A method and device for constructing a coronary artery stenosis detection model

Abstract

This specification provides a method for constructing a coronary artery stenosis detection model, as well as a method, device, medium, and program product based on this detection model for coronary artery stenosis detection, relating to the field of intelligent healthcare. The method of this invention, through a unique network architecture design and frequency domain decoupling analysis mechanism, effectively solves the problems faced by existing coronary artery stenosis detection technologies in clinical applications, such as insufficient small target detection capability, easy loss of deep features, poor anti-interference robustness, and high misjudgment rate in complex backgrounds.

Description

Technical Field

[0001] This invention relates to the field of intelligent healthcare, and more specifically, to a method and apparatus for constructing a coronary artery stenosis detection model. Background Technology

[0002] Coronary artery stenosis is a major pathological basis for cardiovascular disease, and its early and accurate diagnosis is crucial for developing treatment plans and improving patient prognosis. Currently, non-invasive coronary CT angiography (CCTA) image analysis has become a research hotspot in auxiliary diagnosis. However, existing coronary artery stenosis detection technologies still face many technical bottlenecks that urgently need to be addressed in clinical applications: First, the detection capability for small targets is insufficient, and deep features are easily lost. Early coronary artery stenosis lesions occupy a very small proportion in images, representing typical small targets. When existing network models extract deep semantic features through multi-layer downsampling, the spatial resolution is often reduced due to an excessively large receptive field or pooling operations. This causes the weak feature signals of small targets to gradually attenuate or even be completely lost during transmission, resulting in missed detections.

[0003] Secondly, it exhibits poor robustness against interference and a high misclassification rate in complex backgrounds. Coronary angiography images are often accompanied by complex background factors such as calcified plaque interference, uneven contrast agent filling, and imaging noise. Traditional models struggle to effectively distinguish between the true stenosis edge gradient and background noise (such as speckle noise), leading to situations where the model misclassifies noise as lesions or obscures true lesions in low-dose imaging or complex anatomical structures.

[0004] Finally, the models lack interpretability and have low clinical trust. Existing deep learning models often exhibit "black box" decision-making; doctors only receive the final test result but cannot know which features of the image (the overall direction of blood vessels or local edge mutations) the model used to make its judgment. This lack of visibility into the decision-making basis makes it difficult for clinicians to assess the reliability of AI results, limiting its in-depth application in actual diagnostic and treatment processes. Summary of the Invention

[0005] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention provides a method for constructing a coronary artery stenosis detection model, as well as a method and device for coronary artery stenosis detection based on this detection model. The method of this invention effectively solves the problems faced by existing coronary artery stenosis detection technologies in clinical applications through a unique network architecture design and frequency domain decoupling analysis mechanism.

[0006] The first aspect of this application discloses a method for constructing a coronary artery stenosis detection model, including: Obtain a coronary angiography sample containing an image and its corresponding mask label, wherein the label includes a vessel mask label and a stenosis mask label; The image is input into a neural network model containing a decoder and an encoder, and the encoder extracts an initial feature map. The initial feature map is sequentially input into a feature extraction module containing large kernel convolution and depthwise separable convolution and a multi-scale local contrast learning module; the initial feature map is processed by depthwise separable convolution using the large-size convolution kernel in the feature extraction module to obtain an intermediate feature map; the contrast between the center pixel and the surrounding pixels is calculated by the multi-scale local contrast learning module according to local sliding windows with different hole rates to generate a multi-scale attention weight map, and the weight map and the intermediate feature map are weighted and fused to output the target feature map; The target feature map is downsampled at least once to obtain a multi-scale feature map; The multi-scale feature map is processed by a decoder to generate a prediction result; the mask label is used as a supervision signal, and the model is trained based on the difference between the prediction result and the supervision signal to obtain a vascular stenosis detection model.

[0007] The second aspect of this application discloses a method for detecting coronary artery stenosis based on a coronary artery stenosis detection model, comprising: Receive the raw coronary angiography image to be tested; The original coronary angiography image is input into the vascular stenosis detection model trained by the construction method disclosed in the first aspect of this application to obtain an initial class activation heatmap; The initial class activation heatmap is subjected to multi-level discrete wavelet transform for frequency domain decoupling to obtain low-frequency sub-bands and high-frequency sub-bands. The low-frequency sub-bands are used to characterize the global anatomical structure information of blood vessels, and the high-frequency sub-bands are used to characterize the gradient abrupt change features of the blood vessel edges. By fusing the features of the low-frequency subband and the high-frequency subband, a final attention heatmap is generated; The target location of coronary artery stenosis is located based on the final attention heatmap.

[0008] A third aspect of this application discloses a computer device, comprising: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the steps of the above-described method.

