mapping, and precise and robust classification of various cardiomyopathy such as hypertrophic cardiomyopathy and dilated cardiomyopathy is realized by combining dual-channel feature fusion and interactive calibration of a space guidance mask; the problems of efficiently and accurately extracting features from a complex three-dimensional heart image and automatically identifying diseases are solved.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method, apparatus, device, and medium for processing three-dimensional cardiac images. Background Technology

[0002] Accurately and efficiently extracting left ventricular myocardial features from complex 3D cardiac imaging data and classifying cardiomyopathy is a core technical problem that urgently needs to be solved. This problem stems from the complexity and diversity of medical imaging, and the significant differences in pathological features between cardiomyopathy types such as hypertrophic and dilated types. This makes it difficult for traditional methods to simultaneously achieve precision in feature extraction, accuracy in classification, and efficiency in processing in practical scenarios. Specifically, raw cardiac imaging data exhibits a high-dimensional spatial structure under a stereomicroscope, containing noise, individual differences, and data heterogeneity introduced by imaging equipment. Accurately segmenting the inner and outer diameters of the left ventricular myocardium and generating high-quality wall thickness distribution maps and inner diameter surfaces becomes a primary challenge. Furthermore, a key contradiction lies in how to effectively reduce computational complexity through cropping and feature extraction while preserving 3D spatial morphological information, while ensuring the distinguishing ability of features for different types of cardiomyopathy. Moreover, classification models are prone to overfitting or insufficient robustness when faced with small sample data or subtle differences in pathological features, affecting the reliability of detection. These sub-problems all point to a core challenge: how to achieve accurate detection of cardiomyopathy by employing efficient segmentation, feature extraction, and classification processes that balance accuracy, robustness, and computational efficiency within complex and varied cardiac imaging data. This problem is particularly prominent in real-world clinical scenarios because the accuracy of detection directly impacts patient treatment decisions, while efficiency affects the allocation of medical resources and the speed of detection. Summary of the Invention

[0003] Given the limitations of existing complex and variable cardiac imaging data in achieving efficient segmentation, feature extraction, and classification processes while maintaining accuracy, robustness, and computational efficiency for precise detection of cardiomyopathy, this invention provides a processing method based on three-dimensional cardiac images, offering efficient and accurate technical support for cardiomyopathy detection.

[0004] This invention provides a method for processing three-dimensional cardiac images, mainly including the following steps:

[0005] Acquire three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers, wherein the three-dimensional segmentation results of the heart chambers include the left ventricular chambers and the left ventricular myocardium;

[0006] Based on the inner and outer diameters of the left ventricular myocardium, a three-dimensional inner diameter surface is constructed and a myocardial wall thickness mapping is generated, wherein the myocardial wall thickness mapping represents the shortest distance from the inner diameter pixel to the outer diameter.

[0007] Determine the pre-trained classification model and construct the dual-channel input of the pre-trained classification model;

[0008] The mapping between the three-dimensional inner diameter surface and the myocardial wall thickness is input into the first input channel of the pre-trained classification model;

[0009] The dual-channel input is processed by the pre-trained classification model to output the cardiomyopathy classification model result;

[0010] The pre-trained classification model's dual-channel input includes a first input channel relating the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness.

[0011] Furthermore, three-dimensional imaging data of the heart and three-dimensional segmentation results of the heart chambers are acquired, including:

[0012] Obtain raw medical imaging data of the heart;

[0013] The original medical image data is processed using a pre-established segmentation algorithm to generate a three-dimensional segmentation result of the heart chambers;

[0014] Based on the three-dimensional segmentation results of the heart chambers, the boundary structure information of the left ventricular chamber and the left ventricular myocardium is extracted;

[0015] The inner and outer diameters of the left ventricular myocardium are determined based on the boundary structure information. The inner diameter is the boundary between the left ventricular myocardium and the left ventricular chamber, and the outer diameter is the boundary between the left ventricular myocardium and the right ventricle or background.

[0016] Furthermore, determining the inner and outer diameters of the left ventricular myocardium based on the boundary structure information includes:

[0017] The centroid of the left ventricular base is extracted as the first reference point based on the boundary structure information.

[0018] The point in the left ventricle that is farthest from the first reference point is determined as the second reference point;

[0019] The long axis of the left ventricle is constructed using the first reference point and the second reference point;

[0020] Based on the boundary structure information, the inner diameter is determined by the edge of the left ventricular myocardium on the left ventricular chamber side;

[0021] The outer diameter is defined as the boundary that distinguishes the left ventricular myocardium from the right ventricle and the background.

[0022] Furthermore, the step of constructing a three-dimensional inner diameter surface and generating a myocardial wall thickness mapping based on the inner and outer diameters of the left ventricular myocardium includes:

[0023] Iterate through all pixels of the left ventricular myocardial inner diameter, and for each inner diameter pixel, calculate the shortest distance between it and the outer diameter pixel.

[0024] Myocardial wall thickness mapping is generated based on the shortest distance, and the myocardial wall thickness mapping reflects the spatial distribution of inner diameter pixels and the corresponding wall thickness intensity.

[0025] A three-dimensional inner diameter surface is constructed based on the spatial distribution of the inner diameter pixels, and the three-dimensional inner diameter surface preserves the original spatial morphology of the left ventricular myocardium;

[0026] A preset smoothing algorithm is applied to the myocardial wall thickness mapping and the three-dimensional inner diameter surface to obtain a smooth three-dimensional inner diameter surface and myocardial wall thickness mapping.

[0027] Furthermore, the dual-channel input of the pre-trained classification model includes a second input channel regarding the whole heart segmentation result or the original medical image data corresponding to the whole heart segmentation.

[0028] Furthermore, the step of processing the dual-channel input using the pre-trained classification model to output the cardiomyopathy classification model result includes:

[0029] Based on the spatial distribution of the inner diameter of the left ventricular myocardium, a three-dimensional spatial box containing the inner diameter of the left ventricular myocardium is extracted to obtain the cropped three-dimensional data; the three-dimensional spatial box covers the x-axis, y-axis and z-axis directions of the inner diameter;

[0030] The cropped 3D data and myocardial wall thickness mapping are processed by the Adaptive AvgPool3d operator to generate output data in a unified format.

[0031] The output data in the unified format is input into the pre-trained classification model through the first input channel; and

[0032] The whole heart segmentation result or the corresponding raw medical image data is obtained and used as the second channel input to the pre-trained classification model; the second channel input includes chamber morphology information or texture information.

[0033] The dual-channel input is processed by the pre-trained classification model to output the cardiomyopathy classification model result.

[0034] Furthermore, after obtaining the whole heart segmentation result or the corresponding raw medical image data as the second channel input and inputting it into the pre-trained classification model, the process further includes:

[0035] The dual-channel input is subjected to feature extraction using a 3D convolutional network to obtain high-dimensional features, which include the spatial morphology and wall thickness distribution information of the left ventricular myocardium.

[0036] Based on the high-dimensional features, classification processing is performed to generate cardiomyopathy classification results;

[0037] The cardiomyopathy classification results include hypertrophic cardiomyopathy, dilated cardiomyopathy, restrictive cardiomyopathy, amyloidosis, myocarditis, and normal type;

[0038] Based on the cardiomyopathy classification results, a preset optimization algorithm is applied to adjust the classification model parameters to obtain optimized cardiomyopathy classification model results, which are then output as cardiomyopathy classification model results.

[0039] Furthermore, the step of processing the dual-channel input using the pre-trained classification model and outputting the cardiomyopathy classification model result includes:

[0040] Based on the degree of deviation of each sampling point in the myocardial wall thickness mapping from the preset myocardial wall thickness reference value, the spatial weight of each sampling point is determined, and a spatial guidance mask matrix M aligned with the spatial scale of the three-dimensional image data of the heart is generated.

[0041] The three-dimensional image data of the heart is input into the second input channel of the classification model, and a deep spatial feature tensor T containing anatomical texture information is extracted through multidimensional convolution.

[0042] By introducing a dynamic adjustment factor β, the spatial guidance mask matrix M is used to perform weight modulation on the deep spatial feature tensor T on a per-spatial coordinate basis to enhance the feature intensity corresponding to the wall thickness anomaly region and generate a calibrated spatial interaction feature map F.

[0043] The spatial dimension of the spatial interaction feature map F is compressed to extract the core pathological feature vector, and the core pathological feature vector is mapped to the category space to output the cardiomyopathy classification model result.

[0044] Furthermore, a three-dimensional inner diameter surface is constructed based on the spatial distribution of the inner diameter pixels, including:

[0045] Based on the coordinate information of the inner diameter pixels, a three-dimensional mesh structure is generated, which represents the inner diameter surface of the left ventricular myocardium;

[0046] For the aforementioned three-dimensional mesh structure, a preset interpolation algorithm is applied to fill in the missing surface data;

[0047] Based on the spatial distribution of the inner diameter pixels, the geometry of the three-dimensional mesh structure is adjusted to obtain a three-dimensional inner diameter surface that retains the original spatial shape;

[0048] For the three-dimensional inner diameter surface, a corresponding color code is generated, which reflects the spatial position information of the surface.

[0049] It is important to note that the interpolation algorithm uses trilinear interpolation or spline interpolation to fill in missing data; the geometric adjustment of the three-dimensional mesh structure needs to be combined with the spatial distribution of the inner diameter pixels to ensure that the original spatial morphology of the left ventricular myocardium is preserved; color coding can be associated with physiological parameters, for example, red indicates the area with a wall thickness > 15mm and located at the apex of the heart, reflecting both spatial location and wall thickness value.

