Myocardial perfusion image classification method and system based on single resting state

Abstract

The invention discloses a myocardial perfusion image classification method and system based on a single resting state, and belongs to intelligent analysis and auxiliary diagnosis of medical images. The method comprises the following steps: acquiring SPECT three-dimensional voxel data in the single resting state, and performing image reconstruction by adopting an OSEM algorithm; segmenting the myocardial region by using a pre-trained U-Net convolutional neural network, and mapping a segmentation result to a two-dimensional polar coordinate graph conforming to the AHA17 segment model; extracting a multi-dimensional feature vector; a single-phase inference model MSR-Net based on biphase labeling is constructed, in the training stage, segment classification labels of resting-load biphase images are used, in combination with segment consistency indexes, label correction is carried out, and in the inference stage, only resting state features are input, and then a pixel-level perfusion defect distribution diagram and ischemia scores of 17 myocardial segments can be output. Quantitative and segmental evaluation of myocardial ischemia can be completed by using single resting state imaging, exercise or drug load examination is avoided, and cardiovascular adverse events and complication risks are reduced.

Description

Technical Field

[0001] This application relates to the field of intelligent analysis and assisted diagnosis of medical images, and more specifically, to a method and system for classifying myocardial perfusion images based on a single resting state. Background Technology

[0002] Single-photon emission computed tomography (SPECT), an important nuclear medicine functional imaging technique, generates three-dimensional functional images by detecting the distribution of radioactive tracers in tissues. In the field of myocardial perfusion imaging, SPECT technology can acquire blood flow perfusion information of different segments of the left ventricle, generating three-dimensional voxel data reflecting the functional status of the left ventricle myocardium. With the development of medical imaging technology, how to automatically and accurately extract valuable information from SPECT images through advanced image processing and analysis methods has become a research hotspot in the field of medical image processing.

[0003] Traditional SPECT myocardial perfusion image analysis employs a resting-stress biphasic imaging mode, acquiring images separately under resting and stress conditions. Reversible perfusion abnormalities are identified by comparing the differences between the two images. While this mode provides rich pathophysiological information, it also has several drawbacks: 1. High examination burden: Biphasic acquisition requires a long time (usually 3-4 hours), doubling the radiation dose and increasing patient waiting and exposure risks. 2. Stringent implementation conditions: Induction of stress requires specialized equipment and monitoring conditions, increasing operational complexity and potentially inducing adverse cardiovascular events such as angina and arrhythmias. 3. Difficult image registration: Biphasic images require precise registration, but due to factors such as cardiac motion, respiratory motion, and postural changes, spatial alignment is challenging and easily leads to errors in subsequent comparative analysis.

[0004] At the level of image processing technology, existing SPECT myocardial perfusion image analysis methods also face the following limitations:

[0005] First, the biphasic image processing procedure is complex, requiring precise registration of three-dimensional images acquired at two time points. However, due to factors such as heart movement and respiratory displacement, the accuracy of image registration is difficult to guarantee, affecting the results of subsequent comparative analysis.

[0006] Secondly, existing methods are relatively simplistic in feature extraction, mostly focusing only on direct comparison of perfusion intensity, lacking comprehensive utilization of multi-dimensional information such as image texture features and morphological features. SPECT images themselves have limited resolution and low signal-to-noise ratio, making it difficult for a single feature to fully represent complex perfusion patterns.

[0007] Third, traditional image analysis methods rely on manually set thresholds and rules, making it difficult to adapt to image differences generated by different devices and acquisition protocols. Some existing deep learning-based methods, such as patents CN115082409B and CN202111588861.2, although incorporating artificial intelligence technology, still require biphasic images as input and cannot independently complete ischemia assessment under uniphasic conditions.

[0008] Fourth, in terms of spatial positioning accuracy, existing methods mostly use coarse assessments based on myocardial segments, lacking pixel-level fine segmentation capabilities, making it difficult to accurately delineate the boundaries and ranges of abnormal perfusion areas, thus limiting the accuracy of quantitative analysis.

[0009] In summary, how to achieve accurate segmentation, quantitative assessment, and multidimensional feature analysis of myocardial ischemia with only single resting-state imaging data, in order to reduce the examination burden, lower the risk, and maintain diagnostic accuracy, has become a pressing technical problem to be solved in the field of intelligent analysis of myocardial perfusion imaging. Summary of the Invention

[0010] This application aims to overcome the problems of existing SPECT myocardial perfusion image analysis, such as reliance on rest-load biphasic imaging, long examination time, high radiation dose, difficulty in image registration, and insufficient diagnostic accuracy due to single feature extraction. It proposes an intelligent computational method and diagnostic system for myocardial ischemia based on a single resting-state SPECT myocardial perfusion imaging. This method employs a biphasic annotation-based single-phase inference strategy. During model training, the comparison results of rest-load biphasic SPECT images are used as the gold standard for ischemia annotation (based on myocardial ischemia-related indicators interpreted from resting + load myocardial perfusion imaging), and labels are optimized through segmental consistency. In the actual inference stage, only a single resting-state image needs to be input to achieve quantitative and segmental assessment of myocardial ischemia, reducing patient risk, improving diagnostic efficiency, and minimizing the time and safety issues associated with load examinations.