[0009] The fourth aspect of this application discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0010] The fifth aspect of this application discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0011] This application has the following beneficial effects: 1. This application innovatively designs a structure in the feature extraction module of the model construction stage, which includes a large-kernel depthwise separable convolution and a multi-scale local contrastive learning module in series. This improves the extraction and preservation of small target features and solves the problem of feature loss in deep networks. Large-kernel convolution expands the receptive field to capture global contextual information, preventing small targets from getting lost in deep networks. Simultaneously, the multi-scale local contrastive learning module uses sliding windows with different dilation rates to calculate the contrast between the center pixel and the surrounding environment, generating an attention weight map and performing weighted fusion of features. This mechanism can explicitly enhance the discriminative features of small, narrow lesions, ensuring that key small target information is fully preserved even after multiple downsampling, reducing the false negative rate of early small, narrow lesions.

[0012] 2. This application introduces multi-level discrete wavelet transform to decompose image signals into sub-bands of different frequencies. The high-frequency sub-band accurately captures abrupt changes in the edge gradient of narrowed lesions, while specific high-frequency-to-high-frequency components concentrate interference information such as speckle noise. Through differentiated processing of these sub-bands (such as noise suppression steps), the model can effectively suppress the interference of background noise on decision-making. Even in complex anatomical contexts with uneven contrast agents or calcified plaques, it maintains stable and high-precision detection performance, reducing false alarms and false negatives.

[0013] Furthermore, unlike traditional single-scale fusion operations, this application employs a cross-order feature fusion step. It utilizes the coherent vascular tree topology provided by the second-order low-frequency subband (LL2) as "navigation" to guide the search range and avoid vascular breakage. Simultaneously, it uses the fine edge gradient information provided by the first-order high-frequency subband (LH1, HL1) as a "scale" to characterize the physical boundaries of the stenotic lesion. The final attention heatmap generated by fusing these two methods ensures both macroscopic accuracy of the target location and clear microscopic boundaries. This addresses the problems of some methods that can roughly locate stenotic areas but have blurred boundaries, making precise measurement impossible; and other methods that, while capturing edge details, are susceptible to local artifacts due to a lack of global topological information, resulting in breaks or misalignments and hindering sub-pixel-level precise localization.

[0014] 3. This application transforms the previously invisible "black box" decision-making process into visualized frequency domain response evidence. Doctors can intuitively observe whether the model identifies the region based on the overall morphology of the blood vessel (low-frequency component) or confirms stenosis based on drastic changes at the edges (high-frequency component). This clear display of decision-making logic not only provides doctors with credible auxiliary diagnostic evidence and optimizes clinical workflows, but also offers an intuitive tool for explaining conditions to patients. Furthermore, it is cost-effective and helps promote its widespread implementation in primary healthcare institutions and tiered medical service models. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the construction method of the coronary artery stenosis detection model provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the coronary artery stenosis detection method based on the coronary artery stenosis detection model provided in Embodiment 2 of the present invention; Figure 3 This is a visualization of the first-order frequency domain decoupling effect of the original input image provided by the present invention; Figure 4 This is a visualization of the second-order frequency domain decoupling provided by an embodiment of the present invention; Figure 5 This is a schematic diagram of the characteristic responses of different frequency domain channels (from left to right: low frequency, mid frequency, and high frequency) when detecting "narrow" targets, provided by an embodiment of the present invention. Figure 6 This is the final attention heatmap visualization result generated for coronary angiography samples provided in this embodiment of the invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0018] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 like Figure 1 As shown, this embodiment of the invention provides a method for constructing a coronary artery stenosis detection model, specifically including: S101: Obtain a coronary angiography sample containing an image and its corresponding mask label, wherein the label includes a vessel mask label and a stenosis mask label; In some embodiments, the method for obtaining the coronary angiography sample includes: Obtain raw coronary CT angiography sequence images; Keyframe images with clear vascular structures and clear stenotic segments were selected from the original coronary CT angiography sequence images. The keyframe images are annotated to generate vascular mask images with complete vascular regions marked and stenosis mask images with only stenosis regions marked. The keyframe image and its corresponding mask image are subjected to size normalization and data augmentation processing to obtain the coronary angiography sample.

[0021] In some embodiments, the coronary angiography samples in this embodiment are derived from any animal (e.g., mammal), including but not limited to humans, non-human primates, rodents, etc., which will become the recipient of a specific treatment. Preferably, they are derived from humans.

