Deep learning-based three-dimensional medical image classification method for myocardial ischemia

Abstract

The application relates to the technical field of medical image processing, in particular to a myocardial ischemia three-dimensional medical image classification method based on deep learning, which comprises the following steps: acquiring coronary angiography, myocardial perfusion and heart structure image sequences, performing multi-modal registration and cross-modal feature extraction, generating multi-dimensional heart feature volume data which fuses blood vessel, blood flow and tissue features, inputting the three-dimensional feature analysis network to extract functional damage spatial distribution and blood flow dynamics abnormal mode, constructing a risk quantification atlas containing ischemia degree distribution and range label, and outputting a myocardial ischemia image feature classification label through a classification decision engine. Multi-modal registration and fusion make different features correspond accurately in space, provide a consistent analysis carrier, and the spatial distribution and blood flow dynamics mode are jointly analyzed and graphically presented, so that the ischemia state is intuitively quantified and positioned. The application realizes high-precision fusion of multi-modal heart images and ischemia risk spatial visualization, and improves the fineness and interpretability of myocardial ischemia classification.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a deep learning-based method for classifying three-dimensional medical images of myocardial ischemia. Background Technology

[0002] In the clinical assessment of myocardial ischemia, existing methods largely rely on single-modality medical imaging, such as coronary angiography to observe vascular stenosis, myocardial perfusion imaging to detect blood flow defects, and cardiac structural imaging to examine myocardial morphology. These images differ in imaging principles, resolution, and acquisition time, and are often interpreted separately in clinical practice, with physicians relying on their experience to comprehensively determine the location and extent of ischemia. Faced with multi-source information, existing solutions lack a processing flow that can spatially unify data from different modalities, making it difficult to accurately correlate features from different sources and reducing the completeness of the overall understanding of the ischemic area.

[0003] Existing solutions for processing multimodal images typically extract features from each modality independently and then perform manual or simple algorithmic overlay of information. They lack rigorous multimodal registration and fail to create a unified three-dimensional feature representation that integrates vascular, blood flow, and tissue information. In ischemia assessment, conventional methods often directly deduce classification labels from image features, lacking joint mining of the spatial distribution of functional impairment and abnormal hemodynamic patterns. This fails to produce quantifiable maps that intuitively present the distribution and extent of ischemia, limiting the ability to meticulously characterize ischemic states and perform spatial reasoning. This invention aims to address the problem of inaccurate spatial correspondence and feature separation in multimodal images and construct a risk quantification representation that simultaneously reflects the distribution and extent of ischemia, thereby overcoming the limitations of existing methods in multimodal fusion accuracy and ischemia spatial visualization. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-based three-dimensional medical image classification method for myocardial ischemia.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-based three-dimensional medical image classification method for myocardial ischemia, comprising:

[0006] Acquire three-dimensional medical imaging data of the heart of the patient to be analyzed. The three-dimensional medical imaging data of the heart includes coronary angiography image sequence, myocardial perfusion image sequence and cardiac structure image sequence.

[0007] Multimodal registration and feature extraction are performed on the three-dimensional medical imaging scan data of the heart to generate multidimensional cardiac feature data that integrates vascular, blood flow and tissue features;

[0008] The multidimensional cardiac feature data is input into a pre-trained three-dimensional feature parsing network. The three-dimensional feature parsing network is used to extract the spatial distribution features and hemodynamic abnormality patterns of the cardiac function impairment area, including:

[0009] The three-dimensional feature parsing network consists of a spatial encoding path, a feature decoding path, and a multi-scale feature fusion module.

[0010] The spatial coding path progressively downsamples the multidimensional cardiac feature data, extracts the contextual features of the heart at multiple scales, and generates a multi-scale contextual feature map.

[0011] The feature decoding path progressively upsamples the multi-scale context feature map, restoring spatial resolution while fusing features of the same scale from the encoding path through skip connections to generate a high-resolution feature reconstruction map.

[0012] The multi-scale feature fusion module performs cross-scale aggregation on features from different levels of the spatial coding path and features from the feature decoding path to generate a comprehensive representation vector of cardiac function.

[0013] The regions representing local functional decline are parsed from the comprehensive representation vector to generate the spatial distribution features of the cardiac function damage regions. The spatial distribution features include the volume, spatial location, and topological relationship of the damage regions with the coronary artery supply regions.

[0014] The time delay, intensity reduction, and spatial mismatch patterns of hemorrhage perfusion are analyzed from the comprehensive characterization vector to generate the hemodynamic abnormality pattern, which includes the pattern and extent of perfusion defects and their correspondence with abnormal electrocardiogram leads.

[0015] Based on the spatial distribution characteristics and abnormal hemodynamic patterns, a quantitative map of myocardial ischemia risk is constructed, which includes the distribution of ischemia degree and the marking of ischemia range.

[0016] The myocardial ischemia risk quantification map is input into a multi-level classification decision engine, which calculates the ischemia type discrimination probability and outputs myocardial ischemia image feature classification labels.

[0017] As a further aspect of the present invention, the step of performing multimodal registration and feature extraction on the three-dimensional cardiac medical image scanning data to generate multidimensional cardiac feature data incorporating vascular, blood flow, and tissue features includes:

[0018] The coronary angiography image sequence is segmented to extract vessel wall boundaries, and data on changes in vessel centerline and lumen diameter are extracted to generate vessel morphology feature sequences.

[0019] The myocardial perfusion image sequence was subjected to temporal registration and signal intensity analysis. The time to peak perfusion, peak perfusion intensity and area under the perfusion curve of each myocardial segment were calculated to generate a blood flow perfusion dynamics feature map.

[0020] The cardiac structure image sequence is segmented into myocardial wall boundaries, and myocardial wall thickness, myocardial wall motion amplitude and myocardial strain rate data are extracted to generate a myocardial tissue structure feature field.

[0021] A unified spatial coordinate system based on cardiac anatomical landmarks is established, and the vascular morphological feature sequence, blood flow perfusion dynamic feature map and myocardial tissue structural feature field are spatially transformed and aligned to the unified spatial coordinate system.

[0022] The aligned vascular morphology feature sequence, blood flow perfusion dynamics feature map, and myocardial tissue structure feature field are spliced ​​at the voxel level to form multi-channel three-dimensional data containing vascular geometry, blood flow dynamics, and myocardial structure, namely the multi-dimensional cardiac feature data.

[0023] As a further aspect of the present invention, the construction of a quantitative map of myocardial ischemia risk based on the spatial distribution characteristics and abnormal hemodynamic patterns includes:

[0024] The spatial distribution characteristics of the cardiac function damage area are mapped back to the unified spatial coordinate system to generate a damage probability distribution field in the three-dimensional space of the heart.

[0025] The hemodynamic abnormality pattern is quantified into a perfusion abnormality score for each myocardial segment, which is calculated based on the degree of delay in the time to peak perfusion and the percentage decrease in peak perfusion intensity.

[0026] By fusing the damage probability distribution field with the blood perfusion abnormality score, a comprehensive ischemic risk value is calculated for each spatial voxel. The comprehensive ischemic risk value is a weighted sum of the damage probability and the blood perfusion abnormality score.

[0027] Using the comprehensive ischemic risk value as the filler value, a color-coded quantification map of myocardial ischemia risk is generated on a three-dimensional cardiac model.

[0028] In the myocardial ischemia risk quantification map, high-risk ischemic areas and medium-risk ischemic areas are automatically delineated and marked according to the threshold of the comprehensive ischemia risk value, thus completing the ischemia range marking.