[0011] One aspect of this application provides a method for classifying myocardial perfusion images based on a single resting state, comprising: S1, acquiring three-dimensional voxel data (x, y, z) in a single resting state as a SPECT myocardial perfusion image; S2, reconstructing the myocardial perfusion image, and performing denoising and contrast enhancement on the reconstructed image to obtain a preprocessed myocardial perfusion image; S3, segmenting myocardial regions based on the preprocessed myocardial perfusion image using a pre-trained U-Net convolutional neural network; and mapping the segmented myocardial regions to the standard AHA17 mapping algorithm. The segmental model corresponds to a two-dimensional polar coordinate map P(r, θ), where r is the radial coordinate and θ is the angular coordinate; S4, based on the two-dimensional polar coordinate map P(r, θ) and the segmented myocardial region, perfusion intensity features, texture features, and morphological features are extracted through parallel multi-scale convolution to form a multi-dimensional feature vector; S5, a single-phase inference model MSR-Net based on biphasic annotation is constructed, which takes the multi-dimensional feature vector as input, performs ischemia classification on a single resting state myocardial perfusion image, and outputs a pixel-level perfusion defect distribution map and a perfusion score based on myocardial segments.

[0012] Among them, the single-phase inference model MSR-Net based on biphasic annotation uses the classification results corresponding to resting-load biphasic SPECT images as classification labels during the training phase, and performs label correction through label credibility based on segment consistency; during the inference phase, only the multidimensional feature vector of a single resting state needs to be input.

[0013] Further, S2, obtaining the preprocessed myocardial perfusion image includes: using the ordered subset expectation-maximization OSEM algorithm to reconstruct the three-dimensional voxel data (x, y, z) to obtain an initial reconstructed image; based on the initial reconstructed image, obtaining the left ventricular center position (Cx, Cy, Cz) using the minimum error pole fitting algorithm based on the short axis section; using the left ventricular center position (Cx, Cy, Cz) to spatially align the initial reconstructed image using a three-dimensional rigid body registration algorithm to obtain a registered image; using the registered image to perform noise reduction processing on the registered image using adaptive bilateral filtering to obtain a denoised image; and performing multi-scale directional filtering on the denoised image to compensate for edge blurring of the SPECT image and enhance the local contrast of the myocardial boundary and areas of mild perfusion difference to obtain the preprocessed myocardial perfusion image.

[0014] The three-dimensional voxel data (x, y, z) refers to the raw three-dimensional image data acquired by the SPECT imaging device, where each voxel (three-dimensional pixel) represents the concentration value of the radioactive tracer at a certain spatial location of the heart. (x, y, z) are spatial coordinates in a three-dimensional Cartesian coordinate system, where x and y usually represent two directions of the cross-section, and z represents the direction along the long axis of the body.

[0015] A short-axis view is a cross-sectional image perpendicular to the long axis of the left ventricle. In cardiac imaging, short-axis views can display the circular myocardial structure of the left ventricle and are the standard perspective for assessing myocardial perfusion distribution. Multiple consecutive short-axis views can cover the entire left ventricle from the apex to the base of the heart.

[0016] The three-dimensional rigid body registration algorithm refers to a three-dimensional image alignment algorithm that only allows translation and rotation transformations. In this application, the algorithm uses the center position of the left ventricle as a reference to adjust the cardiac image to a standard orientation, ensuring that images acquired by different users or at different times have consistent spatial orientation.

[0017] Multi-scale directional filtering refers to a technique for filtering images at different spatial scales and directions. This filter can adaptively enhance edges based on the directional characteristics of local image structures. In this application, it is used to sharpen myocardial boundaries and enhance subtle perfusion differences within the myocardium while maintaining image smoothness.

[0018] SPECT images refer to Single Photon Emission Computed Tomography (SPECT) images. In this application, it specifically refers to myocardial perfusion images acquired after the injection of a radioactive tracer. These images reflect the blood perfusion status of myocardial tissue, and areas of perfusion deficiency are characterized by reduced or absent radioactive uptake.

[0019] Further, in step S3, the segmented myocardial region is mapped to a two-dimensional polar coordinate map P(r, θ) corresponding to the standard AHA17 segment model. This includes: inputting the preprocessed myocardial perfusion image into a pre-trained U-Net convolutional neural network to output a three-dimensional segmentation mask for the myocardial region; extracting the myocardial contour edge based on the three-dimensional segmentation mask; calculating the gray-level centroid coordinates of the myocardial region based on the myocardial contour edge; determining the myocardial long axis direction by least squares fitting based on the gray-level centroid coordinates of the myocardial region and the left ventricular center position (Cx, Cy, Cz); and adjusting the center point and starting angle of the polar coordinate unfolding according to the myocardial long axis direction and the myocardial contour edge. The resulting polar coordinate map is consistent with the spatial layout of the standard AHA17 segment model. Along the long axis of the myocardium, myocardial region data is sampled in order from the apex to the base of the heart, resulting in multiple short-axis slices. For each short-axis slice, with the determined unfolding center point as the origin, the circular myocardial region is unfolded into a two-dimensional polar coordinate representation, where the radial coordinate r represents the distance from the endocardium to the epicardium, and the angular coordinate θ represents the circumferential position. The polar coordinate representations of all short-axis slices are spatially integrated to generate a complete two-dimensional polar coordinate map P(r, θ), and according to the definition of the standard AHA17 segment model, the polar coordinate map is divided into 17 corresponding myocardial segment regions.

[0020] The standard AHA 17-segment model refers to the 17-segment division of the left ventricle established by the American Heart Association. This model divides the left ventricular myocardium into 17 standardized regions based on anatomical location: 6 segments at the base (anterior wall, anterior septum, inferior septum, inferior wall, inferior lateral wall, and anterior lateral wall), 6 segments in the middle (named the same as the base), 4 segments at the apex (anterior wall, septum, inferior wall, and lateral wall), and 1 segment at the apical cap.

[0021] A two-dimensional polar plot P(r, θ) is a polar coordinate mapping that unfolds three-dimensional myocardial data into a two-dimensional plane representation; it is also called a bull's eye plot. Here, r is the radial coordinate, representing the radial distance from the endocardium to the epicardium; θ is the angular coordinate (0-360 degrees), representing the circumferential position of the myocardium. This plot can completely display the perfusion distribution of the entire left ventricular myocardium on a single two-dimensional image.

