Coronary artery stenosis intelligent diagnosis system and method based on deep learning

Abstract

The invention relates to the technical field of medical information, in particular to a coronary artery stenosis intelligent diagnosis system and method based on deep learning, and realizes efficient and accurate diagnosis through data acquisition, marking, convolutional neural network model construction, evaluation and acceleration modules. The system preprocesses the contrast image, marks the narrow position and degree, constructs a convolutional neural network model combined with an attention degree graph, and optimizes parameters to improve the accuracy; through parallel processing, the diagnosis time is shortened, the problems of an existing diagnosis method are systematically and innovatively solved, and the method has high accuracy, high efficiency and good interpretability and is expected to become an important clinical auxiliary tool for early screening and emergency diagnosis of the coronary heart disease.

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, especially an intelligent diagnosis system and method for coronary artery stenosis based on deep learning. Background Art

[0002] Coronary artery stenosis is one of the main causes of cardiovascular diseases, and its early diagnosis and accurate evaluation are crucial for the treatment and prognosis of patients. Traditional diagnosis of coronary artery stenosis mainly relies on invasive coronary angiography. This method not only has certain risks, but also requires high doctor experience, and the diagnosis results may be subjective and inconsistent.

[0003] With the development of medical imaging technology and artificial intelligence, the diagnosis method of coronary artery stenosis based on deep learning has gradually become a research hotspot. Currently, a variety of methods based on machine learning and deep learning have been proposed for the automatic diagnosis of coronary artery stenosis. These methods usually adopt deep learning models such as convolutional neural networks (CNNs) to automatically extract features from coronary angiography images and perform classification and localization of the stenosis degree.

[0004] However, there are still some key problems in existing intelligent diagnosis methods. First, most methods are insufficient in accuracy when dealing with complex coronary artery morphologies, especially limited in the recognition ability of mild and moderate stenoses. Second, existing models often have the black box problem and lack interpretability, making it difficult to gain the trust of clinicians. Third, many methods face the problem of low computational efficiency in practical applications and cannot meet the needs of real-time diagnosis. In addition, existing systems often lack adaptability to different types of coronary angiography images and have limited generalization ability. Summary of the Invention

[0005] The invention aims to solve the above technical problems and proposes an intelligent diagnosis system and method for coronary artery stenosis based on deep learning. Through innovative model design, data processing and training strategies, the system significantly improves the accuracy, efficiency and interpretability of diagnosis.

[0006] The present invention proposes an intelligent diagnosis system for coronary artery stenosis based on deep learning, including:

[0007] A data acquisition module, used for:

[0008] Obtaining angiography image data of the coronary artery;

[0009] Preprocessing the angiography image data, including steps of grayscale conversion, filtering and normalization, to obtain a sample image to be detected;

[0010] A data annotation module, connected to the data acquisition module, used for:

[0011] Receive the sample image to be detected sent by the data acquisition module;

[0012] Annotate the narrow positions and degrees of the sample image to be detected;

[0013] A model construction module, connected to the data annotation module, is used for:

[0014] Based on the annotated sample image, construct a convolutional neural network model for coronary artery stenosis diagnosis;

[0015] Select a loss function, where the loss function is constructed by combining the attention map mechanism and the cross-entropy loss function;

[0016] An evaluation module, connected to the model construction module, is used for:

[0017] Evaluate the diagnostic accuracy of the convolutional neural network model;

[0018] Based on the evaluation results, optimize the network parameters of the convolutional neural network model;

[0019] An acceleration module, connected to the evaluation module, is used for:

[0020] Reduce the diagnostic time by parallelizing the execution of the diagnostic method of the convolutional neural network model.

[0021] Preferably, the preprocessing steps in the data acquisition module include:

[0022] (1) Grayscale conversion: Perform weighted averaging on the RGB three color channels of each pixel point of the acquired RGB format coronary angiography image, and take the average value as the pixel point value after grayscale conversion;

[0023] (2) Filtering: Obtain the neighborhood average gray value of each pixel point in the image to represent its gray value by performing mean filtering on the image, where the neighborhood is selected as a 5×5 square area;

[0024] (3) Normalization: Subtract 256 from the gray value of each pixel point of the coronary angiography image, then divide by 512, and finally round to form an image from 0 to 1.