[0029] As a further aspect of the present invention, the step of inputting the myocardial ischemia risk quantification map into a multi-level classification decision engine, calculating the ischemia type discrimination probability through the multi-level classification decision engine, and outputting myocardial ischemia image feature classification labels includes:

[0030] The multi-layer classification decision engine includes an ischemia pattern recognition layer, a clinical correlation analysis layer, and a decision fusion layer.

[0031] The ischemia pattern recognition layer receives the myocardial ischemia risk quantification map, uses a pre-trained convolutional neural network to extract the global morphological pattern and local texture pattern of the ischemic region, and outputs the preliminary ischemia pattern category probability.

[0032] The clinical association analysis layer receives patient clinical information data, which includes multi-dimensional parameters such as patient physiological, electrophysiological and biochemical indicators. The clinical information data is logically associated with the preliminary ischemia pattern output by the ischemia pattern recognition layer to calculate the clinical evidence support.

[0033] The decision fusion layer integrates the preliminary ischemia pattern category probability with the clinical evidence support, and calculates the final discrimination probability of each ischemia type through weighted decision rules;

[0034] The ischemia type with the highest probability is selected from the final discrimination probabilities, and combined with the preset classification rules, a myocardial ischemia image feature classification label containing the specific ischemia type, ischemia range, and ischemia degree is generated.

[0035] As a further aspect of the present invention, the step of performing temporal registration and signal intensity analysis on the myocardial perfusion image sequence, calculating the peak perfusion time, peak perfusion intensity, and area under the perfusion curve for each myocardial segment, and generating a hemoperfusion dynamics feature map includes:

[0036] Using the cardiac cycle as a reference, each frame of the myocardial perfusion image sequence is marked with a time point to establish the correspondence between image frames and cardiac phases;

[0037] Based on the aforementioned correspondence, the mutual information maximization criterion is used to perform temporal registration of image sequences of adjacent cardiac cycles, so that myocardial images of the same phase in different cycles are spatially aligned.

[0038] On the image sequence with completed phase registration, the left ventricular myocardium is divided into multiple segments according to the standard myocardial segmentation model;

[0039] For each myocardial segment, the average signal intensity value of each myocardial segment in all temporal images is extracted to form the time-signal intensity curve of the myocardial segment;

[0040] Identify the maximum signal strength from the time-signal strength curve, record the time phase corresponding to the maximum signal strength as the perfusion peak time, and record the maximum signal strength as the perfusion peak intensity;

[0041] The time-signal intensity curve is numerically integrated from the start point to the end point of the curve, and the resulting integral value is the area under the perfusion curve.

[0042] The peak perfusion time, peak perfusion intensity, and area under the perfusion curve of all myocardial segments are arranged according to the spatial location of the segments to generate a two-dimensional blood flow perfusion dynamics feature map, in which the three parameter values ​​of each segment are stored in different channels of the feature map.

[0043] As a further aspect of the present invention, the spatial coding path progressively downsamples the multidimensional cardiac feature data, extracts cardiac contextual features at multiple scales, and generates a multi-scale contextual feature map, including:

[0044] The multidimensional cardiac feature data is used as input and sequentially passed through an encoding module consisting of multiple concatenated convolution and pooling operation layers;

[0045] In each encoding module, a 3D convolution operation is first performed to extract local features, followed by a 3D max pooling operation to reduce the spatial resolution of the feature map and expand the receptive field.

[0046] After processing by a series of encoding modules, a series of feature maps from high resolution to low resolution are obtained, namely the multi-scale context feature maps. The low-resolution feature maps contain richer global context information, while the high-resolution feature maps retain more spatial detail information.

[0047] During the encoding process, the feature map before each pooling operation is recorded for subsequent skip connections with the feature decoding path.

[0048] As a further aspect of the present invention, the step of fusing the damage probability distribution field with the blood perfusion abnormality score to calculate the comprehensive ischemic risk value for each spatial voxel includes:

[0049] The damage probability distribution field is normalized to map the damage probability value of each voxel to a standard interval between zero and one.

[0050] The abnormal blood perfusion score was normalized, and the abnormal blood perfusion score of each myocardial segment was also mapped to a standard interval between zero and one.

[0051] Based on the pre-configured damage probability weight and blood perfusion abnormality weight, the sum of the damage probability weight and blood perfusion abnormality weight is one.

[0052] For each spatial voxel in the unified spatial coordinate system, find its corresponding myocardial segment and obtain the normalized blood perfusion abnormality score of the myocardial segment;

[0053] The sum of the normalized damage probability value of the spatial voxel multiplied by the damage probability weight, plus the normalized perfusion abnormality score of its corresponding myocardial segment multiplied by the perfusion abnormality weight, is the comprehensive ischemic risk value of the spatial voxel.

[0054] As a further aspect of the present invention, the clinical association analysis layer receives patient clinical information data, which includes multi-dimensional parameters containing patient physiological, electrophysiological, and biochemical indicators. The clinical information data is then logically associated with the preliminary ischemia pattern output by the ischemia pattern recognition layer to calculate the clinical evidence support level, including:

[0055] The patient's clinical information data is structured and encoded into a fixed-dimensional clinical feature vector.

[0056] Construct an associated knowledge base containing the correspondence between clinical feature vectors and ischemia pattern categories;

[0057] Calculate the similarity between the current patient's clinical feature vector and the clinical feature vectors of each historical case in the associated knowledge base, filter out a set of similar cases based on the similarity, and generate a clinical prior probability distribution based on the statistical distribution of different ischemic patterns in the set.

[0058] The initial ischemia pattern category probability output by the ischemia pattern recognition layer is combined with the clinical prior probability distribution using Bayesian fusion calculation to update the posterior probability distribution.

[0059] Based on the posterior probability distribution, the strength of support of the current patient's clinical information for each ischemic pattern is calculated, which is the clinical evidence support level. Its value reflects the degree of consistency between clinical information and radiographic ischemic patterns.

[0060] As a further aspect of the present invention, the step of performing Bayesian fusion calculation on the preliminary ischemia pattern category probability output by the ischemia pattern recognition layer and the clinical prior probability distribution to update the posterior probability distribution includes:

[0061] The initial ischemia pattern category probability output by the ischemia pattern recognition layer is used as the likelihood probability;

[0062] The clinical prior probability distribution is used as the prior probability;

[0063] According to Bayes' theorem, the posterior probability numerator of each ischemia mode category is calculated, and the posterior probability numerator is equal to the likelihood probability of the ischemia mode category multiplied by the prior probability of the ischemia mode category.

[0064] Summing the posterior probability numerators for all ischemic pattern categories yields the normalized denominator;

[0065] Divide the posterior probability numerator of each ischemia mode category by the normalized denominator to obtain the normalized posterior probability of the ischemia mode category.

[0066] The normalized posterior probabilities of all ischemic mode categories are combined into a new probability distribution, which is the posterior probability distribution.

[0067] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0068] In cases where cardiac 3D medical imaging data includes coronary angiography sequences, myocardial perfusion sequences, and cardiac structural sequences, multimodal registration is performed to precisely align data from different imaging sources in space. Subsequently, vascular morphology features, blood flow perfusion features, and tissue structural features are extracted and fused into multidimensional cardiac feature volume data in 3D space. This cross-modal synchronous registration and feature fusion ensures that the correlation information between blood vessels, blood flow, and tissues remains consistent within the same spatial framework. It allows for the simultaneous acquisition of descriptions of multiple attributes at the same voxel location, avoiding feature mismatches caused by positional deviations between different modalities. This provides a consistent spatial basis for subsequent comprehensive analysis of cardiac structure and function, enabling direct observation of the coupling relationships between different attributes in the spatial dimension and improving the completeness of localization and identification of ischemia-related changes.

