A myocardial ischemia noninvasive detection model training method, detection method and system

Abstract

The application provides a myocardial ischemia noninvasive detection model training method, a detection method and a system. The detection model training is divided into two stages: in the first stage, a myocardial blood supply area segmentation network is trained, a multi-scale feature aggregation module is used to aggregate U-Net network encoder output, position coding is added, and then the position coding is input into a Transformer encoder, enhanced jump connection features after being separated and restored in size are input into a decoder; in the second stage, a myocardial ischemia prediction network is trained, a wavelet transform convolution is integrated into a three-dimensional dense connection layer as a backbone network, features are learned in a wavelet domain, a position embedding module is introduced to generate reweighting of spatial and channel position weights, a double pooling strategy of an aggregation pooling module is used to aggregate features, and whether a blood supply block is ischemic is detected through a classification head. Overall, under weak supervision, the feature reweighting and saliency guided aggregation improve the performance of the ischemia prediction network, so that the ischemia prediction network focuses on key ischemia information.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a training method, detection method and system for a non-invasive detection model of myocardial ischemia. Background Technology

[0002] Coronary artery disease (CAD) is a leading cause of death worldwide. Cardiovascular disease poses a significant threat to human health and is a global public health challenge. Early diagnosis and timely intervention of CAD are crucial for reducing mortality. Coronary CT angiography (CCTA), as a convenient and non-invasive diagnostic tool, exhibits excellent sensitivity (80-98%) and specificity (75-89%), with a negative predictive value as high as 100%, thus becoming the first-line imaging method for non-invasive screening and assessment of CAD in clinical practice.

[0003] Effective diagnosis of coronary artery disease (CAD) requires a comprehensive assessment of both coronary artery morphology and functional hemodynamic changes. CCTA primarily provides morphological data such as coronary anatomy and cannot assess flow-restricting lesions and associated myocardial ischemia at the functional level. Furthermore, there is no significant correlation between coronary artery stenosis and changes in myocardial blood flow. Clinical trials have shown that in moderate stenosis (50% to 70%), up to 65% of patients do not exhibit ischemia; in severe stenosis (71% to 90%), this proportion is approximately 20%; and even in patients with extremely severe stenosis (91% to 99%), 4% still do not show detectable myocardial ischemia. These differences indicate that morphological assessment alone is insufficient to guide treatment decisions; a functional assessment focusing on flow-restricting lesions and corresponding ischemic areas is essential for diagnosing CAD.

[0004] Currently, the invasive measurement of fractional flow reserve (FFR) is the clinical gold standard for assessing the functional significance of coronary artery stenosis. However, FFR measurement requires invasive cardiac catheterization, which is complex, costly, and carries certain risks, limiting its widespread application. To overcome these limitations, several non-invasive functional assessment techniques based on computer-aided diagnosis (CAD) and coronary artery-centered and myocardial-centered approaches have emerged.

[0005] Coronary artery-centric approaches aim to directly predict coronary artery flow rate (FFR) or assess the functional significance of stenosis using CCTA images. These include CT-FFR techniques based on computational fluid dynamics (CFD) and CT-FFR techniques based on artificial intelligence / deep learning (AI / DL). Traditional CFD-based CT-FFR simulates coronary hemodynamics by combining CCTA-derived anatomical data with CFD modeling. While its diagnostic performance is close to invasive FFR, its computational demands are enormous, typically requiring hours of operation on high-performance servers, limiting its clinical applicability. Recently, the development of deep learning has provided a more efficient alternative to CFD. DL models can directly learn the complex nonlinear relationship between coronary artery anatomy and FFR values ​​from CCTA images and / or clinical data. However, both CFD methods and some DL methods generally still require accurate coronary artery tree segmentation and geometric modeling.

[0006] Given the complexity of directly analyzing coronary arteries, some research approaches have shifted to analyzing myocardial regions directly affected by blood flow changes, forming a "myocardial-centric" approach. This method attempts to determine the presence of myocardial ischemia by assessing the imaging features of the myocardium, thereby diagnosing functionally significant CAD and avoiding difficult vessel segmentation. However, using conventional resting-state CCTA images for myocardial ischemia assessment also faces significant challenges: CT's inherent contrast resolution for soft tissues such as the myocardium is far lower than that of MRI, especially in CCTA, where the myocardium itself is not enhanced by contrast agents, resulting in blurred boundaries and subtle density differences between it and surrounding tissues (such as the right ventricular cavity). This makes it extremely difficult to identify subtle texture or density changes caused by ischemia. Traditional machine learning-based methods, such as relying on hand-designed radiomics features combined with classifiers like random forests and SVMs, have made some progress, but the feature extraction process may not fully capture complex spatial and textural information, and the models have limited interpretability and generalization ability, making it difficult to fully utilize the deep semantic and contextual information of the images to infer the ischemic state. Furthermore, most existing myocardial lateralization methods aim to achieve patient-level prediction (i.e., determining whether the patient has significant lesions or ischemia overall), while clinical practice requires ischemia information at the regional level (specific vascular supply areas) to guide precise intervention, which falls short of clinical needs. Therefore, non-invasive prediction of myocardial ischemia based on conventional resting-state CCTA images faces significant technical challenges in achieving accurate regional-level assessment, overcoming low image contrast, and effectively utilizing weakly supervised information. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide a training method, detection method and system for a non-invasive detection model of myocardial ischemia, which solves the problems of low image contrast and blurred boundaries in existing myocardial ischemia assessment based on CCTA images, which make it difficult to segment the myocardial blood supply area, and the difficulty in regional prediction caused by low efficiency in utilizing weak supervision information.

[0008] One aspect of the present invention provides a method for training a non-invasive detection model of myocardial ischemia based on CCTA images, the method comprising the following steps:

[0009] For CCTA images from multiple user samples, the location mask of multiple blood supply blocks including the left anterior descending artery, right coronary artery and left circumflex artery in the left ventricle, which is labeled by the AHA 17-segment standard, is used as the first label, and whether each blood supply block is ischemic is added as the second label to construct a training sample set.

[0010] The myocardial blood supply area segmentation network is trained using the training sample set and the CCTA image is used as input. The myocardial blood supply area segmentation network is obtained by embedding a multi-scale feature aggregation module into the skip connection path of the U-Net network. After aggregating the outputs of the multi-level encoders in the U-Net network, the multi-scale feature aggregation module adds position encoding and inputs it to the Transformer encoder. The output of the Transformer encoder is separated and restored to its original size to obtain enhanced skip connection features, which are then input to the decoder of the corresponding level in the U-Net network. The parameters of the myocardial blood supply area segmentation network are updated based on the predicted mask of each blood supply block position and the deviation of the first label.

[0011] The myocardial ischemia prediction network is trained using the training sample set. Using the center of the mask corresponding to each blood supply block in the user samples as a reference, a three-dimensional image patch of a set size is extracted from the CCTA image as input. The myocardial ischemia prediction network decomposes the input features into eight sub-band features using a wavelet transform convolutional unit. These sub-band features are then grouped and convolved in the wavelet domain and reconstructed back to the original spatial domain. The reconstructed features are then processed by a location embedding module, which introduces spatial location weights and channel location weights to output a spatially reweighted feature map and a channel-reweighted feature map. A global group max pooling is performed on the spatially reweighted feature map, and a regional group max pooling is performed on the channel-reweighted feature map. These are then summed to obtain an aggregated feature vector. This aggregated feature vector is then processed by a classification head to output a prediction of whether the corresponding blood supply block is ischemic. The loss is constructed based on the deviation from the second label to update the parameters of the myocardial ischemia prediction network.

[0012] The updated myocardial blood supply segmentation network and the myocardial ischemia prediction network are combined to form a non-invasive myocardial ischemia detection model.

[0013] In some embodiments, adding whether each blood supply block is ischemic as a second label includes:

[0014] The second tag is added to detect whether each blood supply block of the user sample is ischemic using invasive-measured fractional flow reserve, immediate wave free ratio, resting full-cycle ratio, diastolic blood pressure ratio, coronary flow reserve and / or microcirculation resistance index.

[0015] In some embodiments, the U-Net network in the myocardial blood supply segmentation network is an nnU-Net network, including multiple encoders for downsampling and corresponding decoders for upsampling. The multi-scale feature aggregation module applies independent 3D convolution to the output of each encoder for fragment embedding, then performs shaping and splicing for aggregation and adds fixed or learnable positional encodings. These are then handed over to the Transformer encoder to use a multi-head self-attention mechanism to capture long-range dependencies within the sequence and interaction relationships between features. The output of the Transformer encoder is subjected to dimensionality separation and reshaped back to the original scale of the encoder output, then upsampled back to the original spatial size of the skip connections and input to the corresponding decoder.