[0025] Preferably, the stenosis degree categories include: no stenosis, mild stenosis, moderate stenosis, and severe stenosis, and each stenosis degree category includes two types of data.

[0026] Preferably, the annotation of the narrow positions and degrees of the preprocessed sample image to be detected in the data annotation module is specifically:

[0027] Use machine annotation or manual annotation, and save the annotation results in a standard format for storage as training data. The annotation content includes the coordinates of the starting point of the stenosis, the coordinates of the ending point, and the stenosis degree category.

[0028] Preferably, the model construction module specifically includes:

[0029] The first sub-module is used to construct a convolutional neural network model, obtain data on coronary angiography images and stenosis degree categories, obtain the human body part information corresponding to the coronary angiography images to determine the corresponding stenosis degree category, and use the coronary angiography images as the training set. For no stenosis, mild stenosis, moderate stenosis, and severe stenosis, data training sets are respectively made and used as the input of the convolutional neural network for coronary angiography images, and recognition models of coronary angiography images for each stenosis degree category are trained.

[0030] The second sub-module is used to select the loss function, select a suitable loss function for the convolutional neural network model, and obtain a convolutional neural network model for diagnosing coronary artery stenosis.

[0031] Preferably, the convolutional neural network model for diagnosing coronary artery stenosis in the first sub-module is the Inception-ResNet-v2 convolutional neural network model; the second sub-module constructs the loss function by combining the attention map mechanism and the cross-entropy loss function.

[0032] Preferably, the evaluation module is specifically:

[0033] Obtain the diagnosis result of the coronary angiography image, compare it with the preprocessed coronary angiography image. If the diagnosis result is consistent with the preprocessed result, it is evaluated that the convolutional neural network model is accurately diagnosed. If the diagnosis result is inconsistent with the preprocessed result, it is evaluated that the convolutional neural network model is misdiagnosed. Evaluate the accuracy rate, and continuously optimize the network parameters of the convolutional neural network model according to the accuracy rate.

[0034] Preferably, it further includes:

[0035] A user interaction interface for the interaction between the user and the system. Through the operation interface, the user can input coronary angiography images and obtain diagnosis results.

[0036] A data preprocessing module for preprocessing coronary angiography images to reduce noise and improve image quality.

[0037] A coronary artery stenosis intelligent diagnosis module for using deep learning algorithms to intelligently diagnose the input coronary angiography images to be diagnosed and output diagnosis results.

[0038] A diagnosis result output display module for displaying the obtained diagnosis results.

[0039] Preferably, the coronary artery stenosis intelligent diagnosis module includes:

[0040] A model intelligent diagnosis unit, configured to send an angiography image to be diagnosed into a trained coronary artery stenosis intelligent diagnosis model, and obtain the severity of coronary artery stenosis through the diagnosis result.

[0041] A deep learning-based coronary artery stenosis intelligent diagnosis method includes the following steps:

[0042] S1. Dataset acquisition: Obtain a publicly available coronary angiography dataset, which contains a large number of angiography images;

[0043] S2. Dataset screening and preprocessing: Process the images in the dataset, remove noise, and remove non-lesion areas in the images to obtain angiography data;

[0044] S3. Model design and training: Design a deep learning-based coronary artery stenosis intelligent diagnosis model, model training data, loss function design, model hyperparameter setting, and model evaluation index design;

[0045] S4. Online diagnosis: Input the angiography image to be diagnosed into the trained intelligent diagnosis model, and obtain the diagnosis result through the result of the intelligent diagnosis model running, so as to obtain the severity of coronary artery stenosis;

[0046] Among them, the step S3 includes:

[0047] S31. Design of coronary artery stenosis intelligent diagnosis model: The coronary artery stenosis intelligent diagnosis model is constructed using a convolutional neural network, including the following steps:

[0048] 1) Send the input image into the convolutional layer for feature extraction;

[0049] 2) Perform global average pooling on the feature map extracted after the convolutional layer operation to convert the feature map into a channel descriptor;

[0050] 3) Use a multi-layer perceptron to map the channel descriptor to the category space;

[0051] 4) Finally, generate the diagnosis result classification probability through an activation function;

[0052] S32. Model training data, loss function design, model hyperparameter setting, and model evaluation index design:

[0053] The training data of the model includes the training dataset, validation dataset, and test dataset. For the setting of model hyperparameters, it includes the design of the neural network structure and the selection of the gradient descent optimizer. For the design of model evaluation metrics, it includes testing the model, and obtaining the diagnostic accuracy through the test, where the diagnostic accuracy is calculated from the accuracy, recall rate, and specificity.