[0069] Multidimensional cardiac feature data is input into a pre-trained 3D feature parsing network to extract the spatial distribution features and hemodynamic abnormality patterns of cardiac functional impairment areas. Based on this, a quantitative atlas of myocardial ischemia risk is constructed, which includes the distribution of ischemia severity and ischemia extent markings. This joint extraction of the spatial distribution of functional impairment and hemodynamic patterns can capture the specific location and morphology of ischemic areas in the myocardium, as well as the dynamic patterns of perfusion abnormalities. The atlas generated by combining these two aspects presents the degree of ischemia in a spatially continuous numerical field, and the boundaries of the affected area are clearly defined through extent markings. This representation differs from simply providing category labels; it directly presents the changes in ischemic intensity and coverage area in space, enabling the observation and inference of ischemic states to rely on an intuitive and computable graphical form, enhancing the interpretability and usability of ischemic spatial features. Attached Figure Description

[0070] Figure 1 This is a flowchart of the deep learning-based three-dimensional medical image classification method for myocardial ischemia described in this invention;

[0071] Figure 2 A flowchart illustrating the operation of a 3D feature parsing network;

[0072] Figure 3 A heatmap showing the distribution of comprehensive ischemic risk values ​​in a two-dimensional cross-section of the heart;

[0073] Figure 4A scatter plot showing the correlation between time to peak myocardial perfusion and area under the curve.

[0074] Figure 5 The resolution variation curve of the encoding-decoding path in the 3D feature parsing network. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0076] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0077] See Figure 1 The method involves acquiring three-dimensional cardiac medical image scans of the patient to be analyzed, including coronary angiography sequences, myocardial perfusion sequences, and cardiac structural sequences. Multimodal registration and feature extraction are performed on the three-dimensional cardiac medical image scans to generate multidimensional cardiac feature data that integrates vascular, blood flow, and tissue features. This multidimensional cardiac feature data is then input into a pre-trained three-dimensional feature parsing network to extract spatial distribution features and hemodynamic abnormality patterns of cardiac function impairment areas. Based on these spatial distribution features and hemodynamic abnormality patterns, a myocardial ischemia risk quantification map is constructed, including ischemia degree distribution and ischemia range markings. Finally, the myocardial ischemia risk quantification map is input into a multi-level classification decision engine to calculate the ischemia type discrimination probability and output myocardial ischemia image feature classification labels.

[0078] In one embodiment of the present invention, the coronary angiography image sequence is processed by segmenting the vessel wall boundaries, extracting data on the vessel centerline and changes in vessel lumen diameter, and generating a vascular morphology feature sequence. Temporal registration and signal intensity analysis are performed on the myocardial perfusion image sequence. Using the cardiac cycle as a reference, each frame in the myocardial perfusion image sequence is time-marked to establish a correspondence between image frames and cardiac phases. Based on this correspondence, the mutual information maximization criterion is used to perform temporal registration on image sequences of adjacent cardiac cycles, ensuring spatial alignment of myocardial images of the same phase within different cycles. On the temporally registered image sequence, the left ventricular myocardium is divided into multiple segments according to a standard myocardial segmentation model. For each myocardial segment, the average signal intensity of each segment in all temporal images is extracted. The time-signal intensity curves of myocardial segments are generated using the intensity values. The maximum signal intensity is identified from these curves, and the time corresponding to this maximum is recorded as the time to peak perfusion and the peak perfusion intensity. The time-signal intensity curve is numerically integrated from its starting point to its ending point, and the resulting integral is the area under the perfusion curve. The time to peak perfusion, peak perfusion intensity, and area under the perfusion curve for all myocardial segments are arranged according to their spatial location to generate a two-dimensional blood flow perfusion dynamics feature map. The three parameter values ​​for each segment are stored in different channels of the feature map. Myocardial wall boundary segmentation is performed on the cardiac structural image sequence, and myocardial wall thickness, myocardial wall motion amplitude, and myocardial strain rate data are extracted to generate a myocardial tissue structural feature field. A unified spatial coordinate system based on cardiac anatomical landmarks is established, and the vascular morphology feature sequence, blood flow perfusion dynamics feature map, and myocardial tissue structural feature field are spatially transformed and aligned to this unified spatial coordinate system. The aligned vascular morphology feature sequences, blood flow perfusion dynamics feature maps, and myocardial tissue structure feature fields are spliced ​​at the voxel level to form multi-channel three-dimensional data containing vascular geometry, blood flow dynamics, and myocardial structure, i.e., multi-dimensional cardiac feature data.

[0079] In a specific implementation, multimodal registration and feature extraction are performed on the three-dimensional cardiac medical imaging scan data to generate multidimensional cardiac feature data that integrates vascular, blood flow, and tissue features. The specific implementation process can be described using an example scenario. In this scenario, the three-dimensional cardiac medical imaging scan data of the patient to be analyzed includes a set of coronary angiography image sequences, a set of myocardial perfusion image sequences, and a set of cardiac structural image sequences. These sequences originate from medical imaging examinations of the same patient at similar time points. The coronary angiography image sequences undergo vessel wall boundary segmentation, extracting data on vessel centerline and lumen diameter changes to generate vascular morphology feature sequences. The myocardial perfusion image sequences undergo temporal registration and signal intensity analysis. The cardiac structural image sequences undergo myocardial wall boundary segmentation, extracting myocardial wall thickness, myocardial wall motion amplitude, and myocardial strain rate data to generate a myocardial tissue structural feature field. In some embodiments, the temporal registration and signal intensity analysis process of myocardial perfusion image sequences is as follows: Using the cardiac cycle as a reference, each frame of the myocardial perfusion image sequence is time-marked, for example, the peak R-wave time is marked as the temporal zero point, establishing a correspondence between each frame and the cardiac phase. Based on the correspondence between image frames and cardiac phases, the mutual information maximization criterion is used to perform temporal registration on image sequences of adjacent cardiac cycles, ensuring that images of the left ventricular myocardium in phases within different cycles, such as at end-systole, are spatially perfectly aligned. On the image sequence with completed temporal registration, the left ventricular myocardium is divided into seventeen segments according to the standard seventeen-segment division model. For each myocardial segment, the average signal intensity value of each myocardial segment in all temporal images is extracted, forming an independent time-signal intensity curve for each myocardial segment. The maximum signal intensity is identified from the time-signal intensity curve, and the phase corresponding to the maximum signal intensity is recorded as the perfusion peak time, and the maximum signal intensity is recorded as the peak perfusion intensity. Numerical integration is performed on the time-signal strength curve from the start point to the end point, and the integral value is calculated as the area under the injection curve. The formula for calculating the area under the injection curve can be understood as follows:

[0080]

[0081] in: Indicates the area under the perfusion curve. This indicates the time phase corresponding to the starting point of the curve. This indicates the time phase corresponding to the end point of the curve. Indicates the phase time The average signal intensity value of the myocardial segment is calculated. The time to peak perfusion, peak perfusion intensity, and area under the perfusion curve of all myocardial segments are arranged according to the standard spatial positions of the seventeen segments to generate a two-dimensional blood flow perfusion dynamic characteristic map of seventeen rows and three columns. Each row represents a myocardial segment, and each column stores the three parameters: time to peak perfusion, peak perfusion intensity, and area under the perfusion curve.