[0016] In some embodiments, in the wavelet transform convolutional unit, the k-channel output after the input passes through the first batch of normalization layers, the first activation function layer, and the 1*1*1 convolutional layer undergoes 3D Haar wavelet transform, and then passes through the second batch of normalization layers, the second activation function layer, and the grouped convolutional layer to apply subband mixing or channel mixing processing to obtain an 8k-channel output. This output is then reconstructed back to the original spatial domain by performing inverse wavelet transform, and finally output by the Dropout layer.

[0017] In some embodiments, the method further includes:

[0018] The parameters of the myocardial blood supply region segmentation network are updated based on the binary cross-entropy loss and Dice loss between the predicted mask and the first label. The total loss is calculated as follows:

[0019] ;

[0020] ;

[0021] in, This represents the total loss from updating the parameters of the myocardial blood supply region segmentation network. This refers to the prediction mask. This refers to the first label. This represents the binary cross-entropy loss. This represents the Dice loss. This is a smoothing factor.

[0022] In some embodiments, the total loss calculation formula for updating the parameters of the myocardial ischemia prediction network by the method is as follows:

[0023] ;

[0024] in, This represents the total loss in updating the parameters of the myocardial ischemia prediction network. This represents the predicted probability of the true category of the second label. It is the focusing parameter. It is a category balance factor;

[0025] The method applies L2 weight decay for regularization during the parameter update process of the myocardial ischemia prediction network.

[0026] On the other hand, the present invention also provides a non-invasive detection method for myocardial ischemia based on CCTA images, the method comprising the following steps:

[0027] Obtain the CCTA image to be analyzed;

[0028] The CCTA image is input into the non-invasive myocardial ischemia detection model in the above-mentioned training method for a non-invasive myocardial ischemia detection model based on CCTA images. The myocardial blood supply area segmentation network in the non-invasive myocardial ischemia detection model identifies multiple blood supply block location masks containing the left anterior descending artery, right coronary artery, and left circumflex artery in the left ventricle. Using the center of the blood supply block location mask as a reference, a three-dimensional image block of a set size is extracted from the CCTA image and input into the myocardial ischemia prediction network in the non-invasive myocardial ischemia detection model, and outputs the identification result of whether myocardial ischemia exists in each blood supply block.

[0029] On the other hand, the present invention also provides a non-invasive detection system for myocardial ischemia based on CCTA images, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the above method.

[0030] On the other hand, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0031] On the other hand, the present invention also provides a computer program product, including a computer program / instructions, characterized in that the computer program / instructions, when executed by a processor, implement the steps of the above-described method.

[0032] The non-invasive detection model training method, detection method, and system for myocardial ischemia described in this invention, for CCTA images of user samples, uses at least three major coronary artery supply regions in the CCTA image based on the AHA 17-segment standard and labels whether the corresponding supply blocks are ischemic as tags to construct a training sample set. In the first stage, a myocardial supply region segmentation network is trained. A multi-scale feature aggregation module is introduced to aggregate the outputs of multi-level encoders in the U-Net network, adds positional encoding, and inputs it to the Transformer encoder. The output of the Transformer encoder is separated and restored to its original size to obtain enhanced skip connection features, which are then input to the corresponding level decoder in the U-Net network. The second stage trains the myocardial ischemia prediction network by introducing wavelet transform convolution into a three-dimensional densely connected layer as the backbone network. Utilizing the three-dimensional discrete Haar wavelet transform and its inverse transform, combined with grouped convolution, feature learning in the wavelet domain is achieved. Subsequently, a location embedding module is introduced, generating spatial and channel location weights through parallel branches for reweighting. The global and regional group max pooling strategies of the aggregation pooling module are used to selectively aggregate features, and a classification head is used to detect whether the blood supply block is ischemic. This invention, under weak supervision without precise lesion annotation, effectively improves the performance of the ischemia prediction network through feature reweighting and saliency-guided aggregation, enabling it to better focus on key ischemia-related information.

[0033] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0034] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0035] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0036] Figure 1 This is a schematic flowchart of a non-invasive detection model training method for myocardial ischemia according to an embodiment of the present invention.

[0037] Figure 2 This is a logical schematic diagram of a non-invasive detection method for myocardial ischemia based on CCTA images according to another embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of the myocardial blood supply region segmentation network in the non-invasive myocardial ischemia detection model training method according to an embodiment of the present invention.

[0039] Figure 4 for Figure 3 Structure diagram of the multi-scale feature aggregation module in the myocardial blood supply region segmentation network.

[0040] Figure 5 This is a schematic diagram of the myocardial ischemia prediction network in the non-invasive myocardial ischemia detection model training method according to an embodiment of the present invention.

[0041] Figure 6 This is a schematic diagram based on the American Heart Association's 17-segment model.

[0042] Figure 7 This is a schematic diagram of the blood supply areas of the left anterior descending artery, right coronary artery, and left circumflex artery within the left ventricle (LV) myocardium, based on standardized labels of the American Heart Association 17-segment model.

[0043] Figure 8 This is a flowchart illustrating the process of generating labels based on the American Heart Association's 17-segment model.

[0044] Figure 9 This is a segmentation effect diagram of the myocardial blood supply area segmentation network in the training method for non-invasive detection model of myocardial ischemia according to an embodiment of the present invention. Detailed Implementation

[0045] 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 embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0046] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0047] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0048] In the clinical practice of cardiovascular disease, coronary CT angiography (CCTA) is a convenient and non-invasive cardiac diagnostic tool, primarily used to provide morphological data such as coronary artery anatomy. However, it cannot assess flow-restricting lesions and related myocardial ischemia at the functional level. Existing CCTA-based myocardial ischemia assessment techniques are mainly divided into two major approaches: coronary artery-centered and myocardial-centered. However, these techniques still have shortcomings in achieving accurate regional-level assessment, overcoming low image contrast, and effectively utilizing weakly supervised information. For example, CT-FFR technology based on computational fluid dynamics and some deep learning (DL) methods require precise coronary artery tree segmentation and geometric modeling, a complex and challenging process. In contrast, myocardial-centered methods struggle to identify subtle texture or density changes in conventional resting-state CCTA images due to blurred boundaries and weak density differences between the myocardium and surrounding tissues. Furthermore, traditional machine learning-based feature extraction methods may not fully capture complex spatial and textural information, and their model interpretability and generalization capabilities are limited. Therefore, there is an urgent need for a new technology that can accurately and non-invasively predict ischemia at the myocardial blood supply level using conventional resting CCTA images under weak supervision, in order to overcome the limitations of existing technologies and meet the clinical need for precise intervention.

[0049] This invention provides a method for training a non-invasive detection model of myocardial ischemia based on CCTA images, such as... Figure 1 As shown, the method includes the following steps S101~S104:

[0050] Step S101: For CCTA images of multiple user samples, use the AHA 17-segment standard label to label the location masks of multiple blood supply blocks in the left anterior descending artery, right coronary artery and left circumflex artery in the left ventricle as the first label, and add whether each blood supply block is ischemic as the second label to construct a training sample set.

[0051] Step S102: Train the myocardial blood supply area segmentation network using the training sample set, with CCTA images as input, such as... Figure 3 As shown, the myocardial blood supply region segmentation network is obtained by embedding a multi-scale feature aggregation module into the skip connection path of the U-Net network; as Figure 4 As shown, the multi-scale feature aggregation module aggregates the outputs of multi-level encoders in the U-Net network, adds positional encoding, and inputs it into the Transformer encoder. The output of the Transformer encoder is separated and restored to its original size to obtain enhanced skip connection features, which are then input into the decoder of the corresponding level in the U-Net network. Based on the deviation between the predicted mask of each blood supply block location and the first label of the myocardial blood supply segmentation network, the parameters of the myocardial blood supply segmentation network are updated.

[0052] Step S103: Train the myocardial ischemia prediction network using the training sample set. Using the center of the mask corresponding to each blood supply block in the user samples as a reference, extract a 3D image patch of a set size from the CCTA image as input, such as... Figure 5 As shown, the myocardial ischemia prediction network decomposes the input features into 8 sub-band features using a wavelet transform convolutional unit. After grouping and convolving the sub-band features in the wavelet domain, they are reconstructed back to the original spatial domain. The reconstructed features are then processed by a location embedding module to introduce spatial location weights and channel location weights, outputting spatially reweighted feature maps and channel-reweighted feature maps. Based on an aggregation pooling module, global group max pooling is performed on the spatially reweighted feature map, and regional group max pooling is performed on the channel-reweighted feature map. The aggregated feature vector is then summed to obtain an aggregated feature vector. The aggregated feature vector is processed by a classification head to output a prediction of whether the corresponding blood supply block is ischemic. The loss is constructed by combining the deviation from the second label to update the parameters of the myocardial ischemia prediction network.