[0050] Furthermore, the step of extracting a three-dimensional bounding box containing the inner diameter of the left ventricular myocardium based on the spatial distribution of the left ventricular myocardium includes:

[0051] Based on the spatial distribution of pixels in the inner diameter, the cropping range in the x-axis, y-axis and z-axis directions is determined;

[0052] For the clipping range, a three-dimensional bounding box containing the inner diameter is generated, and the bounding box covers the main distribution area of ​​the left ventricular myocardial inner diameter;

[0053] The three-dimensional inner diameter surface and myocardial wall thickness mapping are clipped based on the bounding box to obtain the clipped three-dimensional data.

[0054] It is important to note that the cropping range should be set with a 5% redundancy, meaning the cropping range should extend 5% beyond the inner diameter boundary to avoid information loss and reduce redundancy. Multi-region joint cropping is supported, which can simultaneously crop the left ventricular myocardium and parts of the adjacent right ventricle, preserving the positional relationship between chambers. For cases with uneven inner diameter distribution (such as the z-axis being concentrated in the lower half), the boundary can be dynamically adjusted to minimize the data volume.

[0055] The present invention also provides a processing device based on three-dimensional cardiac images, comprising:

[0056] The acquisition module is used to acquire three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers, wherein the three-dimensional segmentation results of the heart chambers include the left ventricular chambers and the left ventricular myocardium;

[0057] The processing module is used to construct a three-dimensional inner diameter surface and generate a myocardial wall thickness mapping based on the inner and outer diameters of the left ventricular myocardium, wherein the myocardial wall thickness mapping represents the shortest distance from the inner diameter pixel to the outer diameter.

[0058] It should be noted that the processing module performs multi-threaded parallel computation to traverse the inner diameter pixels, uses a triangulation algorithm to generate a three-dimensional mesh structure, and optimizes the surface and mapping quality through Gaussian filtering and Laplacian smoothing.

[0059] The training module is used to determine the pre-trained classification model and construct the dual-channel input of the pre-trained classification model;

[0060] The training module is also used to input the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness into the first input channel of the pre-trained classification model;

[0061] It is important to note that before inputting, a 3×3×3 Gaussian smoothing filter with a kernel of 3×3×3 needs to be applied to the three-dimensional inner diameter surface and myocardial wall thickness mapping. After fusion, a composite input reflecting the local thickness and spatial morphology of the left ventricular myocardium is generated.

[0062] The training module is also used to process the dual-channel input through the pre-trained classification model and output the cardiomyopathy classification model result;

[0063] It is important to note that high-dimensional features are extracted through a 3D convolutional network, the output probability distribution is calculated through a fully connected layer, and the highest probability is selected as the classification result after normalization using the Softmax function.

[0064] The pre-trained classification model's dual-channel input includes a first input channel relating the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness.

[0065] The present invention also provides a non-volatile computer-readable storage medium storing computer-executable instructions thereon, which, when executed by a processor, perform the following steps:

[0066] Acquire three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers, wherein the three-dimensional segmentation results of the heart chambers include the left ventricular chambers and the left ventricular myocardium;

[0067] Based on the inner and outer diameters of the left ventricular myocardium, a three-dimensional inner diameter surface is constructed and a myocardial wall thickness mapping is generated, wherein the myocardial wall thickness mapping represents the shortest distance from the inner diameter pixel to the outer diameter.

[0068] Determine the pre-trained classification model and construct the dual-channel input of the pre-trained classification model;

[0069] The mapping between the three-dimensional inner diameter surface and the myocardial wall thickness is input into the first input channel of the pre-trained classification model;

[0070] The dual-channel input is processed by the pre-trained classification model to output the cardiomyopathy classification model result;

[0071] The pre-trained classification model's dual-channel input includes a first input channel relating the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness.

[0072] Optionally, during processor execution, the original medical image data needs to be preprocessed by denoising, contrast enhancement, and spatial normalization. The segmentation algorithm adopts a three-dimensional U-Net architecture, and the classification model needs to perform spatial guidance mask matrix modulation and dynamic adjustment factor weighting operations during processing.

[0073] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any one of the methods described above.

[0074] Optionally, the computer device's memory needs to store a training dataset of more than 500 fully annotated cardiac images. The processor should support multi-threaded parallel computation of the shortest distance between inner diameter pixels during execution, and be compatible with high-dimensional feature extraction and adaptive average pooling 3d (AdaptiveAvgPool3d) operator operations of 3D convolutional networks.

[0075] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0076] 1. This invention discloses a method, device, equipment, and medium for classifying cardiomyopathy based on three-dimensional cardiac imaging data. It addresses the business scenario problem of how to accurately and efficiently extract left ventricular myocardial features from complex cardiac images and achieve cardiomyopathy classification. By constructing a three-dimensional inner diameter surface and mapping it to myocardial wall thickness, and combining it with a pre-trained classification model with dual-channel input, it achieves accurate classification of hypertrophic and dilated cardiomyopathy.

[0077] 2. This invention first processes the original medical images through a segmentation algorithm to extract the inner and outer diameters of the left ventricular myocardium, generating a mapping that reflects the wall thickness distribution and an inner diameter surface that preserves the spatial morphology; then, it uses a three-dimensional bounding box to crop the data, combines the whole heart segmentation results, extracts high-dimensional features through a 3D convolutional network, and finally outputs the classification results.

[0078] 3. This invention optimizes the surface and mapping quality through a smoothing algorithm and applies the Adaptive AvgPool3d operator to ensure data consistency, which significantly improves the robustness and accuracy of the classification model and provides efficient and accurate technical support for cardiomyopathy detection.

[0079] 4. This invention, through a three-dimensional feature mutual induction mechanism, avoids the dimensional collapse problem caused by two-dimensional projection or regional statistics of cardiac geometric features in existing technologies, thus fully preserving the texture details in high-dimensional space. This enables the model to accurately capture subtle pathological changes hidden in areas of abnormal myocardial wall thickness, such as signal disturbances caused by early focal hypertrophy. It fundamentally solves the technical limitations of existing technologies, such as the scheme disclosed in patent document CN121033254A, which cannot identify local discrete lesions due to "dimensionality reduction statistics," and significantly improves diagnostic sensitivity.

[0080] 5. By modulating the original image features point-by-point using a spatially guided mask matrix, this scheme constructs a "navigation mechanism" based on anatomical morphology, effectively filtering physiological noise and background interference from unrelated regions. During classification, feature signals are selectively amplified only at spatial coordinates where geometric deviations highly coincide with image texture anomalies. This mutual inductance alignment ensures that the model can accurately focus on clinically relevant anatomical lesion points, greatly improving the specificity and robustness of cardiomyopathy classification in complex medical scenarios. Attached Figure Description

[0081] Figure 1 This is a flowchart of a method for processing cardiac three-dimensional images according to the present invention.

[0082] Figure 2 This is a schematic diagram of one embodiment of the left ventricular long axis provided in this invention.

[0083] Figure 3 This is a rendering of the left ventricular long axis provided in an embodiment of the present invention.

[0084] Figure 4 This is a schematic diagram of the three-dimensional mapping generation process of left ventricular myocardial wall thickness provided in an embodiment of the present invention.

[0085] Figure 5 This is a diagram illustrating the effect of mapping the three-dimensional inner diameter surface to the myocardial wall thickness in an embodiment of the present invention.

[0086] Figure 6 The diagram shows the whole heart segmentation result and the original cardiac medical image corresponding to the whole heart segmentation result, as provided in the embodiments of the present invention.

[0087] Figure 7 This is a structural diagram of a processing device based on three-dimensional cardiac images provided in an embodiment of the present invention.

[0088] Figure 8 This is a structural diagram of a computing device provided in an embodiment of the present invention. Detailed Implementation

[0089] The technical solutions of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0090] like Figure 1 As shown, this embodiment of the invention provides a processing method based on three-dimensional cardiac images, which may specifically include the following steps:

[0091] S101. Obtain three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers, including the left ventricular chamber and the left ventricular myocardium.

[0092] Acquire three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers, including the left ventricular chambers and left ventricular myocardium, specifically including the following steps S11 to S15:

[0093] S11. Acquire raw medical imaging data of the heart. Raw three-dimensional images of the patient's heart are acquired using cardiac magnetic resonance imaging (MRI) or computed tomography (CT) scanners. The image data includes spatial distribution information of the heart's various chambers and myocardial tissue. The raw image data is stored in DICOM format and includes metadata information such as pixel spacing, slice thickness, and scan parameters.

[0094] S12. Preprocessing operations are performed on the raw medical image data. Preprocessing includes image denoising, contrast enhancement, and spatial normalization. Denoising uses a Gaussian filter to remove random noise from the images, with the filter kernel size set to 3×3×3 pixels. Contrast enhancement uses histogram equalization to improve the contrast difference between myocardial tissue and blood chambers. Spatial normalization unifies the image data from different patients into the same spatial coordinate system, ensuring consistency in subsequent segmentation algorithms.

[0095] S13. The pre-processed image data is segmented into cardiac chambers using a pre-established deep learning segmentation algorithm. The segmentation algorithm employs a 3D U-Net convolutional neural network architecture, consisting of an encoder and a decoder. The encoder extracts multi-scale features from the image through consecutive convolutional and pooling layers, while the decoder restores spatial resolution and generates a segmentation mask through upsampling and skip connections. The network training uses the Dice loss function to optimize segmentation accuracy until it reaches over 95%, thereby generating accurate 3D segmentation results including the left and right ventricles. The training dataset contains 500 fully annotated cardiac image datasets.

[0096] The preprocessed 3D image data is input into the trained U-Net network. The network performs classification prediction for each voxel, outputting a probability distribution map containing multiple categories such as background, left ventricular chamber, left ventricular myocardium, right ventricular chamber, and right ventricular myocardium. An activation function is used to convert the network output into probability values ​​for each category, assigning each voxel to the category with the highest probability.