[0022] S102: Input the image into a neural network model containing a decoder and an encoder, and extract the initial feature map through the encoder; S103: The initial feature map is sequentially input into the feature extraction module containing large kernel convolution and depthwise separable convolution and the multi-scale local contrast learning module; the initial feature map is processed by depthwise separable convolution using the large-size convolution kernel in the feature extraction module to obtain an intermediate feature map; the contrast between the center pixel and the surrounding pixels is calculated by the multi-scale local contrast learning module according to local sliding windows with different hole rates to generate a multi-scale attention weight map, and the weight map and the intermediate feature map are weighted and fused to output the target feature map; In some embodiments, the processing procedure of the feature extraction module comprising large kernel convolution and depthwise separable convolution includes: The initial feature map is sequentially subjected to large kernel depthwise separable convolution, channel expansion point convolution, and channel compression point convolution to construct an inverted bottleneck structure; The initial feature map is residually concatenated with the output feature map after channel compression point convolution.

[0023] In some embodiments, the weighted fusion process of the weight map and the intermediate feature map includes: multiplying the weight map and the intermediate feature map element by element, and then performing channel fusion on the multiplied feature map through a point convolutional layer.

[0024] S104: Perform at least one downsampling on the target feature map to obtain a multi-scale feature map; S105: The multi-scale feature map is processed by the decoder to generate a prediction result; the mask label is used as a supervision signal, and the model is trained based on the difference between the prediction result and the supervision signal to obtain a vascular stenosis detection model.

[0025] In some embodiments, the decoder employs a bilinear interpolation attention module for feature fusion: The multi-scale feature maps are upsampled step by step, and semantic information is filtered out by an attention gate. Then, they are fused with shallow feature maps of the corresponding resolution to generate the prediction result.

[0026] In some embodiments, the neural network model further includes a slice-assisted super-segmentation module, which cuts the feature map fused by the decoder into multiple sub-feature maps with a fixed stride, and inputs them into the prediction convolutional layer for independent prediction, thereby enhancing the model's learning of small targets.

[0027] In some embodiments, the loss function during model training is obtained by weighted summation of the vascular prediction loss function and the stenosis prediction loss function, and both the vascular prediction loss function and the stenosis prediction loss function adopt a hybrid loss function combining binary cross-entropy loss and Dice loss.

[0028] Example 2 like Figure 2 As shown in the figure, this application provides a method for detecting coronary artery stenosis based on a coronary artery stenosis detection model, specifically including: S201, Receive the raw coronary angiography image to be detected; S202, The original coronary angiography image is input into the vascular stenosis detection model trained by the construction method disclosed in Example 1 to obtain an initial class activation heatmap; S203, Perform multi-level discrete wavelet transform on the initial class activation heatmap to decouple it in the frequency domain, and obtain low-frequency sub-band and high-frequency sub-band, wherein the low-frequency sub-band is used to characterize the global anatomical structure information of blood vessels, and the high-frequency sub-band is used to characterize the gradient abrupt change features of blood vessel edges; In some embodiments, the multi-level discrete wavelet transform includes a first-order discrete wavelet transform and a second-order discrete wavelet transform: The initial class activation heatmap is decomposed into a first-order low-frequency approximate subband and multiple first-order high-frequency subbands in multiple directions by the first-order discrete wavelet transform. The first-order low-frequency approximation subband is subjected to a discrete wavelet transform again, i.e., a second-order discrete wavelet transform, to obtain a second-order low-frequency approximation subband and a second-order high-frequency subband with a higher level of semantic abstraction. The secondary low-frequency approximation subband expands the receptive field by aggregating discontinuous vascular fragments into a coherent vascular tree, thereby avoiding vascular rupture caused by uneven contrast agent filling.

[0029] In some embodiments, the plurality of directional first-level high-frequency subbands include: Low-frequency to high-frequency components used to capture high-frequency changes in the vertical direction and vertical edge features; High-frequency and low-frequency components used to capture high-frequency changes in the horizontal direction and horizontal edge features; And, high-frequency-high-frequency components used to capture high-frequency variations in the diagonal direction and isolate speckle noise.

[0030] S204, The features of the low-frequency subband and the high-frequency subband are fused to generate the final attention heatmap; In some embodiments, prior to generating the final attention heatmap, a noise suppression step is included: identifying and suppressing speckle noise features of the high-frequency-high-frequency component (HH1) to improve the signal-to-noise ratio of the model in complex anatomical contexts.

[0031] In some embodiments, the specific steps of fusing the features of the low-frequency subband and the high-frequency subband include: The low-frequency subband characterizing the global anatomical topology of blood vessels is recombined and fused with the high-frequency subband characterizing gradient mutations at the vessel margins. The low-frequency subband is used to aggregate discontinuous vascular segments into a coherent vascular tree, providing global topological guidance for stenosis localization. The high-frequency subband is used to provide the precise physical boundaries of stenotic lesions and characterize the edge contours of the stenotic region.