[0053] Step S104: Combine the updated myocardial blood supply segmentation network and myocardial ischemia prediction network into a non-invasive myocardial ischemia detection model.

[0054] In step S101, the objective of this invention is to identify and segment blood supply blocks from conventional resting CCTA images and to analyze at the block level whether ischemia exists in the blood supply blocks. To achieve accurate segmentation, this invention employs a standardized method based on the 17-segment model recommended by the American Heart Association (AHA) to define the blood supply regions of the three main coronary arteries (LAD, RCA, LCX) of the left ventricular (LV) myocardium, used to establish segmentation labels.

[0055] The entire tag generation process is as follows: Figure 6 , 7 As shown in Figure 8, the key steps include:

[0056] 1. Acquisition and Optimization of Left Ventricular Myocardial Mask: First, using mature medical image post-processing software, such as Siemens Syngovia, a preliminary left ventricular (LV) myocardial mask was automatically segmented. Subsequently, the automatic segmentation results were meticulously verified and adjusted to correct potential segmentation errors and ensure the accuracy of the mask.

[0057] 2. Short-Axis Image Reconstruction: Since the definition and analysis of the AHA 17-segment model are typically based on the short-axis view of the heart (parallel to the long axis of the left ventricle), while routine clinical scans and initial annotations are mostly performed in the axial view, image reconstruction is necessary. This embodiment utilizes the open-source software ITK-SNAP (version 4.2 and above supports multi-planar reconstruction) to allow radiologists to interactively determine the standard cardiac short-axis orientation relative to the original axial image and record the corresponding spatial orientation information. This orientation information essentially defines an affine transformation matrix. Subsequently, a script written using the Python library (SimpleITK) applies affine transformation and resampling to the original axial CCTA image and its corresponding LV myocardial mask based on the saved orientation information, generating an image and mask in the standard short-axis view. During this process, image boundaries are appropriately adjusted or padding is applied to ensure that the rotated cardiac tissue is completely contained within the image domain.

[0058] 3. Automatic AHA 17-Segment Division: The acquired short-axis LV myocardial mask is automatically divided according to the AHA 17-segment standard. This process is implemented by a custom algorithm: First, the apical region is identified and separated, i.e., the most distal myocardial portion excluding the left ventricular chamber, defined as segment 17. The remaining myocardial body (whose cross-section is typically annular, surrounding the left ventricular chamber) is automatically divided into three equal layers along its long axis based on spatial location: basal, mid-cavity, and apical (the portion close to the true apex but still containing the heart chamber). On representative slices of each layer (specified by experts at the start or end slices of the layer; the algorithm interpolates the calculation of intermediate slices), two attachment points of the right ventricle and left ventricular septum are determined. Based on these two key points, the myocardial ring is divided into standard segments using angular division: the basal and middle portions are each divided into 6 segments (segments 1-6 and 7-12, respectively), and the apex (excluding the top part of segment 17) is divided into 4 segments (segments 13-16). For example, the area between the two attachment points at the basal level is equally divided angularly into the anterior septal segment (segment 2) and the inferior septal segment (segment 3), while the remaining portion is equally divided angularly into the anterior segment (segment 1), the inferior segment (segment 4), the inferior lateral segment (segment 5), and the anterior lateral segment (segment 6). The middle and apical portions are divided using a similar rule (the lateral walls of the apex are usually merged into one segment 16). Specific division details are as follows... Figure 6 As shown. This process is automatically completed by an algorithm based on mask data.

[0059] 4. Blood Supply Area Mapping: Based on the standard correspondence between AHA segments and coronary blood supply areas, the 17 pre-divided segments are merged to generate masks for three target blood supply areas: LAD blood supply area mapping segments 1, 2, 7, 8, 13, 14, and 17; RCA blood supply area mapping segments 3, 4, 9, 10, and 15; and LCX blood supply area mapping segments 5, 6, 11, 12, and 16.

[0060] 5. Inverse Coordinate System Transformation: Finally, using the inverse transformation of the orientation information recorded in step 2, the three blood supply region masks generated under the short axis are transformed back to the original axis coordinate system to obtain axis labels for training and evaluation of the segmentation network.

[0061] While some research focuses on developing fully automated workflows that include automatic orientation calibration and keypoint detection, this embodiment employs the method described above, combining standardized models, automated algorithms, and necessary expert interaction (optimizing the initial mask and specifying keypoint orientations) to generate high-quality, clinically compliant training labels. This approach is more robust and reliable than methods that rely on complex and potentially unstable coronary artery tree segmentation to define the blood supply region. Furthermore, the end-to-end segmentation model trained in this embodiment can directly predict on axial images more commonly used in clinical practice, avoiding the need for short-axial reconstruction during the inference phase, thereby improving the clinical applicability and efficiency of the method.

[0062] Furthermore, the fractional flow reserve, immediate wave free ratio, resting full-cycle ratio, diastolic blood pressure ratio, coronary flow reserve and / or microcirculation resistance index based on invasive measurements were used to detect ischemia in each blood supply block of the user sample and a second label was added.

[0063] Fractional flow reserve (FFR), an invasive measurement, is the gold standard for assessing the functional significance of coronary artery stenosis. It is defined as the ratio of the maximum blood flow in the presence of stenosis to the maximum blood flow achievable assuming no stenosis. It is calculated by measuring the ratio of blood pressure (Pd) distal to the stenosis to aortic blood pressure (Pa), i.e., FFR = Pd / Pa. Measurements are performed under maximal coronary congestion conditions, typically via intravenous or intracoronary injection of drugs such as adenosine, to ensure complete dilation of downstream resistance vessels and accurately assess the impact of stenosis on blood flow. Lesions with an FFR ≤0.80 are generally considered functionally significant lesions causing myocardial ischemia and may require revascularization treatments such as percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG).

[0064] Immediate Free Ratio (iFR) does not require the use of vasodilators such as adenosine. It calculates the ratio of distal coronary artery pressure to aortic pressure by measuring the pressure during the free phase of the diastolic wave in the coronary arteries. The recommended cutoff value is 0.89, and it shows no significant difference in clinical outcomes compared to FFR. Resting Whole Cycle Ratio (RFR) also does not require the use of drugs such as adenosine. It calculates the minimum ratio of distal coronary artery pressure to aortic pressure over five complete cardiac cycles. Diastolic Pressure Ratio (PR) is also an adenosine-free functional assessment method, calculating the average pressure ratio at a specific time point or throughout diastole. Coronary Flow Reserve (CFR) is calculated by measuring the ratio of blood flow at rest and under maximal congestion using either the thermodilution method or the Doppler flow velocity method. The cutoff value for normal values ​​is 2.0 or 2.5, depending on the method used. The Microcirculation Resistance Index (IMR) is primarily used to assess coronary microcirculation function and is calculated by measuring pressure and Doppler flow velocity within the coronary arteries.

[0065] In step S102, as the first stage of the present invention, the aim is to accurately segment the left ventricular (LV) myocardium and the blood supply areas of its major coronary arteries (left anterior descending artery LAD, right coronary artery RCA, and left circumflex artery LCX) from conventional resting CCTA images. Considering the low contrast and blurred boundaries between the LV myocardium and adjacent chambers and tissues in CCTA images, making accurate segmentation extremely challenging, this stage employs and improves upon the high-performance three-dimensional medical image segmentation framework nnU-Net.

[0066] In some embodiments, the nnU-Net network includes multiple encoders for downsampling and corresponding decoders for upsampling. The multi-scale feature aggregation module applies independent 3D convolution to the output of each encoder for fragment embedding, then performs reshaping and splicing for aggregation and adds fixed or learnable positional encodings. These are then handed over to the Transformer encoder to use a multi-head self-attention mechanism to capture long-range dependencies within the sequence and interactions between features. The output of the Transformer encoder is subjected to dimensionality separation and reshaped back to the scale of the original encoder output, then upsampled back to the spatial size of the original skip connections and input to the corresponding decoder.