[0097] The segmentation results output by the network are post-processed for optimization. Post-processing includes two steps: connected component analysis and morphological operations. Connected component analysis identifies and preserves the largest connected regions and removes isolated small noise regions. Morphological operations use opening and closing operations to smooth the segmentation boundaries; opening operations remove small connected parts, and closing operations fill in small internal holes.

[0098] S14. Extract the boundary structure information of the left ventricular chamber and left ventricular myocardium based on the 3D segmentation results of the heart chambers. The boundary structure information includes a set of 3D coordinate points of the inner and outer surface boundaries. The inner surface boundary is defined as the interface between the left ventricular myocardium and the left ventricular chamber, and the outer surface boundary is defined as the interface between the left ventricular myocardium and the right ventricle or background region. Boundary extraction uses a gradient calculation method, finding the locations in the segmentation mask where the label values ​​change as boundary points.

[0099] Calculate the gradient values ​​of the segmentation mask in three spatial directions. Gradient calculation uses an operator to perform a convolution operation on the segmentation mask to obtain gradient components in the x, y, and z directions. Locations with gradient magnitudes greater than a preset threshold are identified as potential boundary points; the threshold is set to the 95th percentile of the gradient magnitude distribution.

[0100] The identified boundary points are precisely located and classified. Boundary point classification determines the boundary type based on the label values ​​of adjacent voxels: left ventricular myocardium is labeled with 2, left ventricular chamber with 1, and background with 0. When the label value of an adjacent voxel changes from 1 to 2, the boundary point belongs to the inner surface boundary. When the label value of an adjacent voxel changes from 2 to 0 or to another ventricular label, the boundary point belongs to the outer surface boundary.

[0101] S15. Determine the inner and outer diameter geometries of the left ventricular myocardium based on boundary structure information. The inner diameter geometry is represented by a three-dimensional surface formed by the boundary points of the inner surface, and the outer diameter geometry is represented by a three-dimensional surface formed by the boundary points of the outer surface. The geometry is constructed by using a triangulation method to connect discrete boundary points into a continuous triangular mesh surface.

[0102] In one embodiment, the deep learning segmentation algorithm in step S13 further includes a multi-scale feature fusion mechanism. After the network extracts features at different resolution levels, it fuses the multi-scale information through a feature pyramid structure. Low-resolution features contain global semantic information, which helps distinguish different heart chambers. High-resolution features contain fine boundary information, which helps accurately locate the myocardial boundaries. Feature fusion is achieved through element-wise addition and channel attention mechanisms, with attention weights dynamically adjusted according to the importance of the features.

[0103] For example, in a typical computed tomography coronary angiography dataset, the original image size is 512×512×320 voxels with a voxel spacing of 0.5×0.5×0.35 mm. After preprocessing, the image is resampled to a uniform resolution of 0.3×0.3×0.3 mm. The left ventricular myocardial segmentation region output by the U-Net network contains approximately 80,000 voxels, with approximately 15,000 boundary points on the inner surface and approximately 25,000 boundary points on the outer surface. The 3D coordinate accuracy of the boundary points reaches the sub-voxel level, providing an accurate geometric basis for subsequent myocardial wall thickness calculations.

[0104] It should be noted that the quality of the 3D segmentation results of the heart chambers directly affects the accuracy of subsequent cardiomyopathy classification. By employing deep learning methods and meticulous post-processing steps, high-precision segmentation results of the left ventricular chambers and left ventricular myocardium can be obtained, providing a reliable data foundation for myocardial wall thickness mapping and 3D surface construction.

[0105] In another optional implementation, the inner diameter is the boundary between the left ventricular myocardium and the left ventricular chamber, and the outer diameter is the boundary between the left ventricular myocardium and the right ventricle or background. The expansion method of the inner and outer diameters of the left ventricular myocardium is determined based on the boundary structure information, specifically including:

[0106] The centroid of the left ventricular base is extracted based on the boundary structure information and used as the first reference point. This centroid is obtained by calculating the average coordinates of all pixels in the base plane.

[0107] The point in the left ventricle chamber that is farthest from the first reference point is determined as the second reference point. The Euclidean distance formula is used to calculate the distance from all points to the centroid and the maximum value is selected.

[0108] The major axis of the left ventricle is constructed using the first and second reference points. This major axis is a straight line vector connecting the two points and is used to standardize the direction of the heart.

[0109] Based on the boundary structure information, the edge of the left ventricular myocardium on the left ventricular cavity side is determined as the inner diameter, which is obtained by the intersection of the boundary point cloud;

[0110] The outer diameter is determined by the edge that distinguishes the left ventricular myocardium from the right ventricle and the background, and the boundary of adjacent regions is detected using label difference detection.

[0111] In one embodiment, for normal cardiac images, the extracted centroid coordinates are (100, 150, 200), the second reference point is (100, 150, 50), and the constructed major axis is approximately 150 mm long. This allows for precise definition of the inner diameter as the contact surface between the myocardium and the heart chamber, and the outer diameter as the external boundary of the myocardium, which is beneficial to the accuracy of subsequent wall thickness calculations. For example, in one possible implementation, for images of patients with dilated cardiomyopathy, the constructed major axis is used to calibrate the heart orientation by rotation, ensuring that the number of inner diameter edge points exceeds 10,000, and the outer diameter filters background noise through its intersection with the right ventricular label. This improves the reliability of the classification model input and has the beneficial effect of reducing the misdiagnosis rate.

[0112] In one embodiment, for cardiac images with amyloidosis, the inner diameter determination process takes into account the expansion and deformation of the cardiac chambers. Noise points are filtered by adjusting the distance threshold to 0.5 mm, while the outer diameter is extracted based on the edges with a background gray-level gradient greater than 50. This ensures the integrity of the boundary information and is beneficial to the accuracy of cardiomyopathy detection.

[0113] Please see Figure 2 This illustrates a schematic embodiment of the left ventricular long axis, such as... Figure 2 As shown, the long axis of the left ventricle is the line connecting the base (point A) to the apex (point B) of the left ventricle. The base (point A) is the centroid of the cross-section (base) of the left ventricular chamber and the left atrium. The apex (point B) is the point on the long axis, which is the point of the left ventricle farthest from point A.

[0114] Please see Figure 3 It shows a rendering of the long axis of the left ventricle, such as Figure 3 The image on the left shows a top-down view of the long axis of the left ventricle, as shown in the diagram. Figure 3 The image on the right shows the side view of the long axis of the left ventricle.

[0115] S102. Based on the inner and outer diameters of the left ventricular myocardium, construct a three-dimensional inner diameter surface and generate a myocardial wall thickness mapping. The myocardial wall thickness mapping represents the shortest distance from the inner diameter pixel to the outer diameter.

[0116] Specifically, the process iterates through all pixels of the left ventricular myocardial inner diameter, calculating the shortest distance between each inner diameter pixel and the outer diameter pixels. A three-dimensional coordinate system is established for the inner diameter pixels, representing each pixel as a coordinate point in three-dimensional space, with coordinate values ​​including x-axis, y-axis, and z-axis position information. The coordinates of the inner diameter pixels are obtained based on the boundary between the left ventricular myocardium and the left ventricular chamber in the three-dimensional segmentation of the heart chambers, and the position of all pixels on the boundary is identified through a boundary extraction algorithm.

[0117] A spatial index structure for the outer diameter point set is constructed, and all pixels on the edge distinguishing the left ventricular myocardium from the right ventricle and the background are organized into a data structure that facilitates fast querying.

[0118] For each inner diameter pixel, a nearest neighbor search algorithm is applied to calculate its shortest Euclidean distance to the outer diameter point set. The distance calculation formula is the straight-line distance between two points in three-dimensional space. The nearest outer diameter point is quickly located by traversing the outer diameter point set or using a spatial index structure. For an inner diameter pixel with coordinates (xi, yi, zi), its distance to all outer diameter point coordinates (xj, yj, zj) is calculated, and the minimum value among them is taken as the wall thickness value of the inner diameter pixel.

[0119] The shortest distance value and coordinates of the nearest outer diameter point are recorded for each inner diameter pixel, forming a one-to-one correspondence between inner diameter pixels and wall thickness values. The calculation results are stored as a data structure containing pixel coordinates and corresponding wall thickness values, providing basic data for the subsequent generation of myocardial wall thickness mapping.

[0120] Myocardial wall thickness mapping is generated based on the shortest distance. The myocardial wall thickness mapping reflects the spatial distribution of inner diameter pixels and the corresponding wall thickness intensity.

[0121] A mapping relationship between wall thickness values ​​and color codes is established, converting the calculated distance values ​​into a visual color representation. Using rainbow or heatmap color gamuts, the minimum wall thickness value is mapped to blue or a cool hue, the maximum wall thickness value to red or a warm hue, and intermediate values ​​are interpolated using linear or non-linear functions.

[0122] A 3D texture mapping data structure is constructed, associating the spatial coordinates of each inner diameter pixel with its corresponding wall thickness color code. The texture mapping uses a 3D texture coordinate system to ensure that color information can accurately correspond to specific locations in 3D space, forming a continuous visualization effect of wall thickness distribution.

[0123] Interpolation algorithms are applied to fill the gaps between inner diameter pixels, generating a continuous and smooth wall thickness mapping surface. Trilinear interpolation or spline interpolation methods are used to calculate the estimated wall thickness at intermediate positions based on the wall thickness values ​​of adjacent pixels, ensuring the continuity and smoothness of the mapping surface.

[0124] The numerical range and statistical information of the wall thickness mapping are generated, including the maximum wall thickness, minimum wall thickness, average wall thickness, and standard deviation of the wall thickness distribution. This statistical information is used for feature extraction in subsequent classification models and identification of abnormal wall thickness regions, providing a quantitative reference indicator for cardiomyopathy detection.