[0032] In some preferred embodiments, the low-frequency subband is a second-order low-frequency approximate subband (LL2) obtained by second-order discrete wavelet transform, and the high-frequency subband is a first-order high-frequency subband (LH1, HL1) obtained by first-order discrete wavelet transform.

[0033] S205, locate the target location of coronary artery stenosis based on the final attention heatmap.

[0034] In some embodiments, the step of locating the target coronary artery stenosis further includes: The narrow target region is highlighted in the final attention heatmap, and the corresponding narrow confidence score is also output.

[0035] Example 3 This application provides a computer device, which may include one or more processors and one or more memories; wherein the memories store computer-readable code, which, when run by the one or more processors, can perform the methods described above.

[0036] The processor in this embodiment can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, operations, and logic block diagrams disclosed in this embodiment. The general-purpose processor can be a microprocessor or any conventional processor, and can be based on an x86 or ARM architecture.

[0037] In general, the various exemplary embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, firmware, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device. When aspects of embodiments of this disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as non-limiting examples in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0038] For example, the methods or apparatus according to embodiments of this disclosure can also be implemented using the architecture of a computing device. A computing device may include a bus, one or more CPUs, read-only memory (ROM), random access memory (RAM), a communication port connected to a network, input / output components, a hard disk, etc. Storage devices in the computing device, such as ROM or hard disk, may store various data or files used for processing and / or communication of the methods provided in this disclosure, as well as program instructions executed by the CPU. The computing device may also include a user interface. Of course, the architecture described above is merely exemplary, and one or more components of the computing device described above may be omitted as needed when implementing different devices.

[0039] Example 4 This application also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by a processor, the methods disclosed in this embodiment can be performed. The computer-readable storage medium in this embodiment can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0040] Example 5 This disclosure also provides a computer program product or system, including a computer program that, when executed by a processor, implements the steps of any of the methods described above.

[0041] Example 6 This application provides a system for constructing a coronary artery stenosis detection model, used to implement the steps of the method described in Embodiment 1, including: An angiography sample acquisition module is used or configured to acquire a coronary angiography sample containing an image and its corresponding mask label, wherein the label includes a vessel mask label and a stenosis mask label; An encoder feature extraction module is used or configured to input the image into a neural network model containing a decoder and an encoder, and extract an initial feature map through the encoder; An image processing and fusion module is used or configured to sequentially input the initial feature map into a feature extraction module containing large kernel convolution and depthwise separable convolution and a multi-scale local contrast learning module; use the large-size convolution kernel in the feature extraction module to perform depthwise separable convolution processing on the initial feature map to obtain an intermediate feature map; use the multi-scale local contrast learning module to calculate the contrast between the center pixel and the surrounding pixels according to local sliding windows with different hole rates, generate a multi-scale attention weight map, and perform weighted fusion of the weight map and the intermediate feature map to output the target feature map; An image downsampling module is used or configured to downsample the target feature map at least once to obtain a multi-scale feature map; The model training module is used or configured to process the multi-scale feature map through a decoder to generate prediction results; and to train the model based on the difference between the prediction results and the supervision signals, using the mask labels as supervision signals, to obtain a vascular stenosis detection model.

[0042] In some embodiments, the neural network model further includes a slice-assisted super-segmentation module, which cuts the feature map fused by the decoder into multiple sub-feature maps with a fixed stride, and inputs them into the prediction convolutional layer for independent prediction, thereby enhancing the model's learning of small targets.

[0043] In some embodiments, the loss function during model training is obtained by weighted summation of the vascular prediction loss function and the stenosis prediction loss function, and both the vascular prediction loss function and the stenosis prediction loss function adopt a hybrid loss function combining binary cross-entropy loss and Dice loss.

[0044] Example 7 This application provides a coronary artery stenosis detection system based on a coronary artery stenosis detection model, used to implement the steps of the method described in embodiment 2, including: The image acquisition module is used or configured to receive the raw coronary angiography image to be detected; The feature extraction and segmentation module is used or configured to input the original coronary angiography image into the vascular stenosis detection model trained by the construction method disclosed in the first aspect of this application to obtain an initial class activation heatmap; The frequency domain decoupling analysis module is used or configured to perform multi-level discrete wavelet transform on the initial class activation heatmap to perform frequency domain decoupling, thereby obtaining a low-frequency sub-band and a high-frequency sub-band, wherein the low-frequency sub-band is used to characterize the global anatomical structure information of blood vessels, and the high-frequency sub-band is used to characterize the gradient abrupt change features of the blood vessel edge. A low-frequency and high-frequency sub-band fusion module is used or configured to fuse the features of the low-frequency sub-band and the high-frequency sub-band to generate a final attention heatmap. A target localization module is used or configured to locate the target location of coronary artery stenosis based on the final attention heatmap.