[0067] Specifically, the myocardial blood supply region segmentation network in this stage is based on the nnU-Net architecture, which can use a residual encoder version under 3D full resolution configuration. A novel Multi-Scale Feature Aggregation (MSFA) module is embedded in its standard skip connections. Traditional skip connections directly fuse shallow features from the encoder with corresponding features from the decoder, which may not effectively utilize multi-scale information and global context. The MSFA module of this invention explicitly aggregates multi-scale features from different levels of the encoder and introduces a Transformer mechanism to optimize and interact with these cross-scale features. This design enables the network to better model global anatomical context information and capture complex dependencies between features at different scales. By enhancing the understanding of the overall image structure and effectively integrating information at different levels of detail, the MSFA module significantly improves the segmentation accuracy of myocardial blood supply regions in low-contrast, blurred-boundary CCTA images. Specifically, this myocardial blood supply region segmentation model uses masks of the three main blood supply regions—LAD, RCA, and LCX—obtained by dividing and merging based on the AHA 17-segment standard as training labels.

[0068] In some embodiments, the method further includes: updating the parameters of the myocardial blood supply segmentation network based on the binary cross-entropy loss and Dice loss between the predicted mask and the first label, wherein the total loss is calculated as follows:

[0069] ;

[0070] ;

[0071] in, This represents the total loss due to the updating of network parameters for the myocardial blood supply region segmentation. Indicates the prediction mask, Indicates the first tag, This represents the binary cross-entropy loss. This indicates Dice's loss. This is a smoothing factor.

[0072] The final segmentation effect of the myocardial blood supply area segmentation network is as follows: Figure 9 As shown.

[0073] In step S103, the second stage of the present invention can use the location masks of each blood supply block labeled with the first label in the training sample set and the original CCTA image blocks as input, or it can receive the prediction masks of each myocardial blood supply block output by the myocardial blood supply area segmentation network of the first stage and the original CCTA image blocks as input, and predict whether functional ischemia exists in each major blood supply area based on these. In order to improve the prediction accuracy and overcome the challenge of CCTA images under weak supervision conditions where pixel-level accurate ischemia annotation is lacking, this stage design includes a backbone network and an attention mechanism module.

[0074] Specifically, this stage constructs a myocardial ischemia prediction network, whose input consists of concatenating the corresponding CCTA image patch and the predicted blood supply region mask along the channel dimension. The core of the network is a backbone network WTDenseNet3D developed in this application. WTDenseNet3D creatively integrates the Wavelet Transform Convolution (WTC) module into a 3D dense layer. The WTC unit utilizes the 3D discrete Haar wavelet transform and its inverse transform, combined with grouped convolution, to achieve feature learning in the wavelet domain. This approach can simultaneously capture information from images at different resolutions and effectively expand the network's receptive field, thereby better extracting and recognizing subtle textures and spatial patterns related to myocardial ischemia. Specific implementations of the WTC unit may include grouped convolution schemes such as sub-band mixing or channel mixing.

[0075] In some embodiments, in the wavelet transform convolutional unit serving as the backbone network, the k-channel output after the input passes through the first batch of normalization layers, the first activation function layer, and the 1*1*1 convolutional layer undergoes 3D Haar wavelet transform, and then passes through the second batch of normalization layers, the second activation function layer, and the grouped convolutional layer to apply subband mixing or channel mixing processing to obtain an 8k-channel output. This output is then reconstructed back to the original spatial domain through inverse wavelet transform and finally output by the Dropout layer.

[0076] To further optimize the network's prediction performance under weak supervision and guide the network to focus on the regions and channels most relevant to ischemia when precise lesion location information is unavailable, this stage introduces a Position Embedding (PE) module and an Aggregation Pooling (AP) module after the WTDenseNet3D backbone network. The PE module generates spatial and channel position weights through parallel branches, reweighting the features extracted by the backbone network to highlight features of important regions and channels. The AP module combines Global Group Pooling (GGP) and Region Group Pooling (RGP) strategies to selectively aggregate the features output by the PE module, further enhancing the network's ability to focus on the most indicative features.

[0077] Finally, the prediction network outputs the probability of functional ischemia in each blood supply area (binary classification) through a classification head, and uses invasive FFR measurements as the gold standard label for training. The entire second-stage network can be trained end-to-end.

[0078] In some embodiments, the total loss calculation formula for updating the parameters of the myocardial ischemia prediction network by the method is as follows:

[0079] ;

[0080] in, This represents the total loss from updating the parameters of the myocardial ischemia prediction network. This represents the predicted probability of the true category of the second label. It is the focusing parameter. It is a category balance factor.

[0081] The method describes the application of L2 weight decay for regularization during the parameter update process of the myocardial ischemia prediction network.

[0082] In step S104, the trained myocardial blood supply area segmentation network and myocardial ischemia prediction network are combined to obtain a non-invasive myocardial ischemia detection model that can segment myocardial blood supply blocks in CCTA images and identify whether myocardial ischemia exists in each block.

[0083] On the other hand, the present invention also provides a non-invasive detection method for myocardial ischemia based on CCTA images, such as... Figure 2 As shown, the method includes the following steps S201~S202:

[0084] Step S201: Obtain the CCTA image to be analyzed;

[0085] Step S202: Input the CCTA image into the non-invasive myocardial ischemia detection model in the training method of the non-invasive myocardial ischemia detection model based on CCTA image in steps S101 to S104 above. The myocardial blood supply area segmentation network in the non-invasive myocardial ischemia detection model identifies multiple blood supply block location masks containing the left anterior descending artery, right coronary artery and left circumflex artery in the left ventricle. Using the center of the blood supply block location mask as a reference, extract a three-dimensional image block of a set size from the CCTA image and input it into the myocardial ischemia prediction network in the non-invasive myocardial ischemia detection model and output the identification result of whether myocardial ischemia exists in each blood supply block.

[0086] On the other hand, the present invention also provides a non-invasive detection system for myocardial ischemia based on CCTA images, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the above method.

[0087] On the other hand, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0088] On the other hand, the present invention also provides a computer program product, including a computer program / instructions, characterized in that the computer program / instructions, when executed by a processor, implement the steps of the above-described method.

[0089] The present invention will now be described with reference to a specific embodiment:

[0090] This embodiment proposes a two-stage deep learning pipeline network based on conventional resting-state CCTA images, which can achieve end-to-end myocardial ischemia prediction at the myocardial blood supply region level. The first stage is a myocardial blood supply region segmentation network based on multi-scale feature aggregation, and the second stage is a region ischemia prediction network based on multi-resolution feature extraction and saliency-guided analysis. Specifically, it includes the following steps:

[0091] S1. Data Acquisition: Two batches of data were collected from different centers, including 86 patients. All of these patients underwent CCTA and FFR examinations.

[0092] S2. Dataset Preprocessing: To train an end-to-end deep learning model capable of directly segmenting the myocardial supply territory on CCTA images, this embodiment first needs to generate accurate segmentation labels. This embodiment uses a standardized method based on the 17-segment model recommended by the American Heart Association (AHA) to define the three main coronary artery (LAD, RCA, LCX) supply territories of the left ventricular (LV) myocardium.

[0093] A blood supply region segmentation dataset was constructed: each patient included CCTA scans in both axial and short-axis orientations, resulting in 172 independent 3D image volumes. These volume data, along with their corresponding blood supply region masks generated using the label generation process described above (serving as segmentation labels), constituted the dataset used to train and evaluate the blood supply region segmentation network. To ensure the reliability of the model's generalization ability evaluation, the dataset was randomly divided into training and test sets at approximately a 4:1 ratio, ultimately yielding 137 training samples and 35 test samples.

[0094] A myocardial ischemia prediction dataset was constructed: This dataset aims to train and evaluate a model capable of predicting ischemia at the region-level. The construction process is as follows: First, the previously trained region segmentation model is used to infer the predicted blood supply region masks from all 172 CCTA image volumes. Given that the heart is primarily supplied by three coronary arteries (LCX, RCA, LAD), corresponding to three main blood supply regions, theoretically, each CCTA volume can generate three region samples. For each predicted blood supply region mask, a 255×255×255 cubic pixel three-dimensional image patch is extracted based on its center. Zero-padding is used for cropped regions exceeding the original image boundaries. This step initially generates 172×3=516 image patches and their predicted masks corresponding to specific blood supply regions. However, since not every patient underwent invasive FFR examination (as a reference standard for ischemia) for every coronary artery, this embodiment excluded image patches lacking corresponding FFR labels (i.e., those whose ischemia could not be determined). Additionally, image patches from two patients with poorly segmented blood supply areas due to severe hard artifacts were excluded. After screening, 432 three-dimensional image patch data points with ischemia labels (corresponding to 216 vessels) from 84 patients were ultimately retained. These data were used to train and test a regional ischemia prediction network, whose inputs are image patches and corresponding predicted blood supply area masks, and whose output is a binary ischemia label (0 or 1). To prevent information leakage between the training and test sets due to data from the same patient, the ischemia prediction dataset was also strictly divided according to patient level. Maintaining an approximately 4:1 ratio, 342 image patches from 68 patients were ultimately assigned to the training set, and 90 image patches from 16 patients were assigned to the test set.