[0125] A three-dimensional inner diameter surface is constructed based on the spatial distribution of inner diameter pixels, which preserves the original spatial morphology of the left ventricular myocardium.

[0126] A triangular mesh topology is generated based on the 3D coordinates of the inner diameter pixels. A triangulation algorithm then connects the discrete pixels into continuous surface segments. The triangulation process considers the spatial adjacency of the pixels to ensure that the generated triangular mesh accurately represents the geometry of the left ventricular diameter.

[0127] A surface reconstruction algorithm is applied to optimize the geometric quality of the triangular mesh, eliminating sharp edges and irregular triangles. Methods such as Laplacian smoothing or bilateral filtering are employed to improve the mesh's regularity and smoothness while preserving the main geometric features of the surface.

[0128] The normal vector and curvature information of the 3D inner diameter surface are calculated to provide geometric properties for subsequent lighting rendering and morphological analysis. The normal vector is calculated based on the cross product operation of adjacent triangles, and the curvature is calculated using discrete differential geometry to obtain the principal curvature and mean curvature values ​​at each point on the surface.

[0129] A parametric coordinate system for the surface is established to map the 3D surface to a 2D parameter domain, facilitating texture mapping and feature analysis. The parametric process employs conformal or area-preserving mapping methods to ensure accurate preservation of geometric information during the 3D-to-2D mapping.

[0130] A pre-defined smoothing algorithm is applied to process the myocardial wall thickness mapping and the three-dimensional inner diameter surface to obtain a smooth three-dimensional inner diameter surface and myocardial wall thickness mapping.

[0131] A Gaussian filter was applied to the myocardial wall thickness mapping to eliminate the influence of noise and outliers on the wall thickness distribution. The kernel size and standard deviation parameters of the filter were set according to the typical range of variation of myocardial wall thickness, in order to maintain the authenticity of the wall thickness distribution while eliminating abnormal fluctuations caused by segmentation errors.

[0132] The Laplacian smoothing algorithm is applied to a 3D inner diameter surface to reduce local irregularities. The smoothing process iteratively adjusts the position of each vertex, moving it towards the centroid of adjacent vertices. The smoothing intensity is controlled by the number of iterations and the step size parameter.

[0133] In one embodiment, the traversal process of the inner diameter pixels adopts a multi-threaded parallel computing method, dividing the inner diameter pixel set into multiple subsets, with each thread responsible for calculating the shortest distance from the pixels in a subset to the outer diameter.

[0134] Specifically, for a left ventricle containing 10,000 inner diameter pixels, it can be divided into 10 subsets containing 1,000 pixels each, and distance calculation can be performed simultaneously by 10 parallel threads, which significantly improves computational efficiency.

[0135] For example, the color coding for myocardial wall thickness mapping uses the HSV color space, with a hue value ranging from 240 degrees (blue) to a minimum wall thickness of 2 mm, and a linear variation to 0 degrees (red) to a maximum wall thickness of 15 mm. Saturation and brightness values ​​are kept constant to ensure that color variations primarily reflect differences in wall thickness values. This coding method allows physicians to visually identify abnormal areas of myocardial wall thickness; normal myocardial wall thickness is typically displayed in the green to yellow range.

[0136] In one possible implementation, during the construction of the three-dimensional inner diameter surface, an adaptive mesh density is used for the different geometric features of the left ventricular base and apex. The base, with its relatively gentle geometric changes, uses a larger triangular mesh, while the apex, with its dramatic curvature changes, uses a denser triangular mesh, ensuring that the accuracy of the surface reconstruction meets the geometric complexity requirements of different regions.

[0137] It should be noted that the wall thickness mapping process takes into account the influence of myocardial fiber orientation. By introducing anisotropic distance metrics, the wall thickness calculation is made more consistent with the physiological structural characteristics of the myocardium. The anisotropic distance calculation adjusts the Euclidean distance according to the principal direction of the myocardial fibers, with a smaller weight for the fiber direction and a larger weight for the distance perpendicular to the fiber direction, reflecting the physiological characteristic of myocardial lesions spreading along the fiber direction.

[0138] Please see Figure 4 This demonstrates the three-dimensional mapping generation process of myocardial wall thickness, such as... Figure 4 As shown, based on determining the long axis of the left ventricle and the inner and outer diameters of the myocardium through boundary structure information, one possible implementation involves traversing each pixel on the inner diameter surface and calculating its distance to the nearest point on the outer diameter surface. This distance is defined as the myocardial wall thickness at the corresponding location. In this way, a three-dimensional wall thickness map is constructed, where different colors represent different thicknesses, such as purple for 6 mm and red for 20 mm. This map visually displays the spatial distribution of myocardial thickness, helping to identify abnormal areas, such as thickened or thinned sections.

[0139] It should be noted that this mapping not only provides a visual basis but also offers data support for feature extraction in cardiomyopathy classifiers, significantly improving detection accuracy. Combining the above steps, the significance of constructing a wall thickness mapping after determining the major axis and inner and outer diameters lies in its ability to comprehensively capture the detailed features of the heart structure.

[0140] For example, using the major axis as a spatial reference line for the heart, combined with wall thickness distribution, can help analyze the health status of different regions of the myocardium segmentally, such as the differences between the interventricular septum and non-interventricular septum regions. This method extracts features such as mean and variance through spatial dimensionality reduction and regional statistics, providing reliable input for the training of subsequent intelligent classifiers, ultimately improving the efficiency and accuracy of cardiomyopathy detection.

[0141] S103. Determine the pre-trained classification model and construct the dual-channel input of the pre-trained classification model; input the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness into the first input channel of the pre-trained classification model.

[0142] Please see Figure 5 , Figure 5 The image shows a three-dimensional inner diameter surface and a myocardial wall thickness mapping in the lower left corner.

[0143] In one embodiment, determining a pre-trained classification model and constructing a dual-channel input for the pre-trained classification model specifically includes:

[0144] A classification model based on a 3D convolutional neural network was selected. This model was pre-trained on a large dataset of 3D cardiac images to identify cardiomyopathy types. The training process involved inputting labeled cardiac segmentation data and corresponding lesion labels, and adjusting the weights through backpropagation to minimize classification error.

[0145] Based on the characteristics of 3D cardiac images, a dual-channel input structure was designed. The first channel is dedicated to processing the 3D inner diameter surface and myocardial wall thickness mapping, while the second channel is used for whole-heart segmentation results or raw image data to capture morphological and texture information.

[0146] Please see Figure 6 , Figure 6 The original image data and the whole heart segmentation results in the lower left corner are displayed.

[0147] Verify the compatibility of the dual-channel input and ensure data dimension consistency, for example, by adjusting all inputs to a uniform resolution such as 128x128x128 voxels, and construct a complete dual-channel input for subsequent processing.

[0148] For example, in cardiac magnetic resonance imaging, the dual-channel input constructed in this step can integrate wall thickness distribution and overall chamber morphology, improving the sensitivity of the classification model for hypertrophic cardiomyopathy and helping to reduce false positive results.

[0149] In one embodiment, the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness is input into the first input channel of a pre-trained classification model, specifically including:

[0150] The three-dimensional inner diameter surface and myocardial wall thickness mapping are preprocessed, such as by applying a Gaussian smoothing filter to remove noise, where the filter kernel size is 3x3x3, in order to preserve the surface geometric details and wall thickness intensity distribution.

[0151] The composite input for the first channel is generated by fusing the surface and mapping data. For example, the voxel values ​​of the surface are multiplied point by point with the distance values ​​of the wall thickness mapping to form an enhanced feature representation that reflects the local thickness and spatial morphology of the left ventricular myocardium.

[0152] The fused data is input into the first channel of the classification model, and features are extracted through the model's convolutional layers, such as using a 3x3x3 convolutional kernel for downsampling, to generate a high-dimensional feature map for subsequent classification.

[0153] By combining the input from the second channel, such as the whole heart segmentation results, the dual-channel features are integrated through the model's fusion layer, for example, by using element-wise addition, to achieve comprehensive analysis and output cardiomyopathy classification results.

[0154] For example, when dealing with cases of dilated cardiomyopathy, the wall thickness mapping input in this step shows a uniform thinning distribution. When combined with the surface, the model can accurately distinguish the normal type. The beneficial effect is to improve the detection accuracy, especially in images with large variations in wall thickness.

[0155] In one embodiment, the process is further extended to consider image data of different resolutions, detect the resolution of the input surface and the mapping, and if it is lower than a standard such as 1mm equidistant, apply trilinear interpolation to upsample to the matching dimension to ensure input consistency.

[0156] Statistical features of the wall thickness mapping, such as mean thickness and standard deviation, are calculated and appended as auxiliary metadata to the first channel input to enhance the model’s ability to identify anomalous wall thicknesses.

[0157] For example, this extension allows the model to process low-resolution CT images, and the statistical features of wall thickness mapping help highlight local thickening patterns in restrictive cardiomyopathy, which is beneficial for adapting to diverse device outputs in clinical applications and improving classification robustness.

[0158] In one embodiment, implemented in another scenario, threshold filtering is applied to the myocardial wall thickness mapping, retaining only areas with a thickness exceeding a preset value, such as more than 2 mm, as significant feature inputs to reduce noise interference.

[0159] The filtered mapping is superimposed on the surface to form the tensor input of the first channel, and the classification probability is generated through the forward propagation of the model.

[0160] For example, in the detection of amyloid cardiomyopathy, this filtering highlights uneven wall thickness, and the beneficial effect is that the model can better capture texture anomalies, resulting in an improvement in classification accuracy of about 15%, as verified on a simulated dataset.