[0054] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0055] First of all, from a macroscopic perspective, the present invention constructs a complete intelligent diagnosis ecosystem. This system not only includes the core diagnostic algorithm, but also covers the full-process solution from data acquisition, preprocessing to result display. This system-level innovation greatly improves the overall efficiency and reliability of diagnosis, providing a powerful support tool for clinical practice.

[0056] At the system architecture level, the present invention adopts a modular design, and each functional module works together to form an efficient and flexible diagnostic pipeline. The design of the data acquisition module and the preprocessing module ensures the quality and consistency of the input data, laying a solid foundation for the subsequent diagnostic process. The innovative design of the model construction module and the evaluation module significantly improves the accuracy and reliability of diagnosis. The introduction of the acceleration module effectively solves the problem of computational efficiency, enabling the system to meet the needs of real-time diagnosis. The close cooperation between these modules not only improves the overall performance of the system, but also enhances its scalability and maintainability.

[0057] From the perspective of technological innovation, the present invention has breakthrough improvements in multiple key aspects. In terms of the model structure, an improved Inception-ResNet-v2 network is adopted, which combines the advantages of multi-scale feature extraction and residual learning, significantly improving the model's ability to identify complex coronary artery morphologies. Especially for the diagnosis of mild and moderate stenosis, the system of the present invention shows obvious advantages.

[0058] In solving the problem of the model black box, the present invention innovatively introduces an attention map mechanism. This not only improves the interpretability of the model, but also further enhances the diagnostic accuracy by guiding the model to focus on key regions. This visual diagnostic process helps to gain the trust of clinicians and promotes the wide application of artificial intelligence technology in the medical field.

[0059] In terms of improving the diagnostic efficiency, the present invention shortens the diagnostic time to 0.3 seconds per image through an optimized GPU acceleration strategy and parallel computing technology, which is more than 4 times faster than the traditional method. This high efficiency can not only improve the work efficiency of medical institutions, but also save precious treatment time in emergency situations.

[0060] In addition, the present invention also has innovations in data processing and model training. By adopting advanced data augmentation techniques and adaptive loss functions, the generalization ability of the model is significantly improved. This enables the system to adapt to coronary angiography images under different devices and imaging conditions, greatly expanding its application scope.

[0061] Finally, the system of the present invention demonstrates excellent comprehensive performance in practical applications. Experiments on large-scale clinical datasets show that the diagnostic accuracy of the system reaches 97.8%, and the sensitivity and specificity are 96.9% and 98.4% respectively, and these indicators are significantly better than existing methods. Such highly accurate diagnostic results can effectively reduce misdiagnosis and missed diagnosis, providing more accurate treatment plans for patients.

[0062] In summary, through system-level innovations and breakthroughs in key technologies, the present invention effectively solves the problems existing in existing intelligent diagnostic methods for coronary artery stenosis. Its high accuracy, high efficiency, and good interpretability make it a clinical auxiliary diagnostic tool with great potential, and it is expected to play an important role in multiple scenarios such as early screening of coronary heart disease and rapid emergency diagnosis, making important contributions to improving the diagnosis and treatment level of cardiovascular diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is the overall block diagram of the system of the present invention;

[0064] Figure 2 is the logic block diagram of the data acquisition module of the present invention;

[0065] Figure 3 is the logic block diagram of the data annotation module of the present invention;

[0066] Figure 4 is the logic block diagram of the model construction module of the present invention;

[0067] Figure 5 is the logic block diagram of the evaluation module of the present invention;

[0068] Figure 6 is the logic block diagram of the acceleration module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0069] Please refer to the attached Figure 1-6 , the present invention provides an intelligent diagnostic system and method for coronary artery stenosis based on deep learning. The system includes multiple functional modules and can realize the full-process automation from data acquisition to final diagnosis. The following will describe the detailed implementation manners of the present invention in detail.

[0070] The intelligent coronary artery stenosis diagnosis system based on deep learning of the present invention includes a data acquisition module 1, a data annotation module 2, a model construction module 3, an evaluation module 4, and an acceleration module 5. These modules are connected and interact through data transmission and control signals to jointly complete the intelligent diagnosis task of coronary artery stenosis.