[0022] The centroid coordinates refer to the weighted average position of all pixels within the myocardial region, with the weights being the grayscale value of each pixel (i.e., the concentration of the radioactive tracer). The calculation formula is: Centroid coordinates = Σ(pixel coordinates × pixel grayscale value) / Σ pixel grayscale value. These coordinates reflect the central location of the radioactive distribution in the myocardium.

[0023] The long axis of the myocardium, referring to the axis pointing from the apex of the left ventricle to the center of the mitral valve, is the main axis of symmetry of the left ventricle. In this application, the optimal long axis direction is obtained by combining the gray-scale centroid coordinates and the center position of the left ventricle, and is used as the reference axis for subsequent image processing by least squares fitting.

[0024] The center point and starting angle of polar coordinate unfolding: Center point: refers to the position of the center of the circle when unfolding the circular myocardium into polar coordinates, usually chosen as the center of the left ventricular cavity on each short-axis slice. Starting angle: refers to the anatomical position corresponding to θ = 0° when unfolding polar coordinates. According to the AHA standard, the junction of the anterior ventricular septum and the right ventricle is usually defined as the starting angle to ensure spatial consistency of images for different users.

[0025] Myocardial segments refer to the myocardial regions divided according to the AHA17 segmental model. In a polar coordinate diagram, each segment is defined by specific radial and angular ranges: six segments each in the base and middle occupy a 60-degree angular range, and four segments in the apex each occupy a 90-degree range, with the apical cap located at the center of the polar coordinate diagram. Each segment corresponds to a specific coronary artery supply area, used to locate and quantify myocardial ischemia.

[0026] Furthermore, S4 constructs a multi-dimensional feature vector, including: constructing a multi-scale convolutional network with three parallel branches; using the first branch to perform a convolution operation on the two-dimensional polar coordinate map P(r, θ) to extract perfusion intensity features; using the second branch to perform a convolution operation on the two-dimensional polar coordinate map P(r, θ) to extract texture features; using the third branch to perform a convolution operation on the three-dimensional segmentation mask of the myocardial region to extract morphological features; and concatenating the perfusion intensity features, texture features, and morphological features through a feature fusion layer to obtain a multi-dimensional feature vector.

[0027] Furthermore, the perfusion intensity features are extracted using the first branch, including: calculating the mean perfusion intensity of all pixels within each of the 17 myocardial segment regions based on the two-dimensional polar coordinate map P(r, θ), as the average perfusion value; the standard deviation of the perfusion intensity within each segment, as the standard deviation of the perfusion value; and the difference in the average perfusion values ​​between adjacent segments, as the perfusion gradient feature between segments; and combining the average perfusion value, the standard deviation of the perfusion value, and the perfusion gradient feature of the 17 segments to obtain the perfusion intensity features.

[0028] Furthermore, the second branch is used to extract texture features, including: calculating the gray-level co-occurrence matrix (GLCM) for each myocardial segment in the two-dimensional polar coordinate image P(r, θ); calculating contrast features based on the GLCM to reflect the degree of local gray-level changes; calculating energy features based on the GLCM to reflect the uniformity of image texture; calculating homogeneity features based on the GLCM to assess the similarity of local image regions; and combining the contrast, energy, and homogeneity features from each stage to obtain the texture features.

[0029] Furthermore, morphological features are extracted using the third branch, including: calculating the distance from the endocardium to the epicardium on each short-axis slice based on the three-dimensional segmentation mask of the myocardial region, as the myocardial thickness; calculating the myocardial volume by summing the myocardial region area and interlayer spacing of each short-axis slice; calculating the roundness and compactness indices of the myocardial shape based on the contour information of the myocardial region; and combining the myocardial thickness, myocardial volume, roundness, and compactness indices to obtain the morphological features.

[0030] Furthermore, in S5, a single-phase inference model MSR-Net based on biphasic annotation is constructed, including: constructing the single-phase inference model MSR-Net, which includes a feature encoder, a multi-task decoder, and a label optimization module; during the training phase, paired resting-load biphasic SPECT images are acquired, and the multidimensional feature vector of the resting-state image and the ischemia annotation information of the load-state image are extracted respectively; wherein, the load-state image represents the myocardial perfusion image acquired under exercise load; the load-state image and the resting-state image are compared to generate a pixel-level ischemia annotation mask and ischemia score labels for 17 myocardial segments; the multidimensional feature vector of the resting state is input into the feature encoder, and high-level features are extracted through multi-layer residual convolutional blocks; High-level features are input into a multi-task decoder for pixel-level segmentation and segment scoring. During training, a label optimization module calculates the consistency index of feature distribution within each stage. For stages with consistency indices below a threshold, cluster centers of similar samples are used to re-estimate the corresponding label values ​​to correct the labels. A weighted loss function is used to optimize the pixel-level segmentation and segment scoring tasks, iteratively updating the model parameters. In the inference stage, only a single resting-state multidimensional feature vector needs to be input. The trained single-phase inference model MSR-Net outputs a pixel-level perfusion defect distribution map and perfusion scoring results for 17 myocardial segments. The perfusion scoring results include: resting score (SRS), total score (SSS), and ischemia score (SDS).

[0031] Among them, the single-phase inference model MSR-Net (Multi-Scale Rest-only Network) refers to a deep learning network model used for ischemia diagnosis of a single resting-state myocardial perfusion image. During training, the diagnostic results of biphasic images are used as supervision signals, but in practical applications, only resting-state images are needed to complete the diagnosis, realizing the strategy of "biphasic training and single-phase inference".