[0161] S104. The dual-channel input is processed by a pre-trained classification model to output the cardiomyopathy classification model result; the dual-channel input of the pre-trained classification model includes the first input channel which maps the three-dimensional inner diameter surface to the myocardial wall thickness.

[0162] In one embodiment, the dual-channel input is processed by a pre-trained classification model to output the cardiomyopathy classification model result, specifically including:

[0163] The inference process of the classification model is initiated based on the input dual-channel data. The first channel provides combined information of the three-dimensional inner diameter surface and myocardial wall thickness mapping. This information captures the morphology and thickness distribution of the left ventricular myocardium and is used by the model to extract spatial features.

[0164] 3D convolution operations are applied to dual-channel data to generate a fused feature map that integrates surface geometry and wall thickness strength, helping the model distinguish different myocardial lesion patterns.

[0165] The fully connected layer is calculated based on the feature map, and the probability distribution is output to determine the type of cardiomyopathy, such as hypertrophic or dilated cardiomyopathy.

[0166] In one embodiment, the dual-channel input of the pre-trained classification model includes a first input channel relating the three-dimensional inner diameter surface to the myocardial wall thickness, specifically including:

[0167] Prepare the first channel data by representing the three-dimensional inner diameter surface as a mesh structure, where each mesh point corresponds to the spatial coordinates of the inner diameter pixel. At the same time, map and superimpose the myocardial wall thickness as an intensity value to form a unified input tensor.

[0168] Ensure that the input size of this channel is standardized by adjusting it to a fixed resolution, such as 32x32x32, through adaptive pooling to match the model architecture.

[0169] For example, when processing cardiac CT images, a three-dimensional inner diameter surface can be constructed from the pixels of the left ventricular myocardial inner diameter, preserving the original shape, while myocardial wall thickness mapping calculates the shortest distance from each inner diameter point to the outer diameter. For example, areas with an average wall thickness of 10 mm are marked with high intensity, which helps the model identify local thickening features of hypertrophic cardiomyopathy.

[0170] In one embodiment, it is understood that the pre-trained classification model adopts a 3D ResNet architecture, which processes 3D data through residual connections. First, it extracts local features of the surface and wall thickness in the first channel, such as a convolution kernel size of 3x3x3, and progressively downsamples to capture multi-scale information. Then, it fuses the whole heart segmentation data from the second channel and finally outputs the classification result. This design improves the detection accuracy because it combines morphological and texture information and avoids the limitations of a single channel.

[0171] For example, in a case of dilated cardiomyopathy, the input three-dimensional inner diameter surface shows the enlarged chamber, and its wall thickness mapping shows uniform thinning, such as an average of 5 mm. After model processing, the output probability is 0.85, which belongs to dilated type. This is more accurate than traditional two-dimensional analysis and can identify ventricular remodeling earlier.

[0172] In one embodiment, the first channel construction can be extended to different imaging modalities, such as MRI data, where the surface construction uses an algorithm to generate isosurfaces from inner diameter pixels, and the myocardial wall thickness mapping calculates the shortest path through Euclidean distance transformation to ensure that the mapping reflects the true thickness distribution, such as in hypertrophic cases where areas with a wall thickness peak of 20 mm are highlighted.

[0173] For example, in high-resolution imaging scenarios, the surface mesh density is increased to 100 points per millimeter, and the wall thickness mapping resolution is correspondingly improved. This allows the model to capture subtle variations, such as the local non-uniform thickness of amyloidosis, thereby outputting more reliable classification results and improving the efficiency of clinical decision-making.

[0174] In one embodiment, the surface and mapping data of the first channel are cropped and input into the model, and the output is used for parameter optimization, such as adjusting weights using cross-entropy loss, which ensures the model's generalization ability on diverse cardiac data.

[0175] For example, in normal cardiac images, the wall thickness mapping shows a uniform distribution of 8-12 mm, and the model correctly classifies it as normal, proving that the dual-channel input can effectively distinguish between pathological and physiological states, bringing higher specificity and sensitivity.

[0176] Step S102 above, which involves constructing a three-dimensional inner diameter surface and generating a myocardial wall thickness mapping based on the inner and outer diameters of the left ventricular myocardium, includes:

[0177] Iterate through all pixels of the left ventricular myocardial inner diameter, and for each inner diameter pixel, calculate the shortest distance between it and the outer diameter pixel.

[0178] Myocardial wall thickness mapping is generated based on the shortest distance. The myocardial wall thickness mapping reflects the spatial distribution of inner diameter pixels and the corresponding wall thickness intensity.

[0179] A three-dimensional inner diameter surface is constructed based on the spatial distribution of inner diameter pixels, and the three-dimensional inner diameter surface preserves the original spatial morphology of the left ventricular myocardium.

[0180] A pre-defined smoothing algorithm is applied to process the myocardial wall thickness mapping and the three-dimensional inner diameter surface to obtain a smooth three-dimensional inner diameter surface and myocardial wall thickness mapping.

[0181] The pre-trained classification model has two input channels, including a second input channel for the whole heart segmentation results or the raw medical image data corresponding to the whole heart segmentation.

[0182] In one embodiment, the construction process of the second input channel includes the following sub-steps to ensure the integrity of the input data and its complementarity with the first channel, thereby improving the accuracy of the model in classifying cardiomyopathy.

[0183] The whole heart segmentation result is obtained by processing the raw medical image data using a pre-established segmentation algorithm, which includes the boundary information of each chamber of the heart.

[0184] The whole heart segmentation results or their corresponding raw medical imaging data are standardized, including adjusting the resolution to a uniform size, such as normalizing the image voxel size to 0.3mm x 0.3mm x 0.3mm, to match the format of the three-dimensional inner diameter surface and myocardial wall thickness mapping of the first channel.

[0185] The standardized data is input into the second channel of the pre-trained classification model. This channel extracts chamber morphology or texture features, which are then fused with the myocardial wall thickness information from the first channel to form high-dimensional features for subsequent classification.

[0186] For example, in cardiac MRI imaging scenarios, whole-heart segmentation results can provide precise boundaries of the left ventricle, right ventricle, and atrium, helping the model capture overall cardiac structural abnormalities, while the original image data supplements grayscale texture details, such as highlighting chamber dilation features in dilated cardiomyopathy. This fusion can improve classification accuracy by up to 15% because it avoids the loss of local information from a single channel.

[0187] In one embodiment, the specific method for obtaining whole-heart segmentation results is further refined to adapt to different imaging modalities, specifically including:

[0188] The entire heart region was extracted from the raw medical imaging data, and pixel-level annotations were performed using a segmentation algorithm based on the U-Net architecture to generate labeled maps including the left ventricle, right ventricle, left atrium, and right atrium.

[0189] To verify the accuracy of the whole heart segmentation results, the Dice coefficient was calculated to ensure that the overlap with the ground truth exceeded 0.9. If it was lower than the threshold, the segmentation was re-implemented.

[0190] For example, in CT images, whole-heart segmentation results can highlight the texture of calcified areas, while in MRI, it emphasizes soft tissue contrast. This validation step ensures the reliability of the second-channel input, enabling the model to better identify wall thickness heterogeneity when dealing with hypertrophic cardiomyopathy, resulting in higher detection sensitivity.

[0191] The dual-channel input is processed by a pre-trained classification model to output the cardiomyopathy classification model results, including:

[0192] Based on the spatial distribution of the left ventricular myocardium's inner diameter, a three-dimensional bounding box containing the inner diameter of the left ventricular myocardium is extracted to obtain the cropped three-dimensional data; the three-dimensional bounding box covers the x-axis, y-axis, and z-axis directions of the inner diameter.

[0193] The Adaptive AvgPool3d operator is applied to process the clipped 3D data and myocardial wall thickness mapping to generate output data in a unified format.

[0194] The output data in a uniform format is input into the pre-trained classification model through the first input channel; and,

[0195] The system acquires whole-heart segmentation results or corresponding raw medical image data, which are used as the second channel input and fed into a pre-trained classification model. The second channel input includes chamber morphology information or texture information. The pre-trained classification model processes the dual-channel input and outputs the cardiomyopathy classification model results.

[0196] In one embodiment, based on the spatial distribution of the left ventricular myocardium's inner diameter, a three-dimensional bounding box containing the inner diameter of the left ventricular myocardium is cropped to obtain the cropped three-dimensional data; the bounding box covers the x-axis, y-axis, and z-axis directions of the inner diameter, specifically including:

[0197] The coordinate distribution of pixels in the left ventricular myocardial diameter is analyzed to determine the minimum and maximum boundary values ​​of the x-axis, y-axis, and z-axis. These boundary values ​​are calculated by traversing the three-dimensional coordinates of all inner diameter pixels. For example, the minimum x-value is the minimum x-coordinate of all pixels, and the maximum x-value is the maximum x-value, and so on, to ensure that the bounding box completely surrounds the inner diameter structure.

[0198] A three-dimensional bounding box is generated based on the determined boundary values, and the three-dimensional inner diameter surface data is cropped to obtain cropped three-dimensional data containing only the left ventricular myocardial inner diameter region. This cropping reduces irrelevant background information and improves the efficiency of subsequent processing.

[0199] For example, in the processing of 3D cardiac images, if the inner diameter pixels are distributed within the range of 0 to 100 on the x-axis, 50 to 150 on the y-axis, and 20 to 80 on the z-axis, then the bounding box is set to these ranges. After cropping, the data volume is reduced, which is beneficial for the standardization of model input.

[0200] In one embodiment, the Adaptive AvgPool3d operator is applied to process the clipped 3D data and myocardial wall thickness mapping to generate output data in a uniform format, specifically including:

[0201] The cropped 3D data and myocardial wall thickness mapping are used as input;

[0202] The AdaptiveAvgPool3d operator is an adaptive average pooling operation that dynamically adjusts the pooling step size and kernel size based on the specified output size to downsample the 3D volume data. For example, if the input data has an arbitrary shape, the operator will calculate the average value of each output voxel to ensure that the output is a tensor of a fixed size, such as 32x32x32.