[0071] The data acquisition module 1 is used to acquire angiographic image data of the coronary artery and preprocess these data. Specifically, the data acquisition module 1 first acquires a large number of coronary angiographic images from the hospital's imaging system or public datasets. These images are usually in DICOM format and contain relevant patient information and image data. Subsequently, the data acquisition module 1 preprocesses these original images, including three steps: grayscale conversion, filtering, and normalization.

[0072] In the grayscale conversion step, the system converts the RGB format color image into a grayscale image. This step can reduce the data volume while retaining the main structural information in the image. The filtering step is mainly used to remove noise in the image and improve the image quality. The present invention preferably uses a Gaussian filter, which can effectively smooth the image and retain edge information. The normalization step maps the pixel values of the image to the range of 0-1, which helps the training and inference of subsequent deep learning models.

[0073] After preprocessing, the data acquisition module 1 obtains standardized sample images to be detected. These images are transmitted to the data annotation module 2 for further processing.

[0074] The data annotation module 2 receives the sample images to be detected from the data acquisition module 1 and annotates the stenosis position and degree of these images. The annotation process can be carried out in the way of manual annotation or semi-automatic annotation. In manual annotation, experienced radiologists carefully review each image and mark the stenosis position and degree. Semi-automatic annotation combines a computer-aided diagnosis (CAD) system and manual verification, which can improve the annotation efficiency.

[0075] The annotation content mainly includes the starting coordinates, ending coordinates of the stenosis, and the stenosis degree category. The present invention classifies the stenosis degree into four categories: no stenosis, mild stenosis, moderate stenosis, and severe stenosis. This classification method is consistent with clinical practice and can provide more accurate diagnostic references for doctors. The annotation results are saved in a standard format as the training data of subsequent deep learning models.

[0076] The model construction module 3 is the core part of the present invention and is responsible for constructing a convolutional neural network model for coronary artery stenosis diagnosis. This module first constructs a deep learning model based on the labeled sample images. The present invention adopts an improved Inception-ResNet-v2 network structure, which combines the multi-scale feature extraction ability of the Inception module and the residual learning ability of ResNet, and can effectively extract multi-scale features of coronary artery stenosis.

[0077] During the model training process, the model construction module 3 also needs to select an appropriate loss function. The present invention innovatively proposes a loss function that combines the attention map mechanism and cross-entropy. This loss function can be expressed as:

[0078]

[0079] where C represents the number of stenosis degree categories, y i represents the true label, p i represents the predicted probability, N represents the number of pixels in the feature map, A j represents the value of the attention map, and α is a balance factor. This loss function not only considers the classification accuracy but also introduces the attention map mechanism, which can guide the model to focus on the key areas of coronary artery stenosis and improve the diagnosis accuracy.

[0080] The evaluation module 4 is used to evaluate the diagnostic accuracy of the convolutional neural network model and optimize the model parameters based on the evaluation results. Specifically, the evaluation module 4 will use an independent test set to evaluate the trained model. The evaluation metrics include accuracy, sensitivity, specificity, AUC value, etc. For example, the calculation formula for accuracy is:

[0081]

[0082] where TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative.

[0083] If the evaluation result does not meet the preset threshold (for example, the accuracy is lower than 90%), the evaluation module 4 will trigger the model optimization process. The optimization methods include adjusting the learning rate, increasing the regularization strength, adding data augmentation, etc. This process will be repeated until the model performance meets the requirements.

[0084] The main function of the acceleration module 5 is to reduce the diagnosis time through parallel execution. The present invention adopts a parallel computing method based on GPU to load the convolution operation in the deep learning algorithm onto the GPU accelerator for calculation. For example, for an image with an input size of H×W×3, using 64 3×3 convolution kernels for convolution operation can be expressed as:

[0085]

[0086] Among them, O represents the output feature map, I represents the input image, and K represents the convolutional kernel. Through GPU parallel computing, the computing speed of this process can be significantly improved.

[0087] The preprocessing steps in the data acquisition module 1 include three specific steps: grayscale conversion, filtering, and normalization. In the grayscale conversion step, the system performs weighted averaging on each pixel point of the RGB format coronary angiography image. The specific formula is:

[0088] Gray = 0.299R + 0.587G + 0.114B,

[0089] This formula takes into account the sensitivity of the human eye to different colors and can obtain a visually more natural grayscale image.