[0045] In some embodiments, the system further includes: The narrow confidence score output module is used or configured to output a highlighted narrow target region in the final attention heatmap and simultaneously output the narrow confidence score corresponding to that region.

[0046] In this embodiment, in order to more clearly describe the research process of automatic detection of coronary artery stenosis based on the coronary artery stenosis detection model, this embodiment takes the study of raw coronary CT angiography (CCTA) sequence images (DICOM format) obtained in clinical practice as an example for exemplary description. However, this description is only one use case and is not intended to limit the scope of protection of this data processing method.

[0047] 1. Model training phase, which is the specific operation steps of Example 1 above. The purpose of this phase is to train a robust CN-UNet network model using a large amount of labeled data.

[0048] Step 1.1: Sample Set Construction and Preprocessing (Data Preparation Module) Input: Raw coronary CT angiography (CCTA) sequence images (DICOM format).

[0049] process: Data Reading and Filtering: Read CCTA image sequences from multiple patients. Select sequences with high image quality and no severe artifacts as candidates.

[0050] Keyframe selection: From each patient's sequence, keyframe images with clear vascular structures and the most obvious stenotic segments are selected as the main processing objects.

[0051] Labeling (Generate Labels): Using professional labeling tools, manually outline the coronary arteries on the selected image and specifically mark the locations of stenotic segments. Generate two binary mask images: Blood vessel mask: The grayscale value of the entire blood vessel area is marked as 255, and the background is 0.

[0052] Narrow mask image: Only the narrow segment of the image is marked with a grayscale value of 255, while the background is 0. The training objective of the network in this invention is to predict both mask images simultaneously.

[0053] Data preprocessing: The original image and the corresponding mask label image are processed as follows: Resampling and normalization: The image is resized to a uniform size, for example, 512×512 pixels. The pixel values ​​of the original image are normalized, linearly mapping them to the [0, 1] interval.

[0054] Data augmentation: To improve model robustness, the training set data is augmented in real time, including random rotation (-10° to 10°), random horizontal / vertical flipping, and random brightness and contrast adjustment.

[0055] Output: The completed sample set. Each sample contains: a preprocessed original image, and its corresponding blood vessel mask label and stenosis mask label. The sample set is randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0056] Step 1.2: Construct the CN-UNet deep neural network model Input: Images (512×512×1) in the training set.

[0057] Process: The CN-UNet network (i.e., the model built in Part 1 of Example 1) is constructed using the deep learning framework (PyTorch). This network is one of the core components of this invention, and its internal data flow is as follows: Encoder stage (feature extraction): First stage: The input image X (512×512) is first downsampled by 4× through a Stem module, and the output feature map M1 (128×128×64) is output.

[0058] The second stage involves feeding M1 into a structure consisting of a ConvNeXt module and a multi-scale local contrastive learning module (MSL-CLM) connected in series.

[0059] ConvNeXt module data flow: Input M1 (128×128×64). First, it passes through a 7×7 depthwise convolutional layer (Depthwise Conv) to capture a wide range of contextual information, enhancing the global receptive field and addressing the issue of small object features being easily lost in deep networks. Next, it passes through two 1×1 pointwise convolutional layers (Pointwise Conv). The first expands the number of channels to four times the original (128×128×256), simulating the feature dimension of multi-head attention; the second compresses the number of channels back to the original dimension (128×128×64), forming an inverted bottleneck structure. Finally, a residual connection adds the input and output.

[0060] MSL-CLM Module Data Flow: The feature map output from the ConvNeXt module enters the MSL-CLM. This module runs three local contrastive learning units with different dilation rates in parallel or serially. Each unit calculates the contrast between the center pixel and surrounding pixels through a local sliding window, generating an attention weight map that highlights small (narrow) targets. These weight maps are element-wise multiplied with the original feature map and then fused through a 1×1 convolutional layer. The mathematical expression of this process can be simplified as follows: W i It is the contrast weight learned at the i-th scale.

[0061] Output: The final output of this stage is a highly discriminative feature map M2 (128×128×64).

[0062] Stages 3 through 5: M2 continues to be downsampled using a regular ResNet module, generating feature maps M3 (64×64×128), M4 (32×32×256), and M5 (16×16×512), respectively. These feature maps contain multi-scale information ranging from local details to global semantics.

[0063] Decoder stage: Process: The decoder uses a simplified bilinear interpolation attention module (SBAM) for feature fusion.