[0203] The same operator is applied to myocardial wall thickness mapping to convert its wall thickness distribution into a uniform format. In the process, the operator first determines the ratio of the input size to the target output size, and then performs average pooling in each dimension. For example, if the input is 100 and the output is 32 in the x dimension, the step size is about 3, and the kernel size is adjusted accordingly to generate output data that reflects the average distribution of wall thickness.

[0204] By merging the processed 3D data and wall thickness mapping, output data in a unified format is obtained. This unification helps the model handle the variability in cardiac images of different patients.

[0205] For example, assuming the cropped data is 128x128x64, if the Adaptive AvgPool3d operator is set to output 16x16x16, it will sample uniformly, and each element in the output is the average value of the corresponding input region. In this way, even if the original image resolution is different, it can produce a consistent input, which is beneficial to the generalization ability of the classification model.

[0206] In one embodiment, uniformly formatted output data is input to a pre-trained classification model via a first input channel. The generated uniformly formatted output data is directly fed into the model's first channel, which is dedicated to processing inner diameter surface and wall thickness information, ensuring a smooth data flow into subsequent convolutional layers.

[0207] For example, this input method preserves the three-dimensional characteristics of the data and avoids information loss.

[0208] In one embodiment, the whole-heart segmentation result or the corresponding raw medical image data is obtained and used as the second channel input to a pre-trained classification model. The second channel input includes chamber morphology information or texture information. Specifically, this includes extracting the whole-heart segmentation result from a pre-stored database. This result is a labeled volume generated by a segmentation algorithm, containing morphological information of chambers such as the left ventricle and right ventricle, such as volume and shape features. Alternatively, the corresponding raw medical image data can be extracted, containing pixel-level texture information such as grayscale gradients. When using the whole-heart segmentation result, the second channel input emphasizes chamber morphology, such as the degree of left ventricular dilation. When using raw data, texture information, such as the density distribution of myocardial fibers, is input. The selected data is then input into the model's second channel for parallel processing with the first channel.

[0209] Before input, the second channel data undergoes similar preprocessing as the first channel, such as normalization, to match the model's desired format.

[0210] For example, in the hypertrophic cardiomyopathy scenario, using the whole heart segmentation results as the second channel can capture the thickened morphology of the ventricular wall, which is helpful in distinguishing it from the normal type; while in the myocarditis scenario, using the original image data to highlight the changes in inflammatory texture, this dual-selection input enhances the model's sensitivity to different cardiomyopathy types and improves classification accuracy.

[0211] In one embodiment, a pre-trained classification model processes the dual-channel input to output a cardiomyopathy classification model result. Specifically, after receiving the dual-channel input, the model extracts features through a 3D convolutional layer. The first channel provides the inner diameter and wall thickness distribution, and the second channel supplements the morphology or texture. After fusion, a high-dimensional feature vector is generated.

[0212] High-dimensional features are classified using fully connected layers, and the output probability distribution corresponds to the cardiomyopathy type, such as hypertrophic cardiomyopathy with a probability of 0.8 and normal with a probability of 0.1.

[0213] The probability is normalized using the Softmax function, the highest probability is selected as the final result, and the model parameters can be optimized through backpropagation.

[0214] For example, in one possible implementation, for patients with dilated cardiomyopathy, dual-channel input captures the morphology of ventricular enlargement and wall thinning, and the model outputs a classification result of dilated cardiomyopathy, which helps in early detection.

[0215] It should be noted that this processing method integrates information from multiple sources, reduces the risk of misjudgment from a single channel, and improves the overall classification robustness.

[0216] In one embodiment, if the inner diameter of the three-dimensional spatial frame extracted in the above steps is not uniformly distributed, such as being concentrated in the lower half on the z-axis, the boundary of the three-dimensional spatial frame can be dynamically adjusted to minimize the volume.

[0217] For example, for an image with a ventricular major axis of 150 pixels, setting the z-axis boundary to 0 to 150 will focus the data on the core region after cropping, which is beneficial to computational efficiency.

[0218] For example, in the Adaptive AvgPool3d application, the average input wall thickness mapping is 8 mm, and the output after pooling reflects the average wall thickness of the region as 6 mm. This captures local variations and is used to classify hypertrophic cases.

[0219] Specifically, for the second channel selection, if the patient's image has high noise, the whole heart segmentation result is used first to avoid texture interference and the output result is more stable.

[0220] In one embodiment, when the model processes dual channels, it first extracts the features of each channel independently, and then fuses them through a concat operation. For example, the first channel has a feature dimension of 512, the second channel has a feature dimension of 512, and after fusion, the result is 1024, which is used to classify restrictive cardiomyopathy.

[0221] For example, this fusion improved accuracy from 85% to 92%.

[0222] For example, when the output includes amyloidosis type, combining wall thickness and texture information can distinguish subtle differences from myocarditis, which is beneficial for precision medicine.

[0223] After obtaining the whole heart segmentation results or the corresponding raw medical image data as the second channel input, and feeding them into the pre-trained classification model, the process also includes:

[0224] High-dimensional features are obtained by extracting features from dual-channel inputs using a 3D convolutional network. These high-dimensional features include the spatial morphology and wall thickness distribution information of the left ventricular myocardium.

[0225] Classification is performed based on high-dimensional features to generate cardiomyopathy classification results;

[0226] The cardiomyopathy classification results include hypertrophic cardiomyopathy, dilated cardiomyopathy, restrictive cardiomyopathy, amyloidosis, myocarditis, and normal type;

[0227] Based on the cardiomyopathy classification results, a preset optimization algorithm is applied to adjust the classification model parameters to obtain optimized cardiomyopathy classification model results, which are then output as cardiomyopathy classification model results.

[0228] In one embodiment, a 3D convolutional network is used to extract features from the dual-channel input to obtain high-dimensional features. These high-dimensional features include spatial morphology and wall thickness distribution information of the left ventricular myocardium. Specifically, the process involves receiving dual-channel input, where the first channel is a combination of data mapping a three-dimensional inner diameter surface to myocardial wall thickness, and the second channel is the whole-heart segmentation result or the corresponding raw medical image data. A 3D convolutional layer is applied to the input data to perform convolution operations, extracting local spatial features. For example, multiple 3x3x3 convolutional kernels are used to slide and scan the data, capturing the surface geometry and wall thickness variations of the left ventricular myocardium. Pooling layers are introduced to reduce feature dimensionality while retaining key information, such as max pooling merging the values ​​of adjacent voxels. The convolution and pooling processes are repeated to form a multi-layer network structure, gradually abstracting high-dimensional features. These features integrate spatial morphology, such as the curvature of the myocardial surface, and wall thickness distribution, such as local thickness differences, to obtain a vector representation of the overall cardiac structure.

[0229] Classification is performed based on high-dimensional features to generate cardiomyopathy classification results.

[0230] The cardiomyopathy classification results include hypertrophic cardiomyopathy, dilated cardiomyopathy, restrictive cardiomyopathy, amyloidosis, myocarditis, and normal type.

[0231] Based on the cardiomyopathy classification results, a preset optimization algorithm is applied to adjust the classification model parameters to obtain optimized cardiomyopathy classification model results, which are then output as cardiomyopathy classification model results, specifically including:

[0232] Compare the classification results with the true labels and calculate the loss function value, such as using cross-entropy loss to quantify the prediction bias.

[0233] Based on the loss value, the gradient is calculated using the backpropagation algorithm, and the network weights are updated in reverse. For example, the error is propagated from the output layer to the input layer, and the convolution kernel parameters are adjusted to minimize the error.

[0234] The iterative optimization process continues until the model converges, generating optimized classification results. This adjustment improves the model's accuracy in classifying cardiac images.

[0235] A three-dimensional inner diameter surface is constructed based on the spatial distribution of inner diameter pixels, including:

[0236] Based on the coordinate information of the inner diameter pixels, a three-dimensional mesh structure is generated, which represents the inner diameter surface of the left ventricular myocardium.

[0237] For the three-dimensional mesh structure, a preset interpolation algorithm is applied to fill in the missing surface data;

[0238] Based on the spatial distribution of the inner diameter pixels, the geometry of the three-dimensional mesh structure is adjusted to obtain a three-dimensional inner diameter surface that retains the original spatial shape.

[0239] For a three-dimensional inner diameter surface, a corresponding color code is generated, which reflects the spatial position information of the surface.

[0240] In one embodiment, a three-dimensional mesh structure is generated based on the coordinate information of the inner diameter pixels, the three-dimensional mesh structure representing the inner diameter surface of the left ventricular myocardium.

[0241] Specifically, by collecting the three-dimensional coordinate data of the inner diameter pixels, such as their x, y, and z axis positions, these points are connected to form a triangular mesh, creating a preliminary surface representation. This mesh structure captures the overall contour of the left ventricular myocardial inner diameter, providing a basic surface model for subsequent processing.

[0242] In one embodiment, the step of processing the dual-channel input using a pre-trained classification model and outputting the cardiomyopathy classification model result includes:

[0243] S201. Calculate the absolute deviation of the original wall thickness value of each sampling point in the myocardial wall thickness mapping relative to the preset standard value D of the myocardial wall thickness; normalize the absolute deviation to determine the spatial weight of each sampling point; map the spatial weight to a three-dimensional spatial coordinate system and perform spatial interpolation to generate a spatial guidance mask matrix M consistent with the spatial scale of the three-dimensional image data of the heart; specifically:

[0244] First, sampling points on the inner diameter surface of the left ventricle are obtained, and the shortest distance from each sampling point to the outer diameter surface along the normal direction is calculated to obtain the original wall thickness value d of each sampling point. Then, a preset standard value D for myocardial wall thickness is introduced. The preset standard value D for myocardial wall thickness is the average wall thickness constant obtained based on the statistics of healthy heart samples of the same population.