[0090] The filtering step uses a 5×5 mean filter. For each pixel point in the image, the average value of all pixels within its surrounding 5×5 area is calculated as the new value of this point. This method can effectively remove Gaussian noise but may cause slight blurring of the image. In practical applications, the filter size can be adjusted according to the image quality and noise level.

[0091] The normalization step maps the image pixel values to the range of 0 to 1. The specific formula is:

[0092]

[0093] where I represents the original pixel value, and I normalized represents the normalized pixel value. This formula assumes that the pixel value range of the original image is 0 - 1023. If the pixel value range of the actual image is different, the parameters in the formula can be adjusted accordingly.

[0094] The present invention classifies the stenosis degree into four categories: no stenosis, mild stenosis, moderate stenosis, and severe stenosis. Each category includes two types of data, corresponding to different stenosis degree ranges. For example, the no stenosis category can include completely normal (0% stenosis) and slight changes (1% - 25% stenosis); mild stenosis can include stenosis degrees of 26% - 50% and 51% - 70%; moderate stenosis can include stenosis degrees of 71% - 90% and 91% - 99%; severe stenosis includes two cases of nearly complete occlusion (99%) and complete occlusion (100%).

[0095] This detailed classification method can provide more accurate diagnostic references for clinicians and help formulate personalized treatment plans. At the same time, from the perspective of machine learning, this classification method can also help the model learn more fine-grained features and improve diagnostic accuracy.

[0096] Through the collaborative work of the above modules, the intelligent coronary artery stenosis diagnosis system based on deep learning of the present invention realizes the full-process automation from data acquisition, preprocessing, model training to final diagnosis. This system can not only improve the diagnosis efficiency and accuracy, but also provide interpretable diagnosis results for doctors, and is expected to play an important role in clinical practice.

[0097] In an embodiment of the present invention, the data annotation module 2 uses machine annotation or manual annotation to annotate the stenosis location and degree of the preprocessed sample image to be detected. The annotation results are stored in a standard format and used as training data. The specific annotation content includes the stenosis start coordinate, end coordinate and stenosis degree category.

[0098] In a preferred embodiment, the present invention adopts a semi-automatic annotation method. First, a computer-aided diagnosis (CAD) system based on traditional image processing technology is used to perform a preliminary analysis on the image and automatically mark the suspicious stenosis areas. Subsequently, experienced radiologists will review and correct these automatically annotated results. This method can significantly improve the annotation efficiency while ensuring the annotation quality.

[0099] During the annotation process, doctors can use a specially designed graphical user interface (GUI) tool. This tool allows doctors to directly draw the stenosis area on the image and select the stenosis degree category through a drop-down menu. To improve the consistency of annotation, the system of the present invention also provides a standardized stenosis degree reference atlas for doctors to consult and compare at any time.

[0100] The annotation results are usually saved in JSON or XML format, and each annotation item contains the following information:

[0101] 1. Image ID: uniquely identifies each coronary angiogram image;

[0102] 2. Stenosis start coordinate: (x1, y1);

[0103] 3. Stenosis end coordinate: (x2, y 2) ;

[0104] 4. Stenosis degree category: 0-3, corresponding to no stenosis, mild stenosis, moderate stenosis and severe stenosis respectively;

[0105] 5. Confidence: the degree of confidence of the annotator in this annotation, ranging from 0-1;

[0106] This structured annotation data is convenient for subsequent model training and verification.

[0107] The model construction module 3 of the present invention includes two key sub-modules: the first sub-module 31 is used for constructing a convolutional neural network model, and the second sub-module 32 is used for selecting a loss function.

[0108] The first sub-module 31 first obtains coronary angiography images and corresponding stenosis degree category data. To improve the diagnostic accuracy of the model, this sub-module also obtains the human body part information corresponding to the coronary angiography images. This is because the coronary arteries in different parts (such as the left anterior descending artery, left circumflex artery, right coronary artery, etc.) may present different morphological characteristics, and considering this information can help the model make more accurate judgments.

[0109] Preferably, the system of the present invention makes the training data set into four sub-data sets according to the stenosis degree categories (no stenosis, mild stenosis, moderate stenosis, and severe stenosis) respectively. This strategy can help the model better learn the characteristics of each stenosis degree. During the training process, the system will adopt the method of balanced sampling to ensure that each batch of training data contains balanced samples of various categories, so as to prevent the model from deviating due to class imbalance.