[0095] S3. Construct a myocardial blood supply region segmentation network based on a multi-scale feature aggregation module.

[0096] To achieve subsequent myocardial ischemia prediction based on blood supply regions, this embodiment first requires a model capable of accurately segmenting the three main coronary artery supply regions (LAD, RCA, LCX) of the left ventricular myocardium. Considering the three-dimensional spatial characteristics of blood supply region definition, a single 2D slice often cannot fully reflect the regional attribution based on myocardial anatomical location (e.g., distinguishing whether a slice in a short-axis image belongs to the base or the middle, or determining the boundary of the blood supply region on an axial image). Figure 6 In this embodiment, a 3D segmentation network is constructed.

[0097] This embodiment uses the nnU-Net framework, which performs exceptionally well in the field of medical image segmentation, as its basic architecture. Specifically, it adopts the residual encoder version under its 3D full-resolution configuration (corresponding to nnUNetResEncUNetM or a similar configuration in nnU-Net). This network itself has powerful medical image feature extraction capabilities.

[0098] However, myocardial segmentation in CCTA images presents unique challenges. Unlike MRI's high resolution for soft tissues, CT primarily relies on imaging of tissue density differences. In CCTA, contrast agents mainly enhance vascular cavities, while the myocardium itself is not enhanced or only minimally enhanced. This results in weak density differences between the left ventricular myocardium and surrounding structures not filled with contrast agent (such as the right ventricular cavity), often leading to blurred boundaries, especially in areas like the interventricular septum, which can be visually indistinguishable. This makes accurate segmentation not only require local features but also heavily depends on an understanding of the overall three-dimensional anatomy of the heart and the utilization of spatial contextual information. While standard U-Net captures multi-scale features through encoder-decoder and skip connections, the simple feature transfer mechanism in its skip connections may not be sufficient to fully integrate and utilize cross-scale contextual information to effectively address the challenges of low contrast and blurred boundaries in CCTA images.

[0099] To solve the above problems, such as Figure 3 and 4 As shown, this embodiment proposes and introduces a Multi-Scale Feature Aggregation (MSFA) module, embedding it into the skip connection path of nnU-Net. The core idea of ​​this module is to explicitly aggregate multi-scale features from different levels of the encoder and leverage the powerful global dependency modeling capabilities of the Transformer to interact and refine cross-scale information, ultimately passing the enhanced features to the corresponding level of the decoder. This embodiment aims to improve the network's comprehensive understanding of global spatial structure and multi-scale features, thereby achieving more robust and accurate myocardial (i.e., blood supply area) segmentation in low-contrast regions of CCTA images.

[0100] Similar to nnU-Net, the MSFA module in this embodiment also has a dynamic network structure, which can be freely configured according to the number of stages in nnU-Net. The MSFA module receives skip connection feature maps from the n stages output of the nnU-Net encoder. Taking nnUNetResEncUNetM as an example, it has 6 downsampling stages (including the bottleneck layer), so n=6, where Let i be the skip connection feature map output by the encoder in the i-th stage. Let be the number of channels, height, width, and depth of this 3D feature map. For each Applying independent 3D convolutions for patch embedding reduces spatial dimensions. Unified adjustment of spatial dimensions to the bottleneck layer (nth stage) And adjust the number of channels to the embedding dimension. In the settings of this embodiment, .

[0101] ;

[0102] All embedded feature maps By concatenating along the channel dimension, an aggregated feature map is obtained. Then flatten it into a sequence , where the sequence length Feature Dimension Then came Adding standard sinusoidal positional encoding enables the Transformer to perceive feature information from different positions within the sequence and from different original scales. The sequence with added positional encoding is then input into a database containing... In the Transformer encoder of the layer. This encoder uses a multi-head self-attention mechanism (number of heads...). The embedded dimension (D) is used to capture long-range dependencies within the sequence and complex interactions between features, outputting the processed sequence. Next, the aggregated multi-scale features need to be separated and reshaped. Split along feature dimension D into n parts, each part corresponding to an original stage (dimension D). Then, the sequence length L is reshaped back to the spatial dimension. ,get ,in, Finally, trilinear interpolation was used to... Upsample back to the original skip connection spatial dimensions The enhanced skip connection features are obtained. ,in, ,These The input will be fed into the corresponding layer of the nnU-Net decoder.

[0103] The MSFA module significantly improves the feature representation capability of segmentation networks by introducing explicit multi-scale fusion and cross-scale global modeling in the skip connection path. Specifically, the module first performs spatial alignment and unified embedding of the outputs of each encoding layer, and completes the centralized aggregation of multi-scale features in advance during the skip connection stage. Compared with nnU-Net, which only performs corresponding layer concatenation at the decoding end, this "early fusion" can fully preserve cross-layer semantic information in subsequent inference. Subsequently, the aggregated feature sequence is fed into a Transformer structure containing positional encoding. The self-attention mechanism not only establishes long-range dependencies, but also explicitly learns the interrelationships between features from different scale levels, enabling deep, low-resolution global structural information to directly guide shallow, high-resolution layers in judging blurred boundaries, and vice versa. The dynamic weighting characteristic of self-attention further completes feature selection and suppression, strengthens the representation of key regions of the segmentation task, and filters irrelevant noise, thereby passing more refined contextual information to the decoder. The addition of positional encoding helps the model accurately preserve the spatial source and scale level information of each feature when performing complex global feature interactions, thus avoiding confusion between features of different scales during the fusion process. Overall, MSFA elevates traditional skip connections from "shallow splicing" to "deep interaction," bridging the semantic gap between the codec and the decoder while maintaining positional and scale distinctions. It provides richer and more task-relevant multi-scale priors for the decoding stage, ultimately achieving effective gains in segmentation performance.

[0104] S4. Myocardial Ischemia Prediction Network

[0105] 4.1 Wavelet Transform Convolution (WTC)

[0106] Before introducing the ischemia prediction model, it is necessary to introduce a crucial component of the network: Wavelet Transform Convolution (WTC). Predicting myocardial ischemia directly from CCTA image patches is highly challenging. Ischemia can cause subtle changes within the myocardium or its relationship with surrounding structures, but these changes are often difficult to observe directly in CCTA images, instead manifesting as image features at different scales and frequencies. Standard CNN architectures, due to their limited receptive field, may be limited in capturing these subtle ischemia-related patterns that span multiple scales and may have long-range dependencies. Simply increasing the convolution kernel size introduces parameter and computational problems, while wavelet transform (WT) provides a multi-resolution analysis approach that preserves spatial locality while decomposing frequency information, making convolution operations in the wavelet domain an attractive option. Utilizing the multi-scale decomposition capabilities and potential receptive field expansion effect of wavelet transform, the network's ability to extract ischemia-related, multi-scale contextual information and subtle texture features from CCTA image patches can be enhanced.

[0107] This embodiment uses the three-dimensional discrete Haar wavelet transform and its inverse transform as its foundation. The Haar wavelet is the simplest and most intuitive type of wavelet, decomposing the signal based on the concepts of mean and difference. For the one-dimensional Haar wavelet transform, assume there is a one-dimensional discrete signal. For two adjacent samples and Calculate their mean and difference to obtain the low-frequency coefficient (mean). And high-frequency coefficients (differences) .

[0108] This decomposes the original signal into a low-frequency coefficient sequence. and high-frequency coefficient sequences Normalization coefficients are typically used in this process. This is to maintain energy conservation. Low-frequency sequences represent the smooth portion of the signal, reflecting its overall characteristics, while high-frequency signals reflect the differential portion, revealing more of the signal's details. In practical calculations, this can be achieved using depthwise convolution with a stride of 2. The convolution kernel can be represented as two FIR filters: and .

[0109] The same operation can be performed on the obtained results to achieve multi-level decomposition.

[0110] In image processing (two-dimensional signals), signal decomposition can be achieved by first applying a 1D Haar transform to the row direction and then applying a 1D Haar transform to the column direction. Specifically, for two-dimensional signals... Each row is subjected to a 1D Haar transform to obtain the low-frequency coefficient matrix and the high-frequency coefficient matrix. Then, a 1D Haar transform is applied to each column of the low-frequency coefficient matrix and the high-frequency coefficient matrix respectively, finally yielding 4 sub-bands:

[0111] ;

[0112] ;

[0113] ;

[0114] .