[0082] A unified spatial coordinate system is established based on cardiac anatomical landmarks, such as the apex of the left ventricle and the midpoint of the mitral valve annulus, to create a three-dimensional coordinate system. Vascular morphology feature sequences, perfusion dynamics feature maps, and myocardial tissue structure feature fields are spatially transformed and aligned to the unified spatial coordinate system. Optionally, the spatial transformation process uses an affine transformation method based on anatomical landmarks. The aligned vascular morphology feature sequences, perfusion dynamics feature maps, and myocardial tissue structure feature fields are then stitched together at the voxel level using feature channels. For example, the vascular morphology feature sequences provide three feature channels, the perfusion dynamics feature maps provide three feature channels after spatial interpolation, and the myocardial tissue structure feature fields provide three feature channels. Finally, these are stitched together to form a nine-channel three-dimensional data set containing vascular geometry, blood flow dynamics, and myocardial structure—i.e., multidimensional cardiac feature volume data.

[0083] In one embodiment of the present invention, see [reference] Figure 2The 3D feature parsing network consists of a spatial encoding path, a feature decoding path, and a multi-scale feature fusion module. The spatial encoding path progressively downsamples the multi-dimensional cardiac feature volume data. This process involves taking the multi-dimensional cardiac feature volume data as input and sequentially passing it through a series of cascaded encoding modules consisting of convolutional and pooling operations. Within each encoding module, a 3D convolution operation is first performed to extract local features, followed by a 3D max pooling operation to reduce the spatial resolution of the feature map and expand the receptive field. This series of encoding modules results in a series of feature maps from high to low resolution, i.e., multi-scale context feature maps. The low-resolution feature maps contain richer global context information, while the high-resolution feature maps retain more spatial detail. During the encoding process, the feature map before each pooling operation is recorded for subsequent skip connections with the feature decoding path. The feature decoding path progressively upsamples the multi-scale context feature maps, restoring the spatial resolution while fusing features of the same scale from the encoding path through skip connections to generate a high-resolution feature reconstruction map. The multi-scale feature fusion module aggregates features from different levels of the spatial encoding path and the feature decoding path across scales to generate a comprehensive representation vector of cardiac function. The regions representing localized functional impairment are extracted from the comprehensive representation vector, generating spatial distribution characteristics of the cardiac function impairment region. These characteristics include the volume, spatial location, and topological relationship of the damaged region with the coronary artery supply area. Furthermore, the temporal delay, intensity reduction, and spatial mismatch patterns of hemorrhage perfusion are extracted from the comprehensive representation vector, generating hemodynamic abnormality patterns. These patterns include the pattern and extent of perfusion defects and their correspondence with abnormal electrocardiogram leads.

[0084] In a specific implementation, multidimensional cardiac feature volume data is input into a pre-trained three-dimensional feature parsing network, which then extracts the spatial distribution features and hemodynamic abnormality patterns of the cardiac function impairment area. The specific implementation process can be described using an example scenario. In this scenario, the input multidimensional cardiac feature volume data to the three-dimensional feature parsing network is a three-dimensional data volume with nine feature channels and a spatial size of 128x128x64, generated by the steps described in the embodiment. The three-dimensional feature parsing network consists of a spatial encoding path, a feature decoding path, and a multi-scale feature fusion module. In the specific implementation, the spatial encoding path progressively downsamples the multidimensional cardiac feature volume data. This process involves taking the multidimensional cardiac feature volume data as input and sequentially passing it through an encoding module composed of multiple concatenated layers of convolution and pooling operations. In some embodiments, the spatial encoding path comprises three downsampling stages. In each encoding module, a 3D convolution operation is first performed to extract local features, for example, using a 3x3x3 convolution kernel and the ReLU activation function. Then, a 3D max pooling operation is performed to reduce the spatial resolution of the feature map and expand the receptive field, for example, using a 2x2x2 pooling window. After processing by a series of encoding modules, a series of feature maps from high resolution to low resolution are obtained, i.e., multi-scale context feature maps. The low-resolution feature maps contain richer global context information, while the high-resolution feature maps retain more spatial detail information. It can be understood that after the three downsampling stages, the spatial size of the feature map may become 64x64x32, 32x32x16, and 16x16x8, respectively. During the encoding process, the feature map before each pooling operation is recorded, for example, feature maps with sizes of 128x128x64, 64x64x32, and 32x32x16, for subsequent skip connections with the feature decoding path.

[0085] The feature decoding path progressively upsamples the multi-scale context feature maps, restoring spatial resolution while fusing features from the same scale from the encoding path via skip connections to generate a high-resolution feature reconstruction map. In specific implementations, the feature decoding path includes three upsampling stages corresponding to the encoding path. Each stage upsamples through a 3D transposed convolution operation, and then concatenates the upsampled feature map with the corresponding scale feature map from the spatial encoding path, followed by a convolution operation to fuse information. It can be understood that the spatial resolution of the final generated feature reconstruction map is restored to the same level as the input multi-dimensional cardiac feature volume data. In some embodiments, the multi-scale feature fusion module performs cross-scale aggregation on features from different levels of the spatial encoding path and features from the feature decoding path to generate a comprehensive representation vector of cardiac function. The multi-scale feature fusion module receives feature maps from three different levels of the spatial encoding path and the feature map output from the feature decoding path. It performs global average pooling on these feature maps, compressing each 3D feature map into a 1D feature vector, and then concatenates the four 1D feature vectors to form a comprehensive feature vector, i.e., the comprehensive representation vector. The generation process of the comprehensive representation vector can be represented as follows:

[0086]

[0087] in: Represents the comprehensive representation vector. Indicates a connection operation. , , These represent feature vectors from the three levels of the spatial encoding path, respectively. This represents the feature vector from the final output of the feature decoding path.

[0088] The comprehensive representation vector is used to extract regions representing local functional decline, generating spatial distribution features of cardiac function impairment areas. In specific implementation, the comprehensive representation vector is mapped to a low-dimensional latent spatial code through a fully connected layer. This code is then decoded by a three-dimensional deconvolutional network into a single-channel probability map with the same spatial dimensions as the original heart. This probability map represents the probability distribution of injury within the three-dimensional space of the heart. Based on this probability map, thresholding processing can extract the volume, spatial location, and topological relationship of the injured region with the coronary artery supply area, constituting the spatial distribution features. Optionally, the topological relationship information of the coronary artery supply area is derived from a pre-defined atlas of coronary vessels and myocardial segments. The time delay, intensity reduction, and spatial mismatch patterns of hemorrhage perfusion are extracted from the comprehensive representation vector to generate hemodynamic abnormality patterns. Understandably, the integrated representation vector, through another parallel fully connected layer and decoder, generates a three-dimensional parameter map representing hemodynamic abnormalities. This parameter map includes parameters such as the peak perfusion time offset and peak intensity attenuation coefficient at each voxel location. By analyzing the spatial distribution patterns of these parameters, the pattern and extent of perfusion defects and their correspondence with abnormal ECG leads can be identified. Optionally, the correspondence with abnormal ECG leads is determined by projecting the three-dimensional spatial model of the heart onto the standard ECG lead coordinate system.