[0110] The second sub-module 32 is responsible for selecting an appropriate loss function. The present invention innovatively proposes a combined loss function, which combines the cross-entropy loss and the focal loss (FocalLoss). This loss function can effectively handle the class imbalance problem and guide the model to focus on difficult-to-classify samples. The mathematical expression of the loss function is as follows:

[0111] L = -α(1 - p t ) γ log(p t ),

[0112] where p t is the predicted probability of the model for the correct category, α is the category weight factor, and γ is the focusing parameter. In the preferred embodiment of the present invention, α is set to 0.25 and γ is set to 2. These parameter values are obtained through a large number of experiments and can achieve good results in the coronary artery stenosis diagnosis task.

[0113] The first sub-module 31 of the present invention uses Inception-ResNet-v2 as the basic network structure. This network structure combines the multi-scale feature extraction ability of the Inception module and the residual learning ability of ResNet, and is particularly suitable for processing complex features in coronary angiography images.

[0114] In an embodiment of the present invention, the specific structure of the Inception-ResNet-v2 network is as follows:

[0115] 1. Initial convolutional layer: 3 3x3 convolutional layers with a stride of 2;

[0116] 2. 5 Inception-ResNet-A modules;

[0117] 3. Reduction-A module;

[0118] 4. 10 Inception-ResNet-B modules;

[0119] 5. Reduction-B module;

[0120] 6. 5 Inception-ResNet-C modules;

[0121] 7. Average pooling layer;

[0122] 8. Dropout layer (ratio 0.2);

[0123] 9. Fully connected layer (output is the number of stenosis degree categories);

[0124] This network structure can effectively capture multi-scale features in coronary angiography images, and at the same time alleviate the problem of vanishing gradients through residual connections, which is beneficial to training deeper networks.

[0125] The second sub-module 32 adopts a loss function that combines the attention map mechanism and cross-entropy. This loss function not only considers classification accuracy but also introduces the attention map mechanism, which can guide the model to focus on key areas of coronary artery stenosis. The mathematical expression of the loss function is as follows:

[0126]

[0127] Among them, C represents the number of stenosis degree categories, y i represents the true label, p i represents the predicted probability, N represents the number of pixels in the feature map, A j represents the value of the attention map, and λ is a balancing factor. In the preferred embodiment of the present invention, λ is set to 0.1, and this value is determined through repeated experiments, which can achieve a good balance between classification accuracy and attention learning.

[0128] The evaluation module 4 of the present invention is responsible for evaluating the diagnostic accuracy of the convolutional neural network model and optimizing the model parameters based on the evaluation results. The working process of this module is as follows:

[0129] 1. Obtain the diagnostic results of coronary angiography images;

[0130] 2. Compare the diagnostic results with the preprocessed coronary angiography images;

[0131] 3. If the diagnostic results and the preprocessing results are consistent, evaluate that the convolutional neural network model is diagnostically accurate;

[0132] 4. If the diagnostic results and the preprocessing results are inconsistent, evaluate that the convolutional neural network model is diagnostically incorrect;

[0133] 5. Calculate the accuracy rate and continuously optimize the network parameters of the convolutional neural network model according to the accuracy rate;

[0134] In a preferred embodiment of the present invention, the evaluation module 4 uses multiple indicators to comprehensively evaluate the model performance. These indicators include:

[0135] 1. Accuracy:

[0136] 2. Sensitivity:

[0137] 3. Specificity:

[0138] 4. F1 score:

[0139] 5. AUC (Area Under the ROC Curve);

[0140] Among them, TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative.

[0141] The system of the present invention will set a performance threshold. For example, it is required that the accuracy rate is not less than 95%, and both the sensitivity and specificity are not less than 90%. If the model performance does not reach these thresholds, the evaluation module 4 will trigger an optimization process. The optimization methods include but are not limited to:

[0142] 1. Adjust the learning rate: Adopt a learning rate decay strategy, set the initial learning rate to 0.001, and decay by 10% every 10 epochs;

[0143] 2. Increase regularization: Add L2 regularization before the fully connected layer, and set the coefficient to 0.001;

[0144] 3. Data augmentation: Apply data augmentation techniques such as random rotation (±15°), translation (±10%), and scaling (0.9 - 1.1);

[0145] 4. Adjust the batch size: Gradually increase from the initial 32 to 128 to improve the training stability;

[0146] 5. Fine-tune the network structure: Such as increasing or decreasing the number of neurons in certain layers, or adjusting the convolutional kernel size;

[0147] Through this iterative optimization process, the system of the present invention can continuously improve the diagnostic accuracy rate of the model and ultimately reach or even exceed the level of human experts.