[0115] in, Both rows and columns are low-pass filtered to represent the overall trend of the image; Rows are low-pass filtered and columns are high-pass filtered to reflect the horizontal edge information of the image; Rows are filtered by a high-pass filter and columns are filtered by a low-pass filter to reflect the edge information in the vertical direction of the image; Both rows and columns are high-pass filtered to include details and noise information.

[0116] The entire process can be understood as the tensor product of a one-dimensional convolution kernel.

[0117] ;

[0118] Right now

[0119] .

[0120] After the transformation, the number of channels in the image becomes four times (each channel produces four sub-bands), and the spatial dimensions (length and width) become half of the original.

[0121] Similarly, this can be extended to 3D scenes. The 3D Haar transform contains 8 convolutional kernels:

[0122] ;

[0123] 8 sub-bands were obtained In this transformation, low-pass filtering is applied to all three directions of LLL to obtain a smooth, approximate global information; LLH, LHL, and HLL are low-pass filtered in two directions and high-pass filtered in one direction, reflecting detailed changes in that single direction; LHH, HLH, and HHL are low-pass filtered in one direction and high-pass filtered in two directions to capture finer local changes; and HHH is high-pass filtered in all three directions, containing the most detailed noise or texture information. Similarly, the spatial dimensions (length, width, and height) of the feature map are halved after the transformation.

[0124] For its inverse transform (IDWT), the Haar wavelet transform is orthogonal, meaning that reconstructing the original signal can be achieved naturally. Taking 1D as an example, the original signal can be reconstructed during the inverse transform using the following formula:

[0125] ;

[0126] In 2D, the inverse transformation is usually performed on the rows first, and then on the columns; similarly, in 3D, the inverse transformation is applied sequentially along each dimension.

[0127] The wavelet transform convolution (WTC) unit proposed in this embodiment combines the above transform with group convolution, and is defined as the following sequence operation:

[0128] ;

[0129] Where X is the input feature map. This represents the feature map containing 8 subbands obtained after applying 3D DWT. , In the wavelet domain (i.e., in The above) performs grouped convolution, and generally, to connect to the inverse transform, the number of input channels for this convolution is equal to the number of output channels. This indicates that 3D IDWT is applied to reconstruct the wavelet coefficients after convolution back to the original spatial domain.

[0130] This embodiment designs two grouping schemes for grouped convolutions, referred to as subband mixing and channel mixing. Subband mixing involves grouping the number of input channels by 8. In this case, the original 8 subbands per channel are grouped together, and each group of convolutions operates within the 8 subband features, promoting interaction between different frequency components decomposed from a single input feature channel. Channel mixing involves grouping 8 subbands, meaning each subband is grouped, and each group of convolutions processes one subband across all channels, focusing on independently learning cross-channel patterns within each frequency component. The choice of these two grouping strategies affects how the network learns features in the wavelet domain. Subband mixing tends to integrate cross-band information in local regions, while channel mixing tends to independently analyze feature representations within each band. To ensure that grouped convolutions work as expected, the arrangement order of the feature channels in the input grouped convolutions needs to be adjusted accordingly to meet the grouping requirements.

[0131] The WTC unit explicitly performs multi-scale (multi-frequency) decomposition (DWT) and reconstruction (IDWT) on the feature map, enabling the network to learn features across different frequency subbands. Theoretically, this allows for better capture of subtle image patterns related to ischemia that may exist at different frequencies and scales. On the other hand, although the wavelet domain convolution kernel size remains 3x3x3, it operates on the feature map downsampled by DWT. After reconstruction by IDWT, this operation effectively integrates a larger spatial range of information at the original resolution, thereby significantly expanding the receptive field of a single layer. This allows the network to integrate broader contextual information to determine the ischemic state of local areas, which may be beneficial for understanding the overall situation of the blood supply area. Moreover, compared to directly using very large convolution kernels to expand the receptive field, wavelet transform-based methods can achieve similar or even larger effective receptive fields with smaller kernels, potentially leading to higher parameter efficiency. Finally, performing convolution in the wavelet domain, especially with a "channel blending" grouping strategy, may enable the network to learn filters that are more sensitive to specific frequency components (such as low-frequency contours or high-frequency textures / edges), facilitating feature decoupling and specialization. This frequency-adaptive feature representation is expected to improve the model's ability to learn more discriminative features of myocardial ischemia when dealing with the challenge of low contrast in CCTA images.

[0132] Since the number of channels and spatial size of the feature map remain unchanged after DWT and IDWT operations, the WTC unit can be used as a plug-and-play module to easily replace the original standard convolutional layers in the network. However, this embodiment does not directly stack WTC units to build the network backbone, but instead integrates them into the Dense Layer of the 3D version of DenseNet to enhance its multi-scale feature learning capability. The structure of the Dense Layer in the standard DenseNet (usually in a pre-activated form) is: BatchNorm (BN) -> ReLU -> Conv (1x1x1, output 4k channels) -> BN -> ReLU -> Conv (3x3x3, output k channels) -> Dropout, where k is the growth rate, and the 1x1x1 convolution is used to reduce the number of channels, acting as a bottleneck. The improved Wavelet-TransformedDense Layer (WTDenseLayer3D) structure in this embodiment is as follows:

[0133] 1. Pre-activation and bottleneck layer: BN -> ReLU -> Conv(1x1x1). Unlike standard layers, this 1x1x1 convolution directly outputs k channels instead of 4k. Let its output be... .

[0134] 2. Wavelet domain processing: For Applying the 3D Haar transform, the expression is as follows:

[0135] ;

[0136] Then, wavelet domain feature learning is performed. Applying BN -> ReLU -> Conv, where both the input and output channels of Conv are 8k, the expression is as follows:

[0137] ;

[0138] 3. Application Refactoring: .

[0139] 4. Dropout: Optional Dropout layer.

[0140] In this way, this embodiment moves the core step of feature learning (the original 3x3x3 convolution) to the wavelet domain, while maintaining the spatial dimension and channel growth characteristics (increasing by k channels) of the entire Dense Layer input and output consistent with the original DenseNet. The network constructed by this Dense Layer is called WTDenseNet3D in this embodiment.

[0141] Extracting depth feature maps using WTDenseNet3D Subsequently, this embodiment introduces two key modules: Position Embedding (PE) and Aggregation Pooling (AP), to further process features and generate the final prediction. The motivation for introducing these modules stems from a common challenge in medical image classification tasks: how to effectively learn and focus on truly pathologically relevant image regions using only image / sample-level classification labels, while suppressing interference from irrelevant background information, in the absence of precise lesion location annotations (such as segmentation masks or bounding boxes). A similar challenge exists for the myocardial ischemia prediction task in this embodiment: although this embodiment has region-level ischemia labels (derived from FFR), the features indicating ischemia on CCTA images may be very subtle and spatially ambiguous. Directly using global pooling or fully connected layers for classification makes it difficult to automatically focus on these key features, and may even learn irrelevant or interfering information. Therefore, this embodiment introduces the idea of ​​guiding the model to focus on key regions using only image-level labels. By introducing the PE and AP modules, this embodiment aims to achieve a feature reweighting and saliency-guided aggregation mechanism based on saliency inference. This mechanism enables the model to learn to infer and enhance the spatial regions and feature channels most relevant to potential lesions (i.e., ischemia-related) without explicit localization supervision, and to focus subsequent aggregation operations more on these high-probability / high-significance features, simulating the behavior of doctors focusing on suspicious areas when reviewing images, thereby more effectively focusing the model's "attention" on high-probability areas.

[0142] The core objective of the PE module is to compute and apply location-related weights to adjust the feature map F output by the backbone network, thereby enhancing features corresponding to potential ischemic regions while suppressing features from background or irrelevant regions. This module contains two parallel branches that generate spatial position weights. Channel position weights ,like Figure 5 As shown.

[0143] Spatial branching aims to identify which locations in space are most likely to contain ischemia-related patterns. It first voxel-by-voxel... Perform a fully connected linear regression along the channel dimension (which can be implemented using a 1×1×1 convolution) and apply a sigmoid function to obtain a pixel-level significance graph. This process simulates a doctor browsing images to determine which regions are most likely to contain lesions. Subsequently, global max pooling and average pooling are performed on F and concatenated, then a 1×1×1 convolution is used to obtain a spatial attention map. and will and Element-wise multiplication yields .

[0144] Channel branching treats each channel as a high-order feature detector, used to identify regions with a high probability of containing lesions at the channel scale. Take the Top-K high response blocks for each channel ( (Empirical values), calculate their average activation value and maximum activation value respectively, and form a vector. And obtained through a shared MLP and Sigmoid. .

[0145] The final module outputs two feature maps that have been reweighted based on location information.