[0089] In one embodiment of the present invention, the spatial distribution characteristics of the cardiac function impairment area are mapped back to a unified spatial coordinate system to generate an impairment probability distribution field in the three-dimensional space of the heart. The hemodynamic abnormality pattern is quantified into a perfusion abnormality score for each myocardial segment, which is calculated based on the degree of delay in the time to peak perfusion and the percentage decrease in peak perfusion intensity. By fusing the injury probability distribution field and the blood perfusion anomaly score, the comprehensive ischemic risk value of each spatial voxel is calculated. The fusion process includes normalizing the injury probability distribution field, mapping the injury probability value of each voxel to a standard interval between zero and one, and normalizing the blood perfusion anomaly score, mapping the blood perfusion anomaly score of each myocardial segment to a standard interval between zero and one. Based on pre-configured injury probability weights and blood perfusion anomaly weights (the sum of the injury probability weights and blood perfusion anomaly weights is one), for each spatial voxel in a unified spatial coordinate system, its corresponding myocardial segment is located, and the normalized blood perfusion anomaly score of the myocardial segment is obtained. The normalized injury probability value of the spatial voxel is multiplied by the injury probability weight, and then added to the normalized blood perfusion anomaly score of its corresponding myocardial segment multiplied by the blood perfusion anomaly weight. The sum is the comprehensive ischemic risk value of the spatial voxel. Using the comprehensive ischemic risk value as the filler value, a color-coded quantitative atlas of myocardial ischemia risk is generated on a 3D cardiac model. In the myocardial ischemia risk quantification map, high-risk and medium-risk ischemic areas are automatically delineated and marked based on the threshold of the comprehensive ischemia risk value, thus completing the ischemia range marking.

[0090] See Figure 3 This figure shows a heatmap of the comprehensive ischemic risk value distribution of a two-dimensional cross-section of the heart, generated based on the myocardial ischemia risk quantification algorithm in this embodiment. Specifically, the X-axis and Y-axis in the figure represent the two-dimensional cross-sectional coordinates of the heart space, respectively. The color coding maps the comprehensive ischemic risk value, where dark blue corresponds to a risk value of 0.0 and dark red corresponds to a risk value of 1.0. The color gradient reflects the continuous change of the risk value. This heatmap relies on the fusion calculation of the damage probability distribution field and the blood perfusion abnormality score: First, the spatial distribution characteristics of the cardiac function damage area are mapped to a unified spatial coordinate system, and the damage probability distribution field is normalized to the [0,1] interval; at the same time, the hemodynamic abnormality pattern is quantified into the blood perfusion abnormality score of the myocardial segment, which is also normalized to the [0,1] interval. Combining the pre-configured damage probability weight and blood perfusion abnormality weight, for each spatial voxel, its normalized damage probability value is weighted and summed with the normalized blood perfusion abnormality score of the corresponding myocardial segment to obtain the comprehensive ischemic risk value of that voxel.

[0091] The spatial distribution characteristics of the cardiac functional impairment area and the hemodynamic abnormality pattern are presented. The spatial distribution characteristics exist as a single-channel three-dimensional injury probability distribution field with a spatial size of 128x128x64. The hemodynamic abnormality pattern exists as a parameter list of seventeen myocardial segments, including the proportion of time delay to peak perfusion and the proportion of decrease in peak perfusion intensity for each segment. The spatial distribution characteristics of the cardiac functional impairment area are mapped back to a unified spatial coordinate system to generate an injury probability distribution field in the three-dimensional space of the heart. In specific implementations, the spatial distribution characteristics of the cardiac functional impairment area are already a three-dimensional data volume in a unified spatial coordinate system, which can be directly used as an injury probability distribution field. Each voxel value represents the probability of functional impairment at that spatial location, with probability values ​​ranging from zero to one. The hemodynamic abnormality pattern is quantified into a blood perfusion abnormality score for each myocardial segment. The blood perfusion abnormality score is calculated based on the degree of time delay to peak perfusion and the proportion of decrease in peak perfusion intensity. In some embodiments, for each myocardial segment, the blood perfusion abnormality score is calculated using a linear weighted formula:

[0092]

[0093] in: Indicates the first Blood perfusion abnormality score of each myocardial segment, Indicates the first The degree of delay in peak perfusion time in each myocardial segment Indicates the first The percentage decrease in peak perfusion intensity in each myocardial segment and These are preset weighting coefficients, and they satisfy... To balance the contributions of time and intensity factors to the anomaly.

[0094] The damage probability distribution field and the blood perfusion anomaly score are fused to calculate the comprehensive ischemic risk value for each spatial voxel. This fusion process involves normalizing the damage probability distribution field, mapping the damage probability value of each voxel to a standard interval between zero and one. Similarly, the blood perfusion anomaly score is normalized, mapping the blood perfusion anomaly score for each myocardial segment to a standard interval between zero and one. Based on pre-configured damage probability weights and blood perfusion anomaly weights, the sum of these weights is one. For each spatial voxel in the unified spatial coordinate system, its corresponding myocardial segment is located, and the normalized blood perfusion anomaly score for that segment is obtained. In practice, a spatial lookup table is pre-defined, defining which of the seventeen standard myocardial segments each voxel coordinate belongs to in the unified spatial coordinate system. The comprehensive ischemic risk value of the spatial voxel is obtained by multiplying its normalized damage probability value by the damage probability weight, and then adding the normalized perfusion abnormality score of its corresponding myocardial segment multiplied by the perfusion abnormality weight. Optionally, the damage probability weight can be set to 0.6, and the perfusion abnormality weight can be set to 0.4 accordingly.

[0095] Using a comprehensive ischemia risk value as filler, a color-coded quantification map of myocardial ischemia risk is generated on a 3D cardiac model. In some embodiments, a continuous color map from blue (low risk) to red (high risk) is used to map the comprehensive ischemia risk value of each voxel to a corresponding RGB color value. Then, the voxel data with color information is overlaid onto a 3D cardiac surface model reconstructed from cardiac structural image sequences for volume rendering, thereby generating an intuitive color 3D quantification map of myocardial ischemia risk. In the quantification map of myocardial ischemia risk, high-risk and medium-risk ischemic regions are automatically delineated and marked according to a threshold of the comprehensive ischemia risk value, completing the ischemia range marking. It can be understood that two risk thresholds are set; for example, voxel regions with a comprehensive ischemia risk value greater than 0.7 are marked as high-risk ischemic regions and delineated with a red outline, while voxel regions with a comprehensive ischemia risk value between 0.4 and 0.7 are marked as medium-risk ischemic regions and delineated with a yellow outline.

[0096] In one embodiment of the present invention, the multi-layer classification decision engine includes an ischemia pattern recognition layer, a clinical association analysis layer, and a decision fusion layer. The ischemia pattern recognition layer receives a quantified myocardial ischemia risk map and uses a pre-trained convolutional neural network to extract global morphological patterns and local texture patterns of the ischemic region, outputting preliminary ischemia pattern category probabilities. The clinical association analysis layer receives patient clinical information data, which is a pre-encoded structured feature vector containing the patient's physiological and electrophysiological indicators. It performs logical association analysis between the clinical information data and the preliminary ischemia patterns output by the ischemia pattern recognition layer to calculate the clinical evidence support. The decision fusion layer integrates the preliminary ischemia pattern category probabilities with the clinical evidence support, calculating the final discrimination probability of each ischemia type using weighted decision rules. The ischemia type with the highest probability is selected from the final discrimination probabilities and, combined with preset classification rules, generates a myocardial ischemia image feature classification label containing the specific ischemia category, ischemia range, and ischemia degree.

[0097] See Figure 4 This figure is a scatter plot of the correlation between perfusion kinetic parameters obtained after myocardial perfusion image sequence analysis, used to quantify the blood perfusion characteristics of different myocardial segments. Specifically, the horizontal axis represents the time to peak myocardial perfusion, indicating the moment when the time-signal intensity curve of the myocardial segment reaches its maximum signal intensity, reflecting the degree of delay in blood perfusion; the vertical axis represents the area under the perfusion curve, representing the integral value of the time-signal intensity curve, reflecting the total perfusion volume of that segment; the scatter plot color uses a continuous color scale to encode the peak perfusion intensity, with the color transitioning from dark purple to bright yellow, visually presenting the differences in peak perfusion intensity between different segments.