[0064] Its data flow is as follows: M4 and M5 are first adjusted by a 1×1 convolution, then SBAM performs 2× bilinear upsampling on M5 to make it consistent with the resolution of M4, and then passes an attention gate to filter out important semantic information, which is added to and fused with M4 to generate (32×32×256).

[0065] Repeat this process: D4 and M3 are merged to generate D3 (64×64×128), D3 and M2 are merged to generate D2 (128×128×64), and D2 and M1 are merged to generate D1 (256×256×64).

[0066] Slice-Assisted Hypersegmentation (SAHS) module: Process: During training, the SAHS module divides D1 into multiple small pieces (e.g., 64×64) with a fixed stride, and feeds them into subsequent 1×1 convolutional layers for prediction to enhance the learning of small targets. During inference, the SAHS module predicts overlapping small pieces and then uses non-maximum suppression (NMS) to stitch the results together to ensure that no small targets are missed.

[0067] Output: Generates an initial class activation heatmap Ocam, the same size as the original input, with dimensions 512×512. This heatmap provides a preliminary prediction of the probability that each pixel in the image belongs to a vascular region.

[0068] Output: Deep neural network model and feature maps of each layer.

[0069] Step 1.3: Model Training and Loss Calculation Input: The initial heatmap Ocam predicted by the model in step 1.2, and the vascular mask label Gvessel and stenosis mask label Gstenosis prepared in step 1.1.

[0070] process: Loss function design: The total loss function, Loss_total, consists of a weighted sum of two parts: Where λ1 and λ2 are weighting coefficients (empirically set to 0.5 and 0.5 respectively). Both Loss_vessel and Loss_stenosis employ a hybrid loss method, combining binary cross-entropy (BCE) loss and Dice loss to address the imbalance between positive and negative samples. , Where P and G are the predicted result and the true label, respectively, p i and g i These represent the predicted probability and the actual value for the corresponding pixel.

[0071] Optimization strategy: Use the Adam optimizer, with an initial learning rate of 1×10⁻⁶. 4 A cosine annealing decay strategy is employed. The batch size is set to 8, and the number of iterations is 100. In each epoch, the training set data is input into the network in batches, the loss_total is calculated, and the network's weight parameters are updated using backpropagation. After each epoch, the model performance is validated using a validation set.

[0072] Output: Weights of the trained CN-UNet network model.

[0073] 2. The inference and detection stage, i.e., the specific operation steps of Embodiment 2 above, uses the trained model to detect stenosis in new images and introduces another core innovation of this invention—frequency domain decoupling analysis—to enhance the interpretability and localization accuracy of the model. This embodiment provides a flowchart for localizing coronary stenosis, specifically including: Step S2.1: Acquire coronary angiography images of the target coronary artery. Input: A new, unlabeled patient coronary angiography image.

[0074] Process: Perform the same standardized preprocessing (size adjustment, normalization, etc.) as step 1.1 of the training phase.

[0075] Output: The preprocessed image to be detected, Xnew.

[0076] Step S2.2: Generate the initial class activation heatmap Input: The preprocessed image to be detected, Xnew.

[0077] Procedure: Input Xnew into the trained CN-UNet network. The data flows through the encoder-decoder forward propagation process described in step 1.2.

[0078] Output: An initial class activation heatmap, Oneu, of the same size as the input, output by the decoder. This heatmap roughly indicates regions that the model considers to be blood vessels.

[0079] Step 2.3: Multi-level frequency domain decoupling analysis (Key improvement point 2) This step provides an in-depth analysis of the model's "black box" decision-making process, greatly enhancing the interpretability of the detection results and the accuracy of the localization.

[0080] Input: The initial class activation heatmap O generated in step 2.2 new .

[0081] Process: Perform multi-level discrete Haar wavelet transform (DHWT) on Oneu. The decoupling principle between its data flow and frequency domain is as follows: First-order Haar wavelet transform: Operation: Move O new It is considered as a two-dimensional discrete signal. DHWT decomposes it into four frequency sub-bands through high-pass and low-pass filters.

[0082] Output: LL1 (Low-frequency component): O new An approximate representation of blood vessels preserves the global anatomical structure and topological information of the vessels.

[0083] LH1 (low-frequency to high-frequency component): captures high-frequency changes in the vertical direction, i.e., vertical edge features (such as the vertical orientation of the blood vessel wall).

[0084] HL1 (High-frequency-low-frequency component): Captures high-frequency changes in the horizontal direction, i.e., horizontal edge features (such as the horizontal transition at the bifurcation of blood vessels).

[0085] HH1 (High-Frequency Component): Captures high-frequency variations along the diagonal direction, primarily corresponding to speckle noise and extremely fine textures in the image.

[0086] Second-order Haar wavelet transform: Operation: Perform DHWT again on the LL1 subband obtained by the first-order transformation.