[0032] The label optimization module corrects labels during training because labels obtained from resting-load biphasic images may contain noise or errors. By calculating the consistency of feature distributions, potentially mislabeled samples can be identified, and these labels are then corrected using cluster centers of similar samples. Consistency index calculation: For each myocardial segment, the density of the distribution of feature vectors of all training samples within that segment in the feature space is calculated (e.g., variance, entropy). When the feature distribution of a segment is too dispersed (consistency index below a preset threshold), it indicates that the segment's labels may contain errors or noise. For detected abnormal segments, methods such as K-means clustering or Gaussian mixture models are used to find the most similar sample groups, and the label center values ​​(e.g., mean or median) of these samples are used to replace the original labels.

[0033] The Summed Rest Score (SRS) is the sum of perfusion defect scores for 17 myocardial segments at rest. Each segment is scored from 0 to 4 (0 = normal, 1 = mild reduction, 2 = moderate reduction, 3 = severe reduction, 4 = no perfusion). The SRS reflects the degree of myocardial perfusion abnormality at rest.

[0034] The Summed Stress Score (SSS) is calculated using a model based on resting-state images in this application. It represents the sum of perfusion defect scores for 17 myocardial segments under predicted stress conditions. A higher SSS value indicates more severe myocardial ischemia.

[0035] The Summed Difference Score (SDS) is the difference between the SSS and SRS (SDS = SSS - SRS), reflecting the degree of reversible myocardial ischemia. An SDS > 0 indicates the presence of reversible myocardial ischemia, while a higher SDS indicates a more severe degree of ischemia. In this application, this value is obtained through model prediction and is used to assess myocardial ischemia requiring revascularization therapy.

[0036] Furthermore, the high-level features are input into the multi-task decoder for pixel-level segmentation and segment scoring, including: The multi-task decoder comprises two parallel branches: The first branch generates a pixel-level perfusion defect probability map through a deconvolutional layer; where the pixel-level perfusion defect probability map refers to the two-dimensional probability distribution image output by the first branch of the deep learning model, whose spatial resolution is the same as the input two-dimensional polar coordinate map P(r, θ). Each pixel location in this image corresponds to a probability value (range 0-1), indicating the likelihood of myocardial perfusion defect at that location. Pixel-level: refers to predicting each individual pixel in the image, rather than just making a rough judgment on regions or segments. Perfusion defect: refers to the manifestation of reduced or absent uptake of radioactive tracers due to insufficient myocardial blood perfusion. The second branch outputs perfusion scores for 17 myocardial segments through a fully connected layer.

[0037] Another aspect of this application provides a method for classifying myocardial perfusion images based on a single resting state, comprising: a data acquisition module for acquiring three-dimensional voxel data (x, y, z) in a single resting state as a SPECT myocardial perfusion image; an image preprocessing module for reconstructing the myocardial perfusion image and performing denoising and contrast enhancement on the reconstructed image to obtain a preprocessed myocardial perfusion image; and a myocardial segmentation and mapping module for segmenting myocardial regions based on the preprocessed myocardial perfusion image using a pre-trained U-Net convolutional neural network and mapping the segmented myocardial regions to a standard... The AHA17 segment model corresponds to a two-dimensional polar coordinate map P(r, θ), where r is the radial coordinate and θ is the angular coordinate. The feature extraction module extracts perfusion intensity features, texture features, and morphological features through parallel multi-scale convolution based on the two-dimensional polar coordinate map P(r, θ) and the segmented myocardial regions, forming a multi-dimensional feature vector. The classification and scoring module includes a bidirectional annotation-based single-phase inference model MSR-Net, which takes the multi-dimensional feature vector as input, performs ischemia classification on a single resting-state myocardial perfusion image, and outputs a pixel-level perfusion defect distribution map and a perfusion score based on myocardial segments.

[0038] Compared to existing technologies, the advantages of this application are:

[0039] To address the problems of long examination time, high radiation dose, and inability of some patients to tolerate stress tests caused by the need for rest-load biphasic acquisition in existing SPECT myocardial perfusion imaging technologies, this application provides a myocardial perfusion image classification method based on a single resting state. It innovatively adopts a biphasic annotation single-phase inference strategy, that is, in the model training stage, the comparison results of rest-load biphasic SPECT myocardial perfusion images are used as the gold standard for ischemia annotation, but in the inference stage of actual application, only a single resting state SPECT image needs to be input.

[0040] First, the resting-state SPECT images are finely preprocessed and segmented into myocardial data, mapping the three-dimensional myocardial data to a standardized two-dimensional polar coordinate representation. Then, multi-dimensional features including perfusion intensity, texture, and morphology are extracted through a parallel multi-scale convolutional network. Finally, the MSR-Net model is constructed, which is trained using the diagnostic results of biphasic images to obtain the mapping relationship from monophasic features to ischemia diagnosis, and the reliability of training is improved through a label optimization strategy.

[0041] This application (1) can automatically diagnose myocardial ischemia with only a single time point of resting state imaging, reducing the examination time from the traditional 3-4 hours to within 45 minutes; (2) avoids the high risk and contraindications of stress tests, improving the safety and applicability of the examination; (3) achieves pixel-level perfusion defect localization and quantitative scoring based on AHA17 segments through deep learning models, providing diagnostic accuracy comparable to biphasic imaging; (4) reduces radiation exposure dose by about 50%, improving the accessibility and patient compliance of SPECT myocardial perfusion imaging. Attached Figure Description

[0042] Figure 1 This is an exemplary flowchart of a SPECT myocardial perfusion image classification method based on a single resting state, as described in this application.

[0043] Figure 2 This is a flowchart of the image preprocessing process for this application.