[0098] In a specific implementation, the myocardial ischemia risk quantification map is input into a multi-layer classification decision engine. The engine calculates the ischemia type discrimination probability and outputs a myocardial ischemia image feature classification label. The specific implementation process can be described using an example scenario. In this scenario, the input myocardial ischemia risk quantification map is a rendered, color-coded 3D image data. Simultaneously, the multi-layer classification decision engine receives patient clinical information data including the patient's age of 65 years, symptom description of "chest pain after activity," ECG findings of "ST segment depression of 0.1 mV in leads II, III, and aVF," and blood biochemical indicator "troponin I 0.03 ng / mL." The multi-layer classification decision engine comprises an ischemia pattern recognition layer, a clinical correlation analysis layer, and a decision fusion layer. In the specific implementation, the ischemia pattern recognition layer receives the myocardial ischemia risk quantification map and uses a pre-trained convolutional neural network to extract the global morphological pattern and local texture pattern of the ischemic region. The pre-trained convolutional neural network is, for example, a 3D ResNet network trained on a large amount of labeled cardiac image data. Its input is a single-channel grayscale version or a three-channel RGB version of a myocardial ischemia risk quantification map. The network output layer is a softmax classifier, which outputs preliminary probability vectors corresponding to four categories: "subendocardial ischemia," "transmural ischemia," "microvascular ischemia," and "no ischemia." In some embodiments, for a specific input, the preliminary ischemia pattern category probabilities output by the ischemia pattern recognition layer may be presented in the form of a numerical list.

[0099] The clinical association analysis layer receives patient clinical information data, which is a pre-coded structured feature vector containing the patient's physiological and electrophysiological indicators. It performs logical association analysis between the clinical information data and the preliminary ischemia patterns output by the ischemia pattern recognition layer to calculate the clinical evidence support level. Essentially, the clinical association analysis layer incorporates a rule base or a trained machine learning model to evaluate the degree of support for different ischemia patterns from specific combinations of clinical information. For example, the combination of the symptom "chest pain after activity" and the ECG "ST segment changes in inferior leads" increases the support for the "transmural ischemia" pattern, while normal troponin levels may decrease the support for the acute ischemia pattern. The output of the clinical evidence support level is also a numerical vector corresponding to the ischemia pattern category.

[0100] The decision fusion layer integrates the initial ischemia pattern category probabilities with the level of clinical evidence support, and calculates the final discrimination probability for each ischemia type using weighted decision rules. The decision fusion process follows the formula:

[0101]

[0102] in: Indicates the type of ischemia The final discrimination probability, This indicates the type of ischemia output by the ischemia pattern recognition layer. The initial probability, This indicates the type of ischemia calculated by the clinical association analysis layer. The degree of support from clinical evidence, and These are the fusion weights assigned to the imaging patterns and clinical evidence. In some embodiments, see Table 1 for an example of the association analysis between clinical information and ischemic patterns.

[0103] Table 1: Correlation between Clinical Information Characteristics and Ischemia Patterns

[0104]

[0105] The ischemia type with the highest probability is selected from the final discrimination probabilities. Combined with preset classification rules, a myocardial ischemia image feature classification label is generated, containing the specific ischemia type, ischemia range, and ischemia degree. In specific implementation, the decision fusion layer calculates the final discrimination probability vector as [0.15, 0.70, 0.10, 0.05], corresponding to "subendocardial ischemia," "transmural ischemia," "microvascular ischemia," and "no ischemia," respectively. The ischemia type with the highest probability is "transmural ischemia." Optionally, the preset classification rules may stipulate that when the probability of "transmural ischemia" exceeds 0.5 and the high-risk ischemic area is greater than 10% of the left ventricular myocardium, the corresponding ischemia type is output as the image feature classification label. The generated classification labels will be integrated into a structured data object, containing, for example, feature information such as "Image feature classification: transmural myocardial ischemia (high probability)," "Ischemia range: inferior wall of the left ventricle, accounting for approximately 15% of the left ventricular myocardial volume," and "Ischemia degree: high-risk area accounts for 70%." It is understandable that the specific data on the extent and degree of ischemia in the data object comes from the markers and quantification information in the myocardial ischemia risk quantification map generated by the implementation example.

[0106] In one embodiment of the present invention, the clinical association analysis layer receives patient clinical information data, which is a pre-encoded structured feature vector containing the patient's physiological and electrophysiological indicators. The patient clinical information data is structured and encoded into fixed-dimensional clinical feature vectors. An association knowledge base containing the correspondence between clinical feature vectors and ischemia pattern categories is constructed. The similarity between the current patient's clinical feature vector and the clinical feature vectors of each historical case in the association knowledge base is calculated. A set of similar cases is selected based on the similarity, and a clinical prior probability distribution is generated based on the statistical distribution of different ischemia pattern categories in this set. The preliminary ischemia pattern category probability output by the ischemia pattern recognition layer is compared with the clinical prior probability distribution. Bayesian fusion computation is used to update the posterior probability distribution. The process involves using the initial ischemia pattern category probability output by the ischemia pattern recognition layer as the likelihood probability and the clinical prior probability distribution as the prior probability. According to Bayes' theorem, the posterior probability numerator for each ischemia pattern category is calculated. The posterior probability numerator equals the likelihood probability of the ischemia pattern category multiplied by its prior probability. The posterior probability numerators of all ischemia pattern categories are summed to obtain a normalized denominator. The posterior probability numerator of each ischemia pattern category is then divided by the normalized denominator to obtain the normalized posterior probability of that category. These normalized posterior probabilities are combined to form a new probability distribution, which is the posterior probability distribution. Based on this posterior probability distribution, the strength of support for each ischemia pattern from the current patient's clinical information is calculated; this is the clinical evidence support level, which reflects the consistency between clinical information and the imaging ischemia pattern.

[0107] See Figure 5 This figure shows the quantitative results of the resolution changes of the spatial encoding path and feature decoding path in a 3D feature parsing network with the downsampling scale, intuitively reflecting the network's multi-scale feature extraction and reconstruction mechanism. Specifically, the horizontal axis represents the downsampling scale, indicating the cumulative factor of pooling operations in the encoding path; the vertical axis represents the feature map resolution, reflecting the spatial detail preservation capability of the feature map. The blue solid line is the encoding path resolution curve, which decreases exponentially with the downsampling scale: the resolution is approximately 510 pixels at the 1× scale, decreasing to approximately 30 pixels at the 16× scale. This change stems from the encoding path gradually reducing the spatial resolution through progressive 3D convolution and max pooling operations to expand the receptive field, thereby extracting richer global contextual features. The red dashed line is the decoding path resolution curve, which increases exponentially with the downsampling scale: the resolution is approximately 30 pixels at the 1× scale, rising back to approximately 510 pixels at the 16× scale. This change stems from the decoding path restoring the spatial resolution through progressive upsampling operations and fusing features of the same scale from the encoding path through skip connections to preserve local spatial details.