[0146]

[0147] in, This represents element-wise multiplication (Hadamard product). It is a feature map adjusted for attention in the spatial dimension. It is a feature map adjusted for attention along the channel dimension.

[0148] The AP module is designed to integrate the two feature maps output by the PE module. and The final classification prediction is then performed. The AP module combines two complementary pooling strategies to simulate a "consultation"-style attention mechanism among multiple radiologists, aiming to fully utilize the location attention information generated by the PE module to extract the final discriminative features from the regions and channels most likely to contain pathological features. Global Group-Max Pooling (GGP) is applied to the spatially reweighted feature map. .for For each channel, the top-K positions with the highest activation values ​​in the spatial dimension are selected, and the average of these values ​​is calculated to obtain a feature vector designed to capture the most significant spatial activation information in each important channel. Region-based Group-Max Pooling (RGP) was applied to the channel-reweighted feature map. It first assesses the importance of each spatial location, and then... The feature vectors formed by the features of all channels at each voxel point are input into a shared fully connected layer. Then, the top-K spatial locations with the highest activation values ​​are selected and their average value is calculated along the channel dimension to obtain a feature vector that aims to capture the comprehensive channel information of the most salient regions in space. Finally, the two feature vectors are added together to obtain the aggregated feature vector. This feature integrates information from the most suspicious spatial regions and feature channels for subsequent ischemic state prediction.

[0149] Therefore, the myocardial ischemia prediction network proposed in this embodiment adopts an end-to-end training method. For example... Figure 5 As shown, its overall architecture includes: (1) a WTDenseNet3D backbone network, used to extract depth and multi-scale features from the input CCTA image patch and the blood supply area mask; (2) a location embedding (PE) module, used to learn and apply spatial and channel attention to highlight potential ischemia-related spatial regions and feature channels; (3) an aggregate pooling (AP) module, used to efficiently aggregate key information from the weighted feature map to generate the final classification feature vector; and (4) a classification head (linear layer + Sigmoid), which outputs the ischemia prediction probability based on the aggregated feature vector. This design aims to improve the accuracy and robustness of the model in predicting myocardial ischemia in the absence of precise lesion location information through implicit, data-driven attention guidance.

[0150] S5. Training Settings

[0151] This embodiment involves two core deep learning models: a blood supply region segmentation network and a myocardial ischemia prediction network. All models are implemented using Python and PyTorch frameworks and trained on a deep learning server equipped with a single NVIDIA A6000 (48GB VRAM) GPU.

[0152] Blood Supply Region Segmentation Network: This network is built upon the mature nnU-Net framework. This embodiment largely follows the default training process and hyperparameter settings of nnU-Net, including its automated data preprocessing (such as normalization and resampling) and image analysis-based patching strategy. Specific preprocessing details can be found in the original nnU-Net paper. Training consists of 800 epochs, with each epoch defined as 250 iterations. The learning rate employs the standard nnU-Net polynomial decay strategy. For the improved model in this embodiment, which includes a multi-scale feature aggregation module, the initial learning rate is set to... In contrast, the default initial learning rate is used when training the original nnU-Net. .

[0153] loss function A combination of binary cross-entropy (BCE) loss and Dice loss is used:

[0154] ;

[0155] in, This represents the total loss due to the updating of network parameters for the myocardial blood supply region segmentation. Indicates the prediction mask, Indicates the first tag, This represents the binary cross-entropy loss. This indicates Dice's loss.

[0156] In nnU-net, the Dice loss is defined as a negative Dice coefficient (DSC) to encourage prediction of the region. With real area Maximize overlap:

[0157] ;

[0158] here It is a smoothing factor used to prevent the denominator from being zero and to improve numerical stability. The specific configuration of the optimizer (usually SGD with Nesterov momentum) follows the recommended settings in the nnU-Net paper.

[0159] Myocardial ischemia prediction network: The input to this network consists of preprocessed CCTA image patches and corresponding (predicted) blood supply region masks. The CT values ​​of the images are cropped to the [-1000, 1000] HU range and then normalized, i.e., mapped to [0,1]. The processed binary mask is used as an auxiliary channel and concatenated with the normalized image patch to form a multi-channel input. Training is performed for 100 epochs, with each epoch iterating through all training samples. Based on observation, the model typically converges around the 40th epoch. An exponential decay strategy is used for the learning rate, with an initial learning rate set to... Every 10 cycles, the learning rate is multiplied by a decay factor of 0.95.

[0160] Considering that the number of positive (ischemic) samples in the dataset is relatively small (approximately one-third), Focal Loss is selected as the loss function in this embodiment to alleviate the class imbalance problem. Focal Loss reduces the weight of easily classified samples, allowing the model to focus more on difficult-to-classify samples.

[0161] ;

[0162] in, It is the model's predicted probability of the true class. It is the focusing parameter. This is the class balance factor. The optimizer uses AdamW. This embodiment uses L2 weight decay as a regularization method, with the decay coefficient set to... According to AdamW's best practices, this weight decay is applied only to all weight parameters (w) of the network, while the bias parameters (b) and the parameters of normalization layers (such as BatchNorm) are not involved in the regularization.

[0163] S6. Verification Indicator Settings

[0164] To quantitatively evaluate the performance of the proposed blood supply area segmentation model, this embodiment uses two standard metrics commonly used in the field of medical image segmentation: the Dice Similarity Coefficient (DSC) and the 95th percentile Hausdorff Distance (HD95).

[0165] The Dice similarity coefficient (DSC) measures the degree of volume overlap between the model-predicted segmented region (P) and the manually labeled ground truth region (G). Its calculation formula is as follows:

[0166] ;

[0167] in, This represents the number of voxels representing the intersection (i.e., correct segmentation) between the predicted region P and the true region G. and These represent the total number of elements in the predicted region and the ground truth region, respectively. The DSC value ranges from [0,1]. A value closer to 1 indicates a higher degree of overlap between the predicted result and the ground truth label, resulting in better segmentation. A value of 0 indicates no overlap at all. This metric focuses on evaluating the overall accuracy of the segmentation.

[0168] The Hausdorff distance 95th percentile (HD95) is used to evaluate the distance deviation between the model's predicted segmentation boundary and the true boundary. The standard Hausdorff distance (HD) is defined as the maximum distance between two point sets, calculated as follows:

[0169] ;

[0170] in, and These represent the boundary point sets of the predicted region and the real region, respectively. Point and points The Euclidean distance between the two boundary profiles. HD reflects the degree of least mismatch between two boundary profiles. However, HD is very sensitive to outliers; a single erroneous boundary point can cause an abnormally large distance value. To improve the robustness of the indicator, HD95 is more commonly used in clinical assessment. It calculates the 95th percentile of the minimum distances from all boundary points to the other set, rather than the maximum value. Its calculation can be conceptually represented as:

[0171] ;

[0172] in, HD95 represents the 95th percentile of the set of minimum distances from all points in point set A to point set B. A smaller HD95 value indicates that the predicted boundary is closer to the true boundary, and the segmentation contour accuracy is higher. The unit for this metric is typically millimeters (mm).

[0173] For the binary classification task of predicting myocardial ischemia, this embodiment employs a series of standard and widely used metrics in the medical field to comprehensively evaluate the model's performance. The evaluation is based on four fundamental elements of the confusion matrix: True Positives (TP, correctly predicted as ischemia), True Negatives (TN, correctly predicted as non-ischemia), False Positives (FP, incorrectly predicted as ischemia), and False Negatives (FN, incorrectly predicted as non-ischemia). Specific evaluation metrics include:

[0174] Accuracy (ACC): A basic metric that measures the overall correctness of a model's predictions, defined as the proportion of correctly predicted samples out of the total number of samples. While intuitive, this metric can be misleading on imbalanced datasets.

[0175] ;

[0176] Sensitivity (SEN), also known as recall or true positive rate (TPR), measures a model's ability to correctly identify all samples that are actually positive (ischemic). In medical diagnosis, high sensitivity means a low rate of missed diagnoses, making it crucial.

[0177] ;

[0178] Specificity (SPE), also known as the True Negative Rate (TNR), measures a model's ability to correctly identify all samples that are actually negative (non-ischemic). High specificity means a low false diagnosis rate, helping to avoid unnecessary follow-up examinations or treatments.

[0179] ;

[0180] F1 Score: This metric represents the accuracy rate. The F1 score is the harmonic mean of the proportion of samples predicted as positive that were actually positive and the sensitivity (Recall). The F1 score provides a comprehensive evaluation of model performance, especially in cases of class imbalance, where it more robustly reflects the model's classification ability than accuracy. The F1 score ranges from [0,1], with higher values ​​indicating better model performance.