[0044] Figure 3 This is a schematic diagram of the automatic segmentation and standardized view reconstruction process of the left ventricular myocardial region based on deep learning in this application;

[0045] Figure 4 This is a schematic diagram of the multidimensional image feature extraction and fusion process of this application;

[0046] Figure 5 This is a schematic diagram of the training and validation process of the machine learning model based on the gold standard of biphasic images in this application. Detailed Implementation

[0047] The present application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] like Figure 1 As shown, this embodiment provides a classification method for SPECT myocardial perfusion images based on a single resting state, including the following steps:

[0050] S1: Data Acquisition. The first step of this application is data acquisition, which aims to obtain high-quality three-dimensional voxel data of myocardial perfusion in a single resting state, so as to provide accurate raw input data for subsequent image reconstruction, feature extraction and intelligent diagnosis.

[0051] In the specific implementation process, the subject is injected with the radioactive tracer technetium-12 via the antecubital vein while at rest. 99m Labeled methoxyisobutylisocyanuric acid ( 99mThe Tc-MIBI solution is used, and the radiochemical purity of the imaging agent is not less than 95%. The preferred injection dose is 740–925 MBq. Imaging acquisition should begin within a suitable distribution time (usually 30–60 minutes) after injection to ensure that the uptake and distribution of the radiotracer in the myocardial tissue reaches a stable state.

[0052] The acquisition equipment is preferably an integrated single-photon emission computed tomography / computed tomography (SPECT / CT) imaging system, equipped with a parallel-aperture low-energy high-resolution collimator, with an energy peak set at 140 keV and an energy window width of 20%. The acquisition matrix size is 128×128, with a magnification of 1.45, and operates with dual probes arranged at a 90° angle, each probe rotating 90° to achieve a total acquisition range of 180°. The detector rotation adopts a stepping method, with a step angle of 6° / frame, and the acquisition time for each frame is 35 seconds to ensure sufficient statistical counting and image signal-to-noise ratio.

[0053] The acquired raw data consisted of three-dimensional voxel data (x, y, z) from a single resting state, where each voxel value was expressed as a count, representing the concentration distribution of the radioactive tracer at that spatial location. In the three-dimensional coordinate system, the x and y axes represented the transverse direction, and the z-axis represented the direction along the longitudinal axis of the body (from the apex to the base of the heart). This data completely covered the subject's left ventricle and surrounding myocardial tissue, ensuring the accuracy and reliability of subsequent image reconstruction and processing.

[0054] S2: Image preprocessing, such as Figure 2 As shown in this application, the image preprocessing module is used to reconstruct, locate, align, denoise, and enhance the acquired single-shot resting-state SPECT three-dimensional voxel data to eliminate acquisition artifacts and noise in the original image and improve the accuracy of subsequent myocardial segmentation and feature extraction.

[0055] First, image reconstruction is performed on the three-dimensional voxel data (x, y, z). Preferably, the Ordered Subsets Expectation Maximization (OSEM) algorithm is used, which can converge quickly within a finite number of iterations and obtain high reconstruction quality. In this embodiment, the number of OSEM reconstruction iterations is set to 8, and the number of subsets is set to 6, in order to control the computation time while ensuring spatial resolution, and obtain the initial reconstructed image.

[0056] Secondly, the left ventricular center is located in the initial reconstructed image. Specifically, in the short-axis section image, based on the myocardial contour information, the position of the left ventricular center (Cx, Cy, Cz) is estimated using a minimum error pole fitting algorithm. This method minimizes the error between the fitted contour and the actual edge by fitting the polar coordinate center of the myocardial annular structure, thereby obtaining stable and robust center coordinates and providing a reference point for subsequent spatial registration.

[0057] Next, spatial alignment is performed. Based on the acquired left ventricular center position, a 3D Rigid Registration algorithm is used to spatially transform the initial reconstructed image, allowing only translation and rotation without introducing scaling or deformation, to align the long axis of the myocardium with the standard anatomical coordinate system. This process eliminates the influence of differences in patient position and changes in cardiac posture on image analysis, ensuring spatial consistency across different cases.

[0058] Denoising is then performed on the spatially aligned image. Preferably, an adaptive bilateral filtering method is used, which smooths noise while preserving edge information of the image. In this embodiment, the standard deviation of the spatial domain kernel σs is set to 2.0, and the standard deviation of the intensity domain kernel σr is set to 25, to balance noise suppression and detail preservation, and reduce the interference of low-count statistical noise in SPECT on perfusion features.

[0059] Finally, contrast enhancement processing is performed. Based on multi-scale oriented filtering technology, edge responses are detected and enhanced at multiple spatial scales (scale = 3 in this embodiment) and multiple directions (number of directions = 8 in this embodiment) to compensate for the inherent spatial resolution limitations of SPECT imaging, sharpen myocardial boundaries, and enhance local contrast in areas of mild perfusion difference, thereby improving the model's sensitivity to early and mild ischemic perfusion changes.

[0060] The multi-stage image preprocessing described above can significantly improve the signal-to-noise ratio, spatial alignment accuracy, and local contrast of resting SPECT myocardial perfusion images, laying a high-quality input data foundation for subsequent deep learning-based myocardial segmentation and multi-dimensional feature extraction.

[0061] S3: As Figure 3 As shown, the third step of this application aims to accurately extract the left ventricular myocardial region from the preprocessed myocardial perfusion image and normalize and map it into a two-dimensional polar coordinate representation, corresponding to the coronary artery distribution area, so as to facilitate subsequent multidimensional feature extraction and segmental analysis.

[0062] First, the preprocessed image is input into a pre-trained U-Net convolutional neural network model, which outputs a 3D segmentation mask for the myocardial region. The U-Net model, through an encoder-decoder structure combined with skip connections, achieves fine segmentation of low signal-to-noise ratio SPECT images, accurately separating myocardial tissue from background tissue.

[0063] Subsequently, the myocardial contour edges were extracted based on a 3D segmentation mask, and the gray-level centroid coordinates of the myocardial region were calculated. Combined with the left ventricular center position (Cx, Cy, Cz), the direction of the left ventricular long axis was obtained by fitting using the least squares method.