[0108] The clinical association analysis layer receives patient clinical information data, which is a pre-coded structured feature vector containing the patient's physiological and electrophysiological indicators. This data is structured and encoded into a fixed-dimensional clinical feature vector. In practice, each piece of clinical information is mapped to a numerical value or vector. For example, the patient's age of 62 years is directly used as a scalar feature; the symptom "exertional chest pain" is mapped to a feature code [1,0,0] according to a pre-defined dictionary; the ECG manifestation of "T-wave inversion in leads V4-V6" is mapped to a feature code [0,1,0,1,1,1] to represent anterior and lateral wall abnormalities; and "low-density lipoprotein 3.8 mmol / L" is standardized and mapped to a scalar feature. All these numerical features are concatenated to form a fixed-length clinical feature vector. A knowledge base is constructed that includes the correspondence between clinical feature vectors and ischemia pattern categories. This knowledge base stores a large number of historical confirmed cases' clinical feature vectors and their final diagnosed ischemia pattern categories.

[0109] The similarity between the current patient's clinical feature vector and the clinical feature vectors of each historical case in the associated knowledge base is calculated. A set of similar cases is selected based on the similarity, and a clinical prior probability distribution is generated based on the statistical distribution of different ischemic pattern categories within this set. In some embodiments, the similarity calculation is performed by calculating the cosine similarity between clinical feature vectors. The top 100 historical cases with the highest similarity are selected as the set of similar cases. The distribution ratio of different ischemic pattern categories in these 100 cases is statistically analyzed. For example, if 70 cases of subendocardial ischemia, 25 cases of transmural ischemia, and 5 cases of microvascular ischemia occur, the generated clinical prior probability distribution is 0.70 for subendocardial ischemia, 0.25 for transmural ischemia, and 0.05 for microvascular ischemia. The preliminary ischemic pattern category probability output by the ischemic pattern recognition layer is fused with the clinical prior probability distribution using Bayesian fusion to update the posterior probability distribution. It can be understood that the Bayesian fusion calculation process uses the preliminary ischemic pattern category probability output by the ischemic pattern recognition layer as the likelihood probability and the clinical prior probability distribution as the prior probability. Calculations are performed based on Bayes' theorem:

[0110]

[0111] in: Indicates the ischemia mode category The posterior probability, This indicates the ischemia pattern category output by the ischemia pattern recognition layer. The likelihood probability (i.e., the initial probability). Indicates the ischemia mode category Clinical prior probability, The total number of ischemic patterns is represented by the denominator, which is the sum of all patterns normalized.

[0112] In practice, for the example data, the posterior probability numerator for subendocardial ischemia is calculated as 0.65 * 0.70 = 0.455, the posterior probability numerator for transmural ischemia is 0.25 * 0.25 = 0.0625, and the posterior probability numerator for microvascular ischemia is 0.10 * 0.05 = 0.005. Summing the posterior probability numerators for all ischemia pattern categories yields a normalized denominator of 0.455 + 0.0625 + 0.005 = 0.5225. Dividing the posterior probability numerator of each ischemic pattern category by the normalized denominator yields the normalized posterior probabilities: 0.455 / 0.5225 ≈ 0.871 for subendocardial ischemia, 0.0625 / 0.5225 ≈ 0.120 for transmural ischemia, and 0.005 / 0.5225 ≈ 0.009 for microvascular ischemia. Combining these normalized posterior probabilities for all ischemic pattern categories creates a new probability distribution, the posterior probability distribution [0.871, 0.120, 0.009]. Based on this posterior probability distribution, the strength of support for each ischemic pattern from the current patient's clinical information is calculated, representing the clinical evidence support level. In some embodiments, the clinical evidence support level can directly use the value of the posterior probability distribution, which reflects the consistency between clinical information and the radiographic ischemic pattern. Alternatively, the posterior probability distribution can be compared to a uniform distribution, and a support score can be generated by calculating the KL divergence or other dissimilarity measures.

[0113] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A deep learning-based three-dimensional medical image classification method for myocardial ischemia, characterized in that, Includes the following steps: Acquire three-dimensional medical imaging data of the heart of the patient to be analyzed. The three-dimensional medical imaging data of the heart includes coronary angiography image sequence, myocardial perfusion image sequence and cardiac structure image sequence. Multimodal registration and feature extraction are performed on the three-dimensional medical imaging scan data of the heart to generate multidimensional cardiac feature data that integrates vascular, blood flow and tissue features; The multidimensional cardiac feature data is input into a pre-trained three-dimensional feature parsing network. The three-dimensional feature parsing network is used to extract the spatial distribution features and hemodynamic abnormality patterns of the cardiac function impairment area, including: The three-dimensional feature parsing network consists of a spatial encoding path, a feature decoding path, and a multi-scale feature fusion module. The spatial coding path progressively downsamples the multidimensional cardiac feature data, extracts the contextual features of the heart at multiple scales, and generates a multi-scale contextual feature map. The feature decoding path progressively upsamples the multi-scale context feature map, restoring spatial resolution while fusing features of the same scale from the encoding path through skip connections to generate a high-resolution feature reconstruction map. The multi-scale feature fusion module performs cross-scale aggregation on features from different levels of the spatial coding path and features from the feature decoding path to generate a comprehensive representation vector of cardiac function. The regions representing local functional decline are parsed from the comprehensive representation vector to generate the spatial distribution features of the cardiac function damage regions. The spatial distribution features include the volume, spatial location, and topological relationship of the damage regions with the coronary artery supply regions. The time delay, intensity reduction, and spatial mismatch patterns of hemorrhage perfusion are analyzed from the comprehensive characterization vector to generate the hemodynamic abnormality pattern, which includes the pattern and extent of perfusion defects and their correspondence with abnormal electrocardiogram leads. Based on the spatial distribution characteristics and abnormal hemodynamic patterns, a quantitative map of myocardial ischemia risk is constructed, which includes the distribution of ischemia degree and the marking of ischemia range. The myocardial ischemia risk quantification map is input into a multi-level classification decision engine, which calculates the ischemia type discrimination probability and outputs myocardial ischemia image feature classification labels.

2. The deep learning-based three-dimensional medical image classification method for myocardial ischemia according to claim 1, characterized in that, The process of performing multimodal registration and feature extraction on the three-dimensional medical image scan data of the heart to generate multidimensional cardiac feature volume data that integrates vascular, blood flow, and tissue features includes: The coronary angiography image sequence is segmented to extract vessel wall boundaries, and data on changes in vessel centerline and lumen diameter are extracted to generate vessel morphology feature sequences. The myocardial perfusion image sequence was subjected to temporal registration and signal intensity analysis. The time to peak perfusion, peak perfusion intensity and area under the perfusion curve of each myocardial segment were calculated to generate a blood flow perfusion dynamics feature map. The cardiac structure image sequence is segmented into myocardial wall boundaries, and myocardial wall thickness, myocardial wall motion amplitude and myocardial strain rate data are extracted to generate a myocardial tissue structure feature field. A unified spatial coordinate system based on cardiac anatomical landmarks is established, and the vascular morphological feature sequence, blood flow perfusion dynamic feature map and myocardial tissue structural feature field are spatially transformed and aligned to the unified spatial coordinate system. The aligned vascular morphology feature sequence, blood flow perfusion dynamics feature map, and myocardial tissue structure feature field are spliced ​​at the voxel level to form multi-channel three-dimensional data containing vascular geometry, blood flow dynamics, and myocardial structure, namely the multi-dimensional cardiac feature data.