[0181] ;

[0182] Area Under the Receiver Operating Characteristic Curve (AUC): The ROC curve plots the model's sensitivity (TPR) relative to the false positive rate at different classification thresholds. The area under the ROC curve (AUC) is one of the most important metrics for measuring the overall discriminative performance of a classifier. The AUC ranges from [0,1], typically between [0.5,1]. An AUC of 0.5 indicates that the model's discriminative ability is equivalent to random guessing, while an AUC of 1 indicates a perfect classifier. A higher AUC value indicates a stronger ability to distinguish between positive and negative samples, and this metric is independent of the choice of a specific threshold.

[0183] Based on the above metrics, the segmentation metrics of the myocardial blood supply region segmentation network based on the traditional nnU-Net3d_fullres, nnU-Net2d, and nnFormer, as well as the myocardial blood supply region segmentation network (ours) of this application, are compared. As shown in Table 1, the myocardial blood supply region segmentation network used in this application has a significant improvement in performance.

[0184] Table 1

[0185]

[0186] Based on the above indicators, the performance of different myocardial ischemia prediction networks was compared, including the traditional DenseNet3D structure, the traditional DenseNet3D+PE&AP structure, the WTDenseNet3D structure based only on the backbone network of this application, and the WTDenseNet3D+PE&AP structure (ours). The results are shown in Table 2 below. It can be seen that the structure used in this application has significantly improved performance.

[0187] Table 2

[0188]

[0189] Corresponding to the above method, the present invention also provides an apparatus / system including a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus / system performs the steps of the method as described above.

[0190] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0191] In summary, the training method, detection method, and system for the non-invasive detection model of myocardial ischemia described in this invention, for CCTA images of user samples, constructs a training sample set by labeling at least three major coronary artery supply regions in the CCTA image based on the AHA 17-segment standard and marking whether the corresponding supply blocks are ischemic. In the first stage, a myocardial supply region segmentation network is trained. A multi-scale feature aggregation module is introduced to aggregate the outputs of multi-level encoders in the U-Net network, adds positional encoding, and inputs it to the Transformer encoder. The output of the Transformer encoder is separated and restored to its original size to obtain enhanced skip connection features, which are then input to the corresponding level decoder in the U-Net network. The second stage trains the myocardial ischemia prediction network by introducing wavelet transform convolution into a three-dimensional densely connected layer as the backbone network. Utilizing the three-dimensional discrete Haar wavelet transform and its inverse transform, combined with grouped convolution, feature learning in the wavelet domain is achieved. Subsequently, a location embedding module is introduced, generating spatial and channel location weights through parallel branches for reweighting. The global and regional group max pooling strategies of the aggregation pooling module are used to selectively aggregate features, and a classification head is used to detect whether the blood supply block is ischemic. This invention, under weak supervision without precise lesion annotation, effectively improves the performance of the ischemia prediction network through feature reweighting and saliency-guided aggregation, enabling it to better focus on key ischemia-related information.

[0192] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0193] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0194] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0195] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A training method for a non-invasive detection model of myocardial ischemia based on CCTA images, characterized in that, The method includes the following steps: For CCTA images from multiple user samples, the location mask of multiple blood supply blocks including the left anterior descending artery, right coronary artery and left circumflex artery in the left ventricle, which is labeled by the AHA 17-segment standard, is used as the first label, and whether each blood supply block is ischemic is added as the second label to construct a training sample set. The myocardial blood supply area segmentation network is trained using the training sample set and the CCTA image is used as input. The myocardial blood supply area segmentation network is obtained by embedding a multi-scale feature aggregation module into the skip connection path of the U-Net network. After aggregating the outputs of the multi-level encoders in the U-Net network, the multi-scale feature aggregation module adds position encoding and inputs it to the Transformer encoder. The output of the Transformer encoder is separated and restored to its original size to obtain enhanced skip connection features, which are then input to the decoder of the corresponding level in the U-Net network. The parameters of the myocardial blood supply area segmentation network are updated based on the predicted mask of each blood supply block position and the deviation of the first label. The myocardial ischemia prediction network is trained using the training sample set. Using the center of the mask corresponding to each blood supply block in the user samples as a reference, a three-dimensional image patch of a set size is extracted from the CCTA image as input. The myocardial ischemia prediction network decomposes the input features into eight sub-band features using a wavelet transform convolutional unit. These sub-band features are then grouped and convolved in the wavelet domain and reconstructed back to the original spatial domain. The reconstructed features are then processed by a location embedding module, which introduces spatial location weights and channel location weights to output a spatially reweighted feature map and a channel-reweighted feature map. A global group max pooling is performed on the spatially reweighted feature map, and a regional group max pooling is performed on the channel-reweighted feature map. These are then summed to obtain an aggregated feature vector. This aggregated feature vector is then processed by a classification head to output a prediction of whether the corresponding blood supply block is ischemic. The loss is constructed based on the deviation from the second label to update the parameters of the myocardial ischemia prediction network. The updated myocardial blood supply segmentation network and the myocardial ischemia prediction network are combined to form a non-invasive myocardial ischemia detection model.

2. The training method for a non-invasive detection model of myocardial ischemia based on CCTA images according to claim 1, characterized in that, Add a second tag indicating whether each blood supply block is blood-deficient, including: The second tag is added to detect whether each blood supply block of the user sample is ischemic using invasive-measured fractional flow reserve, immediate wave free ratio, resting full-cycle ratio, diastolic blood pressure ratio, coronary flow reserve and / or microcirculation resistance index.

3. The training method for a non-invasive detection model of myocardial ischemia based on CCTA images according to claim 1, characterized in that, The U-Net network in the myocardial blood supply segmentation network is an nnU-Net network, which includes multiple encoders for downsampling and corresponding decoders for upsampling. The multi-scale feature aggregation module applies independent 3D convolution to the output of each encoder for fragment embedding, then performs shaping and splicing for aggregation and adds fixed or learnable positional encoding. This is then handed over to the Transformer encoder to use a multi-head self-attention mechanism to capture long-range dependencies within the sequence and interaction relationships between features. The output of the Transformer encoder is subjected to dimensionality separation and reshaped back to the scale of the original encoder output, then upsampled back to the original skip connection spatial size and input to the corresponding decoder.

4. The training method for a non-invasive detection model of myocardial ischemia based on CCTA images according to claim 1, characterized in that, In the wavelet transform convolutional unit, the k-channel output after the input passes through the first batch of normalization layers, the first activation function layer, and the 1*1*1 convolutional layer undergoes 3D Haar wavelet transform, and then passes through the second batch of normalization layers, the second activation function layer, and the grouped convolutional layer to apply subband mixing or channel mixing processing to obtain an 8k-channel output. This output is then reconstructed back to the original spatial domain through inverse wavelet transform and finally output by the Dropout layer.

5. The training method for a non-invasive detection model of myocardial ischemia based on CCTA images according to claim 1, characterized in that, The method further includes: The parameters of the myocardial blood supply region segmentation network are updated based on the binary cross-entropy loss and Dice loss between the predicted mask and the first label. The total loss is calculated as follows: ； ； in, This represents the total loss from updating the parameters of the myocardial blood supply region segmentation network. This refers to the prediction mask. This refers to the first label. This represents the binary cross-entropy loss. This represents the Dice loss. This is a smoothing factor.

6. The training method for a non-invasive detection model of myocardial ischemia based on CCTA images according to claim 1, characterized in that, The total loss calculation formula for updating the parameters of the myocardial ischemia prediction network using the method is as follows: ； in, This represents the total loss in updating the parameters of the myocardial ischemia prediction network. This represents the predicted probability of the true category of the second label. It is the focusing parameter. It is a category balance factor; The method applies L2 weight decay for regularization during the parameter update process of the myocardial ischemia prediction network.

7. A non-invasive detection method for myocardial ischemia based on CCTA images, characterized in that, The method includes the following steps: Obtain the CCTA image to be analyzed; The CCTA image is input into the non-invasive myocardial ischemia detection model in the training method of the non-invasive myocardial ischemia detection model based on CCTA image according to any one of claims 1 to 6, and the myocardial blood supply area segmentation network in the non-invasive myocardial ischemia detection model identifies multiple blood supply block location masks including the left anterior descending branch, the right coronary artery and the left circumflex branch in the left ventricle. Using the center of the blood supply block location mask as a reference, a three-dimensional image block of a set size is extracted from the CCTA image and input into the myocardial ischemia prediction network in the non-invasive myocardial ischemia detection model, and the identification result of whether myocardial ischemia exists in each blood supply block is output.

8. A non-invasive detection system for myocardial ischemia based on CCTA images, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 7.

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

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images