[0064] Next, based on the long axis direction and myocardial contour edge information, the center point and starting angle of the polar coordinate unfolding were determined to ensure that the polar coordinate map was consistent with the spatial layout of the American Heart Association (AHA) 17-segment model. Multi-slice short-axis sections of the myocardial region were sampled along the apex to base of the heart, and the circular myocardium was unfolded into a two-dimensional polar coordinate representation P(r, θ), where r represents the radial distance from the endocardium to the epicardium, and θ represents the circumferential position angle (0° to 360°).

[0065] Finally, all short-axis slice unfolded images were merged to generate a complete two-dimensional polar coordinate map, which was then divided into 17 myocardial segments according to the AHA17 segment model, providing a spatial benchmark for subsequent segmented feature calculation and scoring.

[0066] S4: Multidimensional feature extraction, such as Figure 4 As shown, the fourth step of this application aims to extract comprehensive perfusion intensity, texture and morphological features from standardized two-dimensional polar coordinates and three-dimensional segmentation masks to achieve a multi-dimensional quantitative description of myocardial perfusion status.

[0067] To this end, a multi-scale convolutional network structure with three parallel branches is constructed:

[0068] The first branch (perfusion intensity feature extraction): Input a two-dimensional polar coordinate graph P(r, θ), and calculate the average perfusion value, standard deviation and perfusion gradient of each myocardial segment to reflect the absolute level and trend of local blood flow.

[0069] The second branch (texture feature extraction): Based on the two-dimensional polar coordinate map, the gray-level co-occurrence matrix (GLCM) is calculated segment by segment to extract contrast, energy and homogeneity features, and to characterize the spatial structure and uniformity of local gray-level distribution.

[0070] The third branch (morphological feature extraction): using a three-dimensional segmentation mask, the thickness of the myocardium is measured slice by slice, the myocardial volume is calculated by combining the area and the interlayer spacing, and the roundness and compactness indicators of the myocardial contour are extracted.

[0071] The output feature vectors of the three branches are concatenated through a feature fusion layer to form a complete multidimensional feature vector, which contains both infused quantization information and reflects texture details and morphological features, providing rich input for subsequent intelligent classification.

[0072] S5: Single-phase reasoning classification and scoring based on biphasic annotation, such as Figure 5 As shown, the fifth step of this application constructs a multi-scale rest-only network (MSR-Net) and uses biphasic annotation training to achieve uniphasic inference, thereby completing the automatic classification and scoring of myocardial ischemia.

[0073] During the training phase, paired resting and stressed images are input. Multidimensional feature vectors are extracted from the resting images, and pixel-level ischemia masks and segmental scoring labels (SRS, SSS, SDS) are generated from the stressed images. The feature encoder employs a multi-layer residual convolutional structure to extract high-level features; the multi-task decoder includes pixel-level segmentation and segmental scoring branches, simultaneously outputting perfusion defect probability maps and 17-segment scoring results. The label optimization module identifies and corrects low-confidence labels based on the consistency of segment feature distribution to reduce the impact of manual annotation noise.

[0074] During the inference phase, only the multidimensional feature vector of a single resting-state image is input. The trained MSR-Net can output a pixel-level perfusion defect distribution map, as well as the resting score (SRS), predicted total load score (SSS), and ischemia score (SDS) based on the AHA17 segment, thus achieving quantitative and segmental assessment of ischemia without the need for load testing.

[0075] Example 2

[0076] This embodiment is applicable to patients undergoing routine outpatient or inpatient coronary heart disease screening. It aims to utilize the intelligent calculation method of single-scan resting-state SPECT myocardial perfusion imaging in this application to obtain ischemic diagnostic results comparable to biphasic imaging without increasing the workload of examinations.

[0077] In the specific implementation process, the subject is injected with the radioactive imaging agent technetium-12 via the antecubital vein while at rest. 99m Labeled methoxyisobutylisocyanuric acid ( 99m Tc-MIBI), the injection dose was 740 MBq, and the sample was collected approximately 30 minutes after injection.

[0078] The acquisition was performed using a dual-probe SPECT / CT imaging system. The detector was equipped with a low-energy, high-resolution parallel-aperture collimator, with an acquisition range of 180° and an acquisition matrix of 64×64. The magnification was set according to the system calibration value, and the total acquisition time for a single session was approximately 15 minutes, which ensured that the signal-to-noise ratio of the myocardial region met the requirements for quantitative analysis.

[0079] The obtained raw data is processed through the processing flow of this application, including:

[0080] OSEM image reconstruction: The ordered subset expectation-maximization algorithm is used with 8 iterations and 6 subsets to obtain higher spatial resolution and counting accuracy.

[0081] Three-dimensional rigid body registration: Adjust the long axis direction of the myocardium based on the left ventricular center point parameters to make the image consistent with the standard space;

[0082] Adaptive bilateral filtering and multi-scale directional filtering: enhances myocardial boundary and mild perfusion differences while suppressing random noise, with 8 directions and 3 scales;

[0083] Myocardial segmentation and polar coordinate mapping: The left ventricular myocardial region was segmented using a pre-trained U-Net convolutional neural network and unfolded into a two-dimensional polar coordinate map according to the AHA17 segment model;

[0084] Multidimensional feature extraction: Injection intensity, texture, and morphological features are extracted using a parallel multi-scale convolutional network;

[0085] MSR-Net inference: Input multidimensional features into a biphasic labeled single-phase inference network, and output pixel-level perfusion defect distribution map and 17-segment SRS, SSS, and SDS scores.

[0086] In a clinical validation study of 50 patients undergoing routine screening, the Kappa value of the method in this embodiment was 0.91, which is consistent with the manual diagnosis of biphasic SPECT. The total examination time (from injection to result) was approximately 45 minutes, which is significantly shorter than the traditional biphasic procedure and has no risk of stress test.