[0087] Output: LL2 (low-frequency-low-frequency component): a higher level of semantic abstraction that aggregates discontinuous vascular fragments into a highly coherent vascular tree, expanding the receptive field and avoiding vascular rupture caused by uneven contrast agent filling.

[0088] LH2, HL2, HH2: These correspond to high-frequency features at the semantic level, encoding more macroscopic morphological changes, such as complex texture patterns in patchy regions.

[0089] By setting up a multi-level frequency domain decoupling analysis stage, the following beneficial effects were achieved: (1) Improve small target detection and noise immunity: Through this decoupling process, high-frequency subbands (such as LH1 and HL1 were explicitly separated, precisely pinpointing the edge gradient mutations of the stenotic lesions and providing accurate physical boundaries for stenosis localization. Noise (speckle) was effectively isolated within the HH1 subband. This directly resulted in a higher signal-to-noise ratio and stronger robustness in detecting small targets (early, minute stenosis).

[0090] (2) Improve model interpretability: This step decomposes the model's perception of narrowing into understandable frequency domain components. Doctors can visually visualize this (e.g., Figure 3 , Figure 4 As shown, CAM (the aforementioned activation heatmap) is used to model the search area based on the overall shape of the blood vessel (LL sub-band) and to quantify the degree of stenosis based on edge details (LH, HL sub-bands). Figure 5 As shown, after frequency domain decoupling, the low-frequency channel provides the global background, the mid-frequency channel captures the texture of the blood vessel wall, and the high-frequency channel precisely locates the stenotic lesion. This explicit and visualized decision-making basis enhances doctors' trust in AI-assisted diagnosis.

[0091] Step 2.4: Feature Recombination and Final Target Localization Input: Multi-level frequency sub-bands, especially LL2 with strong semantic information and LH1 and HL1 with strong edge information.

[0092] Process: These key subbands are recombined using a lightweight fusion module (a 1×1 convolutional layer) to generate a final attention heatmap. final .

[0093] Output: Generates a final attention heatmap O that accurately covers the LAD / LCX bifurcation and effectively suppresses background noise. final (like Figure 6As shown in the figure, the red highlighted area is the target location of coronary artery stenosis determined by the model, and a confidence score in the interval [0, 1] can be output at the same time.

[0094] The final heatmap integrates two complementary pieces of information: "where is the target?" provided by low frequency and "where is the target's boundary?" provided by high frequency. This resolves the common contradiction in small target detection: "accurate location but blurry outline" or "clear outline but off-center location," achieving sub-pixel-level precise localization of narrow lesions.

[0095] The exemplary embodiments of this disclosure described in detail above are merely illustrative and not restrictive. Those skilled in the art will understand that various modifications and combinations can be made to these embodiments or their features without departing from the principles and spirit of this disclosure, and such modifications should fall within the scope of this disclosure.

Claims

1. A method for constructing a coronary artery stenosis detection model, characterized in that, include: Obtain a coronary angiography sample containing an image and its corresponding mask label, wherein the label includes a vessel mask label and a stenosis mask label; The image is input into a neural network model containing a decoder and an encoder, and the encoder extracts an initial feature map. The initial feature map is sequentially input into a feature extraction module containing large kernel convolution and depthwise separable convolution and a multi-scale local contrast learning module; the initial feature map is processed by depthwise separable convolution using the large-size convolution kernel in the feature extraction module to obtain an intermediate feature map; the contrast between the center pixel and the surrounding pixels is calculated by the multi-scale local contrast learning module according to local sliding windows with different hole rates to generate a multi-scale attention weight map, and the weight map and the intermediate feature map are weighted and fused to output the target feature map; The target feature map is downsampled at least once to obtain a multi-scale feature map; The multi-scale feature map is processed by a decoder to generate prediction results; Using the mask label as a supervision signal, the model is trained based on the difference between the prediction result and the supervision signal to obtain a vascular stenosis detection model.

2. The method for constructing a coronary artery stenosis detection model according to claim 1, characterized in that, The processing steps of the feature extraction module, which includes large kernel convolution and depthwise separable convolution, include: The initial feature map is sequentially subjected to large kernel depthwise separable convolution, channel expansion point convolution, and channel compression point convolution to construct an inverted bottleneck structure; The initial feature map is residually concatenated with the output feature map after channel compression point convolution. The weighted fusion process of the weight map and the intermediate feature map includes: multiplying the weight map and the intermediate feature map element by element, and then performing channel fusion on the multiplied feature map through a point convolutional layer.