[0245] Next, the wall thickness offset S at each sampling point is calculated using the following formula: In order to convert the deviation of geometric dimension into weight coefficients that the model can handle, a normalization function is used to map the wall thickness offset S to obtain the sampling point weight value W, so that the sampling point weight value W is distributed in the interval [0,1], and the larger the sampling point weight value W, the more the geometric shape of the position deviates from the normal physiological characteristics.

[0246] Subsequently, a three-dimensional empty pixel matrix is ​​constructed that is completely consistent with the spatial resolution and physical size of the original medical image data. The sampling point weight value W corresponding to each sampling point is filled into the corresponding coordinate position in the three-dimensional empty pixel matrix, and a trilinear interpolation algorithm is used to fill the gaps between sampling points, thereby generating a continuously distributed spatial guidance mask matrix M.

[0247] Finally, the generated spatial guidance mask matrix M is smoothed at the boundary, and abrupt changes caused by interpolation are eliminated by a three-dimensional Gaussian filter to ensure that the weight distribution of the spatial guidance mask matrix M in three-dimensional space can truly reflect the continuous deformation trend of myocardial wall thickness, providing spatial dimension probability guidance for subsequent interactive calibration with the features of the original medical image data.

[0248] S202. The three-dimensional image data of the heart is voxel-normalized and input into an extraction path composed of multiple residual convolutional units. The residual convolutional units extract a deep spatial feature tensor T with multi-channel texture features while maintaining the continuity of the anatomical structure. Specifically:

[0249] First, the acquired 3D image data of the heart is isotropically resampled, the voxel spacing is standardized to a preset resolution, and grayscale value standardization is performed to eliminate brightness differences caused by imaging from different devices.

[0250] Next, the processed image data is input into a pre-constructed 3D convolutional path consisting of three sets of residual blocks. Each set of residual blocks contains two 3×3×3 convolutional layers and a cross-layer identity connection. The identity connection directly transmits shallow anatomical information to deep layers, preventing gradient vanishing when extracting complex texture features.

[0251] Subsequently, spatial downsampling is performed by setting convolution operations with a stride of 2 after the first and second residual blocks, so that the final output deep spatial feature tensor T is 1 / 4 of the original three-dimensional image data size of the heart in the depth, height and width dimensions. At this time, the cumulative downsampling factor is recorded as a feedback signal to adjust the size change parameters of the spatial guidance mask matrix M.

[0252] Finally, a deep spatial feature tensor T with dimensions C×D×H×W is extracted from the last residual block; each channel C in the deep spatial feature tensor T represents a specific pathological texture pattern (such as myocardial fiber disorder or local density fluctuation), and its spatial coordinates (D, H, W) can be accurately traced back to the anatomical location of the three-dimensional image data of the heart through linear scaling.

[0253] S203. Expand the channel dimension of the spatial guidance mask matrix M to obtain an enhanced weight tensor M' with the same dimension as the deep spatial feature tensor T; after introducing a dynamic adjustment factor β to weight the enhanced weight tensor M', perform per-spatial coordinate position multiplication modulation with the deep spatial feature tensor T to generate a calibrated spatial interaction feature map F; specifically:

[0254] First, read the spatial guidance mask matrix M and the deep spatial feature tensor T. For the case where the spatial guidance mask matrix M is single-channel data and the deep spatial feature tensor T is C-channel data, copy the spatial guidance mask matrix M C times along the channel dimension to construct an enhanced weight tensor M' with the same dimensions as the deep spatial feature tensor T.

[0255] Next, a preset dynamic adjustment factor β is introduced, which is used to control the intensity of the intervention of abnormal myocardial wall thickness features on the texture features of the three-dimensional image data of the heart; each element in the enhanced weight tensor M' is multiplied by the dynamic adjustment factor β to obtain the weighted mask tensor.

[0256] Subsequently, the weighted mask tensor and the deep spatial feature tensor T are multiplied element-wise at their corresponding spatial coordinates to generate a calibrated spatial interaction feature map F. During the operation, since the weight values ​​in the spatial guidance mask matrix M that reflect the abnormal myocardial wall thickness region are close to 1, while those in the normal region are close to 0, the element-wise multiplication preserves or amplifies the texture features in the deep spatial feature tensor T located at the anatomical location of the abnormal myocardial wall thickness, while weakening the texture features located in the normal myocardial wall thickness region.

[0257] Finally, the generated spatial interaction feature map F is subjected to nonlinear activation processing. The modified linear unit (ReLU) is used to suppress the negative signal that may be generated during the operation, so as to ensure that the output spatial interaction feature map F can centrally represent the deep pathological information guided by geometric abnormalities, and realize the deep mutual coupling of morphological features and texture features in the spatial dimension.

[0258] S204. Perform global spatial dimension pooling aggregation on the spatial interaction feature map F to eliminate spatial redundancy and extract the core pathological feature vector V representing the global pathological intensity; map the core pathological feature vector V to a preset category space, and then output the cardiomyopathy classification model result based on the response weights corresponding to each category; specifically:

[0259] First, the spatial interaction feature map F is received; the spatial interaction feature map F is a multi-channel voxel matrix after being weighted and calibrated by the geometric information of myocardial wall thickness.

[0260] Next, global average pooling is performed on each feature channel of the spatial interaction feature map F to calculate the average activation value of all spatial locations within that channel. This process compresses the three-dimensional spatial tensor of dimension C×D×H×W into a one-dimensional feature sequence of dimension C×1×1×1, yielding the core pathological feature vector V. Each dimension of the core pathological feature vector V represents the global intensity of a specific pathological texture constrained by geometric morphology.

[0261] Subsequently, the core pathological feature vector V is input into a pre-defined fully connected mapping layer; the fully connected mapping layer contains multiple sets of learnable weight parameters, which are used to map the high-dimensional cardiopathological feature vector V to the original score space consistent with the number of cardiomyopathy categories to be classified.

[0262] Finally, the original scores are exponentially normalized using the Softmax function to calculate the probability distribution of each cardiomyopathy type. The category with the highest probability value is selected as the final cardiomyopathy classification model output.

[0263] The entire process achieves a semantic progression from "geometry-guided local texture extraction" to "global pathological feature modeling," ensuring that the decision-making basis has spatial sensitivity and pathological relevance.

[0264] By employing a three-dimensional feature mutual inductance mechanism, this approach avoids the dimensional collapse problem caused by two-dimensional projection or regional statistics of cardiac geometric features in existing technologies, thus fully preserving the texture details in high-dimensional space. This processing method enables the model to accurately capture subtle pathological changes hidden in areas of abnormal myocardial wall thickness, such as signal disturbances caused by early focal hypertrophy. It fundamentally solves the technical limitations of existing technologies, such as the scheme disclosed in patent document CN121033254A, which cannot identify local discrete lesions due to "dimensionality reduction statistics," and significantly improves diagnostic sensitivity.

[0265] Furthermore, such as Figure 7 As shown, this embodiment of the invention also provides a processing device based on three-dimensional cardiac images, including: an acquisition module 710, used to acquire three-dimensional image data of the heart and three-dimensional segmentation results of cardiac chambers, the three-dimensional segmentation results of cardiac chambers including the left ventricular chamber and the left ventricular myocardium; a processing module 720, used to construct a three-dimensional inner diameter surface and generate a myocardial wall thickness mapping based on the inner and outer diameters of the left ventricular myocardium, the myocardial wall thickness mapping representing the shortest distance from the inner diameter pixel to the outer diameter; a training module 730, used to determine a pre-trained classification model and construct a dual-channel input for the pre-trained classification model; the training module 730 is also used to input the three-dimensional inner diameter surface and the myocardial wall thickness mapping into the first input channel of the pre-trained classification model; the training module 730 is also used to process the dual-channel input through the pre-trained classification model and output the cardiomyopathy classification model result; the dual-channel input of the pre-trained classification model includes the first input channel regarding the three-dimensional inner diameter surface and the myocardial wall thickness mapping.

[0266] Furthermore, such as Figure 8As shown, this embodiment of the invention also provides a computing device, which may include: a memory 801 storing executable program code; and a processor 802 coupled to the memory 801, wherein the processor 802 calls the executable program code stored in the memory 801 to execute a program based on... Figure 1 The steps are shown.

[0267] This application also discloses a computer program product that, when run on a computer, causes the computer to execute... Figure 1 Some or all of the steps of any of the disclosed methods.

[0268] This application also discloses an application publishing platform for publishing computer program products. When the computer program product is run on a computer, it causes the computer to execute... Figure 1 Some or all of the steps of any of the disclosed methods.

[0269] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0270] In summary, this invention discloses a method, apparatus, device, and medium for classifying cardiomyopathy based on three-dimensional cardiac imaging data. Addressing the practical challenge of accurately and efficiently extracting left ventricular myocardial features from complex cardiac images and classifying cardiomyopathy, this invention constructs a three-dimensional inner diameter surface mapping to myocardial wall thickness and combines it with a pre-trained classification model using dual-channel input, achieving accurate classification of hypertrophic and dilated cardiomyopathy. First, this invention processes the original medical images using a segmentation algorithm to extract the inner and outer diameters of the left ventricular myocardium, generating a mapping reflecting wall thickness distribution and an inner diameter surface that preserves spatial morphology. Then, it uses a three-dimensional bounding box to crop the data, combines the whole-heart segmentation results, and extracts high-dimensional features through a 3D convolutional network, finally outputting the classification result. This invention optimizes the surface and mapping quality through a smoothing algorithm and applies Adaptive AvgPool3d to ensure data consistency, significantly improving the robustness and accuracy of the classification model and providing efficient and accurate technical support for cardiomyopathy detection.