[0148] The system of the present invention also includes a user interaction interface, a data preprocessing module, a coronary artery stenosis intelligent diagnosis module and a diagnosis result output display module. These modules together constitute a complete intelligent diagnosis system that can provide doctors and patients with convenient and accurate coronary artery stenosis diagnosis services.

[0149] The user interaction interface is a bridge between the system and the user. In a preferred embodiment of the present invention, the interface adopts an intuitive graphical user interface (GUI) design, including the following main functional areas:

[0150] 1. Image import area: Users can import the coronary angiography images to be diagnosed by dragging or clicking. Multiple image formats such as DICOM, JPG, and PNG are supported.

[0151] 2. Image preview area: displays the imported image and provides basic image operation functions, such as zooming, panning, adjusting brightness and contrast, etc.

[0152] 3. Diagnostic control area: Contains buttons such as start diagnosis, pause, reset, etc., and users can control the diagnostic process.

[0153] 4. Result display area: The diagnosis results are displayed in a graphic and textual manner, including information such as the location of stenosis, degree of stenosis, and confidence level.

[0154] 5. History area: Save and display past diagnostic records to facilitate user comparison and tracking.

[0155] The data preprocessing module is responsible for preprocessing the input coronary angiography images to improve the accuracy of subsequent diagnosis. The system of the present invention adopts a series of advanced image processing technologies in this module, mainly including:

[0156] 1. Noise Removal: Use an adaptive median filter to remove salt and pepper noise from the image while preserving edge details. The filter window size is dynamically adjusted based on the local noise level, ranging from 3x3 to 7x7.

[0157] 2. Contrast enhancement: Adopt the adaptive histogram equalization (CLAHE) algorithm, divide the image into 8x8 grids, perform histogram equalization on each grid separately, and then use bilinear interpolation to merge the results. This method can effectively enhance the local contrast while avoiding excessive enhancement of noise.

[0158] 3. Image segmentation: The coronary arteries are segmented using an improved U-Net network to remove background interference. The network uses ResNet-50 as the backbone network in the encoder part to improve feature extraction capabilities.

[0159] 4. Image normalization: Normalize the image pixel values to the range [-1, 1]. The calculation formula is:

[0160]

[0161] where I is the original pixel value, I min and I max are the minimum and maximum pixel values of the image respectively.

[0162] The intelligent coronary stenosis diagnosis module is the core of the system of the present invention, and is responsible for using deep learning algorithms to perform intelligent diagnosis on the preprocessed coronary angiography images. This module contains a model intelligent diagnosis unit, and its working process is as follows:

[0163] 1. Image input: Input the preprocessed angiography image into the trained intelligent coronary stenosis diagnosis model.

[0164] 2. Feature extraction: The model first extracts the deep features of the image through a multi-layer convolutional neural network. In the preferred embodiment of the present invention, an improved DenseNet structure is adopted, and the features of each layer are fully utilized through dense connection, improving the efficiency of feature extraction.

[0165] 3. Multi-scale analysis: Use the Spatial Pyramid Pooling technology to perform multi-scale analysis on the extracted features to capture stenosis features of different sizes. The pyramid levels are set to 1x1, 2x2, and 4x4.

[0166] 4. Attention mechanism: Introduce channel attention and spatial attention mechanisms to help the model better focus on key regions and features. Channel attention is implemented using the Squeeze-and-Excitation block, and spatial attention is achieved by generating a 2D attention map.

[0167] 5. Classification and regression: Finally, perform classification of the stenosis degree and regression of the stenosis position through a fully connected layer. Softmax activation function is used for classification, and linear activation function is used for regression.

[0168] 6. Post-processing: Use the non-maximum suppression (NMS) algorithm to process multiple detection results, eliminate duplicate detections, and obtain the final diagnosis result.

[0169] The diagnosis result output display module is responsible for presenting the results of the intelligent diagnosis to the user in an intuitive and easy-to-understand manner. In an embodiment of the present invention, the output of this module includes the following:

[0170] 1. Overlay display of the original image and the annotation results, and the stenosis area is marked with a bounding box of different colors, and the color depth indicates the stenosis degree.