3. The method for constructing a coronary artery stenosis detection model according to claim 1, characterized in that, The decoder uses a bilinear interpolation attention module for feature fusion. The multi-scale feature maps are upsampled step by step, and semantic information is filtered out by an attention gate. Then, they are fused with shallow feature maps of the corresponding resolution to generate the prediction result.

4. The method for constructing a coronary artery stenosis detection model according to claim 3, characterized in that, The neural network model also includes a slice-assisted super-segmentation module, which cuts the feature map fused by the decoder into multiple sub-feature maps with a fixed stride, and inputs them into the prediction convolutional layer for independent prediction, thereby enhancing the model's learning of small targets.

5. The method for constructing a coronary artery stenosis detection model according to claim 3, characterized in that, The loss function in the model training process is obtained by weighted summation of the blood vessel prediction loss function and the stenosis prediction loss function, and both the blood vessel prediction loss function and the stenosis prediction loss function adopt a hybrid loss function that combines binary cross-entropy loss and Dice loss.

6. A method for detecting coronary artery stenosis based on a coronary artery stenosis detection model, characterized in that, include: Receive the raw coronary angiography image to be tested; The original coronary angiography image is input into the vascular stenosis detection model trained by the construction method according to any one of claims 1-5 to obtain an initial class activation heatmap; The initial class activation heatmap is subjected to multi-level discrete wavelet transform for frequency domain decoupling to obtain low-frequency sub-bands and high-frequency sub-bands. The low-frequency sub-bands are used to characterize the global anatomical structure information of blood vessels, and the high-frequency sub-bands are used to characterize the gradient abrupt change features of the blood vessel edges. By fusing the features of the low-frequency subband and the high-frequency subband, a final attention heatmap is generated. The target location of coronary artery stenosis is located based on the final attention heatmap.

7. The method for detecting coronary artery stenosis based on a coronary artery stenosis detection model according to claim 6, characterized in that, The multi-level discrete wavelet transform includes a first-order discrete wavelet transform and a second-order discrete wavelet transform: The initial class activation heatmap is decomposed into a first-order low-frequency approximate subband and multiple first-order high-frequency subbands in multiple directions by the first-order discrete wavelet transform. The first-order low-frequency approximation subband is subjected to a discrete wavelet transform again, i.e., a second-order discrete wavelet transform, to obtain a second-order low-frequency approximation subband and a second-order high-frequency subband with a higher level of semantic abstraction.

8. The method for detecting coronary artery stenosis based on a coronary artery stenosis detection model according to claim 7, characterized in that, The multiple directional first-level high-frequency subbands include: Low-frequency to high-frequency components used to capture high-frequency changes in the vertical direction and vertical edge features; High-frequency and low-frequency components used to capture high-frequency changes in the horizontal direction and horizontal edge features; And, high-frequency-high-frequency components used to capture high-frequency variations in the diagonal direction and isolate speckle noise.

9. The method for detecting coronary artery stenosis based on a coronary artery stenosis detection model according to claim 7, characterized in that, The specific steps for fusing the features of the low-frequency subband and the high-frequency subband include: The low-frequency subband, which characterizes the global anatomical topology of blood vessels, is recombined and fused with the high-frequency subband, which characterizes gradient mutations at the vessel margins. The low-frequency subband is used to aggregate discontinuous vascular segments into a coherent vascular tree, providing global topological guidance for stenosis localization. The high-frequency subband is used to provide the precise physical boundaries of stenotic lesions and delineate the edge contours of the stenotic region.

10. The method for detecting coronary artery stenosis based on a coronary artery stenosis detection model according to claim 9, characterized in that, The low-frequency subband is a second-order low-frequency approximate subband obtained by second-order discrete wavelet transform, and the high-frequency subband is a first-order high-frequency subband obtained by first-order discrete wavelet transform.

11. The method for detecting coronary artery stenosis based on a coronary artery stenosis detection model according to claim 7, characterized in that, Before fusion, a noise suppression step is also included: identifying and suppressing speckle noise features in the high-frequency-high-frequency components obtained by first-order discrete wavelet transform, thereby improving the signal-to-noise ratio of the model in complex anatomical backgrounds.

12. A computer device, characterized in that, The device includes: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the method for constructing the coronary artery stenosis detection model according to any one of claims 1-5, and / or the method for detecting coronary artery stenosis based on the coronary artery stenosis detection model according to any one of claims 6-11.

13. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the method for constructing a coronary artery stenosis detection model as described in any one of claims 1-5, and / or the method for detecting coronary artery stenosis based on the coronary artery stenosis detection model as described in any one of claims 6-11.

14. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the method for constructing a coronary artery stenosis detection model according to any one of claims 1-5, and / or the method for detecting coronary artery stenosis based on the coronary artery stenosis detection model according to any one of claims 6-11.

Images