[0087] Example 3: Emergency Rapid Testing Procedure

[0088] This embodiment is applicable to chest pain centers and cardiology emergency scenarios, and is used to rapidly assess myocardial ischemia in patients in the early stages of acute chest pain to support emergency interventional decisions.

[0089] Patients must meet the following criteria to be included: acute chest pain and no obvious ST segment elevation on electrocardiogram.

[0090] The acquisition device is a high-sensitivity SPECT system equipped with a CZT (cadmium zinc telluride) detector, which has high detection efficiency and can shorten the acquisition time. The patient is injected while at rest. 99m Tc-MIBI 555MBq was injected immediately after the procedure, with a 180° acquisition range, a 64×64 acquisition matrix, and a total acquisition time of 5 minutes.

[0091] Data processing optimization strategies include:

[0092] Fast OSEM reconstruction: 6 iterations, 4 subsets, reducing computational cost to improve speed;

[0093] GPU-accelerated filtering and segmentation: Adaptive bilateral filtering, 3D rigid body registration and U-Net segmentation are all executed in parallel on the GPU, with a total execution time of <5 seconds;

[0094] Fast inference: The optimized version of MSR-Net completes the entire process from multidimensional feature input to result output in 1 second.

[0095] This process ensures that the total time from imaging agent injection to the issuance of a diagnostic report does not exceed 20 minutes. When the model-predicted ischemia score (SDS) > 4, the patient can directly enter the emergency interventional pathway without additional stress testing.

[0096] In a prospective validation study of 30 emergency patients, the method showed a concordance rate of >88% with manual biphasic diagnosis, significantly shortening decision-making time while ensuring diagnostic accuracy.

[0097] The foregoing illustrative description of the present application and its embodiments is not restrictive and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The accompanying drawings are only one embodiment of the present application, and the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present application, such designs should fall within the scope of protection of this application. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. A method for classifying myocardial perfusion images based on a single resting state, characterized in that, include: S1, acquire three-dimensional voxel data (x, y, z) in a single resting state as SPECT myocardial perfusion images; S2, perform image reconstruction on the myocardial perfusion image, and then perform noise reduction and contrast enhancement on the reconstructed image to obtain the preprocessed myocardial perfusion image; S3, based on the preprocessed myocardial perfusion image, uses a pre-trained U-Net convolutional neural network for myocardial region segmentation; The segmented myocardial regions are mapped to a two-dimensional polar coordinate map P(r, θ) corresponding to the standard AHA17 segment model, where r is the radial coordinate and θ is the angular coordinate. S4. Based on the two-dimensional polar coordinate map P(r, θ) and the segmented myocardial region, perfusion intensity features, texture features and morphological features are extracted through parallel multi-scale convolution to form a multi-dimensional feature vector. S5. Construct a single-phase inference model MSR-Net based on biphasic annotation. Take multi-dimensional feature vectors as input, classify ischemia in a single resting state myocardial perfusion image, and output a pixel-level perfusion defect distribution map and a perfusion score based on myocardial segments. Among them, the single-phase inference model MSR-Net based on biphasic annotation uses the classification results corresponding to resting-load biphasic SPECT images as classification labels during the training phase, and performs label correction through label credibility based on segment consistency; during the inference phase, only the multidimensional feature vector of a single resting state needs to be input.

2. The myocardial perfusion image classification method based on a single resting state according to claim 1, characterized in that: S2, obtaining preprocessed myocardial perfusion images, including: The ordered subset expectation-maximization OSEM algorithm is used to reconstruct the three-dimensional voxel data (x, y, z) to obtain the initial reconstructed image; Based on the initial reconstructed image, the left ventricular center position (Cx, Cy, Cz) is obtained by using the minimum error pole fitting algorithm based on the short axis section; Based on the left ventricular center location (Cx, Cy, Cz), a three-dimensional rigid body registration algorithm is used to spatially align the initial reconstructed image to obtain the registered image; Based on the registered image, an adaptive bilateral filter is used to denoise the registered image to obtain a denoised image. Multi-scale directional filtering was applied to the denoised image to compensate for edge blurring in the SPECT image and enhance the local contrast of the myocardial boundary and areas with mild perfusion differences, resulting in a preprocessed myocardial perfusion image.

3. The myocardial perfusion image classification method based on a single resting state according to claim 2, characterized in that: S3 utilizes a pre-trained U-Net convolutional neural network for myocardial region segmentation, including: The preprocessed myocardial perfusion image is input into a pre-trained U-Net convolutional neural network, which outputs a three-dimensional segmentation mask of the myocardial region. Extract the myocardial contour edge based on the 3D segmentation mask, and calculate the gray-level centroid coordinates of the myocardial region based on the myocardial contour edge; Based on the gray-level centroid coordinates of the myocardial region and the center position of the left ventricle (Cx, Cy, Cz), the direction of the long axis of the myocardium is determined by least squares fitting. Based on the long axis direction of the myocardium and the edge of the myocardial contour, adjust the center point and starting angle of the polar coordinate unfolding so that the unfolded polar coordinate diagram is consistent with the spatial layout of the standard AHA17 segment model. Along the long axis of the myocardium, the myocardial region data were sampled in order from the apex to the base of the heart to obtain multiple short axis slices; For each short-axis slice, with the determined unfolding center point as the origin, the circular myocardial region is unfolded into a two-dimensional polar coordinate representation, where the radial coordinate r represents the distance from the endocardium to the epicardium, and the angular coordinate θ represents the position in the circumferential direction. Spatially integrate the polar coordinate representations of all short-axis slices to generate a complete two-dimensional polar coordinate map P(r, θ), and divide the polar coordinate map into 17 corresponding myocardial segment regions according to the definition of the standard AHA17 segment model.