[0271] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and additions without departing from the principle of the present invention, and these improvements and additions should also be considered within the scope of protection of the present invention.

Claims

1. A method for processing cardiac three-dimensional images, characterized in that, Includes the following steps: Acquire three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers, wherein the three-dimensional segmentation results of the heart chambers include the left ventricular chambers and the left ventricular myocardium; Based on the inner and outer diameters of the left ventricular myocardium, a three-dimensional inner diameter surface is constructed and a myocardial wall thickness mapping is generated, wherein the myocardial wall thickness mapping represents the shortest distance from the inner diameter pixel to the outer diameter. Determine the pre-trained classification model and construct the dual-channel input of the pre-trained classification model; The mapping between the three-dimensional inner diameter surface and the myocardial wall thickness is input into the first input channel of the pre-trained classification model; The dual-channel input is processed by the pre-trained classification model to output the cardiomyopathy classification model result; The pre-trained classification model's dual-channel input includes a first input channel relating the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness.

2. The processing method based on three-dimensional cardiac images as described in claim 1, characterized in that, The acquisition of three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers includes: Obtain raw medical imaging data of the heart; The original medical image data is processed using a pre-established segmentation algorithm to generate a three-dimensional segmentation result of the heart chambers; Based on the three-dimensional segmentation results of the heart chambers, the boundary structure information of the left ventricular chamber and the left ventricular myocardium is extracted; The inner and outer diameters of the left ventricular myocardium are determined based on the boundary structure information. The inner diameter is the boundary between the left ventricular myocardium and the left ventricular chamber, and the outer diameter is the boundary between the left ventricular myocardium and the right ventricle or background.

3. The method for processing cardiac three-dimensional images as described in claim 2, characterized in that, The inner and outer diameters of the left ventricular myocardium are determined based on the boundary structure information, including: The centroid of the left ventricular base is extracted as the first reference point based on the boundary structure information. The point in the left ventricle that is farthest from the first reference point is determined as the second reference point; The long axis of the left ventricle is constructed using the first reference point and the second reference point; Based on the boundary structure information, the inner diameter of the left ventricular myocardium on the left ventricular cavity side is determined; The outer diameter is defined as the boundary that distinguishes the left ventricular myocardium from the right ventricle and the background.

4. The method for processing cardiac three-dimensional images as described in claim 1, characterized in that, Based on the inner and outer diameters of the left ventricular myocardium, a three-dimensional inner diameter surface is constructed and a myocardial wall thickness mapping is generated, including: Iterate through all pixels of the left ventricular myocardial inner diameter, and for each inner diameter pixel, calculate the shortest distance between it and the outer diameter pixel. Myocardial wall thickness mapping is generated based on the shortest distance, and the myocardial wall thickness mapping reflects the spatial distribution of inner diameter pixels and the corresponding wall thickness intensity. A three-dimensional inner diameter surface is constructed based on the spatial distribution of the inner diameter pixels, and the three-dimensional inner diameter surface preserves the original spatial morphology of the left ventricular myocardium; A preset smoothing algorithm is applied to the myocardial wall thickness mapping and the three-dimensional inner diameter surface to obtain a smooth three-dimensional inner diameter surface and myocardial wall thickness mapping.

5. The processing method based on three-dimensional cardiac images as described in claim 1, characterized in that, The pre-trained classification model has a dual-channel input including a second input channel for the whole heart segmentation results or the original medical image data corresponding to the whole heart segmentation.

6. The processing method based on three-dimensional cardiac images as described in claim 5, characterized in that, The dual-channel input is processed by the pre-trained classification model to output the cardiomyopathy classification model results, including: Based on the spatial distribution of the inner diameter of the left ventricular myocardium, a three-dimensional spatial frame containing the inner diameter of the left ventricular myocardium is extracted to obtain the cropped three-dimensional data. The three-dimensional spatial frame covers the x-axis, y-axis, and z-axis directions of the inner diameter; The cropped 3D data and myocardial wall thickness mapping are processed by the Adaptive AvgPool3d operator to generate output data in a unified format. The output data in the unified format is input into the pre-trained classification model through the first input channel; and The whole heart segmentation result or the corresponding raw medical image data is obtained and used as the second channel input to the pre-trained classification model; The second channel input includes chamber morphology information or texture information; The dual-channel input is processed by the pre-trained classification model to output the cardiomyopathy classification model result.

7. The processing method based on three-dimensional cardiac images as described in claim 6, characterized in that, After obtaining the whole heart segmentation result or the corresponding raw medical image data as the second channel input and inputting it into the pre-trained classification model, the process further includes: The dual-channel input is subjected to feature extraction using a 3D convolutional network to obtain high-dimensional features, which include the spatial morphology and wall thickness distribution information of the left ventricular myocardium. Based on the high-dimensional features, classification processing is performed to generate cardiomyopathy classification results; The cardiomyopathy classification results include hypertrophic cardiomyopathy, dilated cardiomyopathy, restrictive cardiomyopathy, amyloidosis, myocarditis, and normal type; Based on the cardiomyopathy classification results, a preset optimization algorithm is applied to adjust the classification model parameters to obtain optimized cardiomyopathy classification model results, which are then output as cardiomyopathy classification model results.

8. The processing method based on three-dimensional cardiac images as described in claim 5, characterized in that, The steps of processing the dual-channel input using the pre-trained classification model and outputting the cardiomyopathy classification model result include: Based on the degree of deviation of each sampling point in the myocardial wall thickness mapping from the preset myocardial wall thickness reference value, the spatial weight of each sampling point is determined, and a spatial guidance mask matrix M aligned with the spatial scale of the three-dimensional image data of the heart is generated. The three-dimensional image data of the heart is input into the second input channel of the classification model, and a deep spatial feature tensor T containing anatomical texture information is extracted through multidimensional convolution. By introducing a dynamic adjustment factor β, the spatial guidance mask matrix M is used to perform weight modulation on the deep spatial feature tensor T on a per-spatial coordinate basis to enhance the feature intensity corresponding to the wall thickness anomaly region and generate a calibrated spatial interaction feature map F. The spatial dimension of the spatial interaction feature map F is compressed to extract the core pathological feature vector, and the core pathological feature vector is mapped to the category space to output the cardiomyopathy classification model result.

9. The processing method based on three-dimensional cardiac images as described in claim 4, characterized in that, A three-dimensional inner diameter surface is constructed based on the spatial distribution of the inner diameter pixels, including: Based on the coordinate information of the inner diameter pixels, a three-dimensional mesh structure is generated, which represents the inner diameter surface of the left ventricular myocardium; For the aforementioned three-dimensional mesh structure, a preset interpolation algorithm is applied to fill in the missing surface data; Based on the spatial distribution of the inner diameter pixels, the geometry of the three-dimensional mesh structure is adjusted to obtain a three-dimensional inner diameter surface that retains the original spatial shape; For the three-dimensional inner diameter surface, a corresponding color code is generated, which reflects the spatial position information of the surface.

10. The method for processing cardiac three-dimensional images as described in claim 6, characterized in that, Based on the spatial distribution of the inner diameter of the left ventricular myocardium, a three-dimensional spatial frame containing the inner diameter of the left ventricular myocardium is extracted, including: Based on the spatial distribution of pixels in the inner diameter, the cropping range in the x-axis, y-axis and z-axis directions is determined; For the aforementioned trimming range, a three-dimensional spatial frame containing the inner diameter is generated, the three-dimensional spatial frame covering the main distribution area of ​​the left ventricular myocardial inner diameter; Based on the three-dimensional spatial frame, the three-dimensional inner diameter surface and myocardial wall thickness mapping are clipped to obtain the clipped three-dimensional data.

11. A processing device based on three-dimensional cardiac images, characterized in that, include: The acquisition module is used to acquire three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers, wherein the three-dimensional segmentation results of the heart chambers include the left ventricular chambers and the left ventricular myocardium; The processing module is used to construct a three-dimensional inner diameter surface and generate a myocardial wall thickness mapping based on the inner and outer diameters of the left ventricular myocardium, wherein the myocardial wall thickness mapping represents the shortest distance from the inner diameter pixel to the outer diameter. The training module is used to determine the pre-trained classification model and construct the dual-channel input of the pre-trained classification model; The training module is also used to input the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness into the first input channel of the pre-trained classification model; The training module is also used to process the dual-channel input through the pre-trained classification model and output the cardiomyopathy classification model result; The pre-trained classification model's dual-channel input includes a first input channel relating the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness.

12. A non-volatile computer-readable storage medium storing computer-executable instructions thereon, characterized in that, When the computer-executable instructions are executed by the processor, the following steps are performed: Acquire three-dimensional image data of the heart and three-dimensional segmentation results of the heart chambers, wherein the three-dimensional segmentation results of the heart chambers include the left ventricular chambers and the left ventricular myocardium; Based on the inner and outer diameters of the left ventricular myocardium, a three-dimensional inner diameter surface is constructed and a myocardial wall thickness mapping is generated, wherein the myocardial wall thickness mapping represents the shortest distance from the inner diameter pixel to the outer diameter. Determine the pre-trained classification model and construct the dual-channel input of the pre-trained classification model; The mapping between the three-dimensional inner diameter surface and the myocardial wall thickness is input into the first input channel of the pre-trained classification model; The dual-channel input is processed by the pre-trained classification model to output the cardiomyopathy classification model result; The pre-trained classification model's dual-channel input includes a first input channel relating the mapping between the three-dimensional inner diameter surface and the myocardial wall thickness.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 10.

Images