3. The deep learning-based three-dimensional medical image classification method for myocardial ischemia according to claim 2, characterized in that, The construction of a quantitative map of myocardial ischemia risk based on the spatial distribution characteristics and abnormal hemodynamic patterns includes: The spatial distribution characteristics of the cardiac function damage area are mapped back to the unified spatial coordinate system to generate a damage probability distribution field in the three-dimensional space of the heart. The hemodynamic abnormality pattern is quantified into a perfusion abnormality score for each myocardial segment, which is calculated based on the degree of delay in the time to peak perfusion and the percentage decrease in peak perfusion intensity. By fusing the damage probability distribution field with the blood perfusion abnormality score, a comprehensive ischemic risk value is calculated for each spatial voxel. The comprehensive ischemic risk value is a weighted sum of the damage probability and the blood perfusion abnormality score. Using the comprehensive ischemic risk value as the filler value, a color-coded quantification map of myocardial ischemia risk is generated on a three-dimensional cardiac model. In the myocardial ischemia risk quantification map, high-risk ischemic areas and medium-risk ischemic areas are automatically delineated and marked according to the threshold of the comprehensive ischemia risk value, thus completing the ischemia range marking.

4. The deep learning-based three-dimensional medical image classification method for myocardial ischemia according to claim 3, characterized in that, The process of inputting the myocardial ischemia risk quantification map into a multi-level classification decision engine, calculating the ischemia type discrimination probability through the multi-level classification decision engine, and outputting myocardial ischemia image feature classification labels includes: The multi-layer classification decision engine includes an ischemia pattern recognition layer, a clinical correlation analysis layer, and a decision fusion layer. The ischemia pattern recognition layer receives the myocardial ischemia risk quantification map, uses a pre-trained convolutional neural network to extract the global morphological pattern and local texture pattern of the ischemic region, and outputs the preliminary ischemia pattern category probability. The clinical association analysis layer receives patient clinical information data, which includes multi-dimensional parameters such as patient physiological, electrophysiological and biochemical indicators. The clinical information data is logically associated with the preliminary ischemia pattern output by the ischemia pattern recognition layer to calculate the clinical evidence support. The decision fusion layer integrates the preliminary ischemia pattern category probability with the clinical evidence support, and calculates the final discrimination probability of each ischemia type through weighted decision rules; The ischemia type with the highest probability is selected from the final discrimination probabilities, and combined with the preset classification rules, a myocardial ischemia image feature classification label containing the specific ischemia type, ischemia range, and ischemia degree is generated.

5. The deep learning-based three-dimensional medical image classification method for myocardial ischemia according to claim 4, characterized in that, The process involves temporal registration and signal intensity analysis of the myocardial perfusion image sequence, calculating the time to peak perfusion, peak perfusion intensity, and area under the perfusion curve for each myocardial segment, and generating a hemoperfusion dynamics feature map, including: Using the cardiac cycle as a reference, each frame of the myocardial perfusion image sequence is marked with a time point to establish the correspondence between image frames and cardiac phases; Based on the aforementioned correspondence, the mutual information maximization criterion is used to perform temporal registration of image sequences of adjacent cardiac cycles, so that myocardial images of the same phase in different cycles are spatially aligned. On the image sequence with completed phase registration, the left ventricular myocardium is divided into multiple segments according to the standard myocardial segmentation model; For each myocardial segment, the average signal intensity value of each myocardial segment in all temporal images is extracted to form the time-signal intensity curve of the myocardial segment; Identify the maximum signal strength from the time-signal strength curve, record the time phase corresponding to the maximum signal strength as the perfusion peak time, and record the maximum signal strength as the perfusion peak intensity; The time-signal intensity curve is numerically integrated from the start point to the end point of the curve, and the resulting integral value is the area under the perfusion curve. The peak perfusion time, peak perfusion intensity, and area under the perfusion curve of all myocardial segments are arranged according to the spatial location of the segments to generate a two-dimensional blood flow perfusion dynamics feature map, in which the three parameter values ​​of each segment are stored in different channels of the feature map.

6. The deep learning-based three-dimensional medical image classification method for myocardial ischemia according to claim 5, characterized in that, The spatial encoding path progressively downsamples the multidimensional cardiac feature volume data, extracting cardiac contextual features at multiple scales to generate a multi-scale contextual feature map, including: The multidimensional cardiac feature data is used as input and sequentially passed through an encoding module consisting of multiple concatenated convolution and pooling operation layers; In each encoding module, a 3D convolution operation is first performed to extract local features, followed by a 3D max pooling operation to reduce the spatial resolution of the feature map and expand the receptive field. After processing by a series of encoding modules, a series of feature maps from high resolution to low resolution are obtained, namely the multi-scale context feature maps. The low-resolution feature maps contain richer global context information, while the high-resolution feature maps retain more spatial detail information. During the encoding process, the feature map before each pooling operation is recorded for subsequent skip connections with the feature decoding path.

7. The deep learning-based three-dimensional medical image classification method for myocardial ischemia according to claim 6, characterized in that, The method of fusing the damage probability distribution field with the blood perfusion anomaly score to calculate the comprehensive ischemic risk value for each spatial voxel includes: The damage probability distribution field is normalized to map the damage probability value of each voxel to a standard interval between zero and one. The abnormal blood perfusion score was normalized, and the abnormal blood perfusion score of each myocardial segment was also mapped to a standard interval between zero and one. Based on the pre-configured damage probability weight and blood perfusion abnormality weight, the sum of the damage probability weight and blood perfusion abnormality weight is one. For each spatial voxel in the unified spatial coordinate system, find its corresponding myocardial segment and obtain the normalized blood perfusion abnormality score of the myocardial segment; The sum of the normalized damage probability value of the spatial voxel multiplied by the damage probability weight, plus the normalized perfusion abnormality score of its corresponding myocardial segment multiplied by the perfusion abnormality weight, is the comprehensive ischemic risk value of the spatial voxel.

8. The deep learning-based three-dimensional medical image classification method for myocardial ischemia according to claim 7, characterized in that, The clinical association analysis layer receives patient clinical information data, which consists of multi-dimensional parameters including patient physiological, electrophysiological, and biochemical indicators. It performs logical association analysis between the clinical information data and the preliminary ischemia pattern output by the ischemia pattern recognition layer to calculate the clinical evidence support level, including: The patient's clinical information data is structured and encoded into a fixed-dimensional clinical feature vector. Construct an associated knowledge base containing the correspondence between clinical feature vectors and ischemia pattern categories; Calculate the similarity between the current patient's clinical feature vector and the clinical feature vectors of each historical case in the associated knowledge base, filter out a set of similar cases based on the similarity, and generate a clinical prior probability distribution based on the statistical distribution of different ischemic patterns in the set. The initial ischemia pattern category probability output by the ischemia pattern recognition layer is combined with the clinical prior probability distribution using Bayesian fusion calculation to update the posterior probability distribution. Based on the posterior probability distribution, the strength of support of the current patient's clinical information for each ischemic pattern is calculated, which is the clinical evidence support level. Its value reflects the degree of consistency between clinical information and radiographic ischemic patterns.

9. The deep learning-based three-dimensional medical image classification method for myocardial ischemia according to claim 8, characterized in that, The step of performing Bayesian fusion calculations on the preliminary ischemia pattern category probability output by the ischemia pattern recognition layer and the clinical prior probability distribution to update the posterior probability distribution includes: The initial ischemia pattern category probability output by the ischemia pattern recognition layer is used as the likelihood probability; The clinical prior probability distribution is used as the prior probability; According to Bayes' theorem, the posterior probability numerator of each ischemia mode category is calculated, and the posterior probability numerator is equal to the likelihood probability of the ischemia mode category multiplied by the prior probability of the ischemia mode category. Summing the posterior probability numerators for all ischemic pattern categories yields the normalized denominator; Divide the posterior probability numerator of each ischemia mode category by the normalized denominator to obtain the normalized posterior probability of the ischemia mode category. The normalized posterior probabilities of all ischemic mode categories are combined into a new probability distribution, which is the posterior probability distribution.

Images