4. The myocardial perfusion image classification method based on a single resting state according to claim 2, characterized in that: S4 constitutes a multidimensional feature vector, including: Construct a multi-scale convolutional network with three parallel branches; The first branch is used to perform a convolution operation on the two-dimensional polar coordinate map P(r, θ) to extract the perfusion intensity features; The second branch is used to perform a convolution operation on the two-dimensional polar coordinate map P(r, θ) to extract texture features; The third branch is used to perform a convolution operation on the three-dimensional segmentation mask of the myocardial region to extract morphological features; Infusion intensity features, texture features, and morphological features are concatenated through a feature fusion layer to obtain a multi-dimensional feature vector.

5. The myocardial perfusion image classification method based on a single resting state according to claim 4, characterized in that: The first branch is used to extract perfusion intensity features, including: Based on the two-dimensional polar coordinate plot P(r, θ), the following calculations were performed for each of the 17 myocardial segment regions: The average perfusion value is the mean value of perfusion intensity of all pixels within each segment. The standard deviation of the grouting intensity within each segment is taken as the standard deviation of the grouting value. The difference in average perfusion values ​​between adjacent segments is used as the perfusion gradient characteristic between segments; The average irrigation value, standard deviation of irrigation value, and irrigation gradient characteristics of the 17 segments are combined to obtain the irrigation intensity characteristics.

6. The myocardial perfusion image classification method based on a single resting state according to claim 4, characterized in that: Extracting texture features using the second branch includes: For each myocardial segment in the two-dimensional polar coordinate map P(r, θ), calculate the gray-level co-occurrence matrix GLCM; Based on the gray-level co-occurrence matrix (GLCM), the contrast features are calculated to reflect the degree of local gray-level changes. Based on the gray-level co-occurrence matrix (GLCM), energy features are calculated to reflect the uniformity of image texture. Based on the gray-level co-occurrence matrix (GLCM), homogeneity features are calculated for the similarity of local regions in the image. The contrast features, energy features, and homogeneity features of each stage are combined to obtain the texture features.

7. The myocardial perfusion image classification method based on a single resting state according to claim 4, characterized in that: Morphological features are extracted using the third branch, including: Based on the three-dimensional segmentation mask of the myocardial region, the distance from the endocardium to the epicardium is calculated on each short-axis slice as the myocardial thickness; The myocardial volume is calculated by summing the myocardial region area and interlayer spacing of each short-axis slice; Based on the contour information of the myocardial region, calculate the roundness and compactness indices of the myocardial shape; Morphological characteristics are obtained by combining indices such as myocardial thickness, myocardial volume, roundness, and compactness.

8. The method for classifying myocardial perfusion images based on a single resting state according to any one of claims 2 to 7, characterized in that: S5, Constructing the single-phase inference model MSR-Net based on biphasic annotation, including: A single-phase inference model MSR-Net is constructed. The MSR-Net model includes a feature encoder, a multi-task decoder, and a label optimization module. During the training phase, paired resting-load biphasic SPECT images were acquired, and the multidimensional feature vector of the resting image and the ischemia annotation information of the load image were extracted respectively; where the load image represents the myocardial perfusion image acquired under exercise load. By comparing images under load and at rest, pixel-level ischemia annotation masks and ischemia score labels for 17 myocardial segments are generated. The resting-state multidimensional feature vector is input into the feature encoder, and high-level features are extracted through multi-layer residual convolutional blocks; High-level features are input into a multi-task decoder for pixel-level segmentation and segment scoring. During training, the consistency index of feature distribution within each stage is calculated through the label optimization module. For stages where the consistency index is lower than the threshold, the corresponding label value is re-estimated using the cluster centers of similar samples to correct the label. A weighted loss function is used to optimize pixel-level segmentation and segment scoring tasks, and the model parameters are updated iteratively. During the inference phase, only a single resting state multidimensional feature vector needs to be input. The trained single-phase inference model MSR-Net outputs a pixel-level perfusion defect distribution map and perfusion score results for 17 myocardial segments. The perfusion score results include: resting score SRS, total score SSS, and ischemia score SDS.

9. The myocardial perfusion image classification method based on a single resting state according to claim 8, characterized in that: High-level features are input into a multi-task decoder for pixel-level segmentation and segment scoring, including: The multi-task decoder contains two parallel branches: The first branch generates a pixel-level perfusion defect probability map through a deconvolution layer; The second branch outputs perfusion scores for 17 myocardial segments through the fully connected layer.

10. A method for classifying myocardial perfusion images based on a single resting state, characterized in that, include: The data acquisition module acquires three-dimensional voxel data (x, y, z) in a single resting state as SPECT myocardial perfusion images; The image preprocessing module reconstructs the myocardial perfusion image, and performs noise reduction and contrast enhancement on the reconstructed image to obtain the preprocessed myocardial perfusion image. The myocardial segmentation and mapping module, based on the preprocessed myocardial perfusion image, uses a pre-trained U-Net convolutional neural network to segment the myocardial region and maps the segmented myocardial region to a two-dimensional polar coordinate map P(r, θ) corresponding to the standard AHA17 segment model, where r is the radial coordinate and θ is the angular coordinate. The feature extraction module extracts perfusion intensity features, texture features, and morphological features respectively based on the two-dimensional polar coordinate map P(r, θ) and the segmented myocardial region through parallel multi-scale convolution, forming a multi-dimensional feature vector; The classification and scoring module includes the MSR-Net single-phase inference model based on bidirectional annotation. It takes multi-dimensional feature vectors as input, performs ischemia classification on a single resting state myocardial perfusion image, and outputs a pixel-level perfusion defect distribution map and a perfusion score based on myocardial segments.

Images