[0171] 2. Quantitative description of the stenosis degree, such as 50% stenosis in the proximal part of the left anterior descending artery.

[0172] 3. Confidence score, which shows the degree of confidence of the model in the diagnosis result in percentage form.

[0173] 4. Heatmap display, which visually shows the areas that the model focuses on and helps to improve the interpretability of the diagnosis.

[0174] 5. Diagnostic suggestions, which give preliminary treatment suggestions based on the stenosis degree, such as suggesting coronary angiography or immediate interventional treatment.

[0175] The present invention also provides an intelligent diagnosis method for coronary artery stenosis based on deep learning. The method includes four main steps: dataset acquisition, dataset screening and preprocessing, model design and training, and online diagnosis.

[0176] In the dataset acquisition step (S1), the method of the present invention obtains publicly available coronary angiography datasets from multiple sources. These datasets usually come from large medical institutions or medical imaging research institutions and contain a large number of angiography images. Preferably, the datasets used in the present invention include but are not limited to the following:

[0177] 1. CADIS (Coronary Artery Disease Imaging Sequence) dataset, which contains more than 5000 coronary angiography sequences.

[0178] 2. CAD-RadAI dataset, which contains more than 10000 well-annotated coronary angiography images.

[0179] 3. Self-built clinical dataset, which contains 3000 coronary angiography images from multiple top-three hospitals.

[0180] The dataset screening and preprocessing step (S2) processes the acquired datasets to improve the data quality. Specifically, it includes:

[0181] 1. Data cleaning: Remove low-quality, blurred or incomplete images. Use an image quality assessment algorithm, such as sharpness assessment based on Laplace transform, to eliminate images with scores lower than a threshold (such as 0.5).

[0182] 2. Denoising processing: Apply the BM3D (Block-Matching and 3D filtering) algorithm to remove image noise. This algorithm can effectively remove Gaussian noise and Poisson noise while retaining edge details.

[0183] 3. Contrast Enhancement: Use the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm to enhance the image contrast. The image is segmented into an 8x8 grid, and the contrast limit threshold for each grid is set to 0.01.

[0184] 4. Vessel Enhancement: Apply the Frangi filter to enhance the vascular structure. This filter is based on the eigenvalue analysis of the Hessian matrix and can effectively enhance tubular structures.

[0185] 5. Image Normalization: Normalize the image pixel values to the range [0, 1], which is beneficial for the training and convergence of the model.

[0186] The model design and training step (S3) is the core of the method of the present invention. This step includes two sub-steps: model design (S31) and model training (S32).

[0187] In the model design sub-step (S31), the present invention adopts a deep learning model based on a convolutional neural network. The specific structure of the model is as follows:

[0188] 1. Input Layer: Accept the preprocessed coronary angiography image, with a size of 512x512x1.

[0189] 2. Feature Extraction Layer: Consists of 5 convolutional blocks, each convolutional block is composed of two 3x3 convolutional layers, a batch normalization layer, and a ReLU activation function. The number of convolutional kernels gradually increases from 64 to 512.

[0190] 3. Global Average Pooling Layer: Convert the feature map into a channel descriptor, greatly reducing the number of parameters.

[0191] 4. Fully Connected Layers: Include two fully connected layers, with the number of neurons being 1024 and 512 respectively.

[0192] 5. Output Layer: Use the softmax activation function to output the classification probability of the stenosis degree.

[0193] In the model training sub-step (S32), the present invention adopts the following strategies:

[0194] 1. Data Augmentation: Apply techniques such as random rotation (±15°), translation (±10%), scaling (0.9 - 1.1), and horizontal flipping to expand the training set.

[0195] 2. Loss Function: Use focal loss, and its mathematical expression is:

[0196] FL(p t )=-α t (1 - p t) γ log(p t ),

[0197] where p t is the probability predicted by the model, α t is the class weight, and γ is the focusing parameter. In the present invention, α t is set to 0.25 and γ is set to 2.

[0198] 3. Optimizer: The Adam optimizer is adopted, the initial learning rate is set to 0.001, and it decays by 10% every 50 epochs.

[0199] 4. Batch size: It is set to 32 and dynamically adjusted during training to balance computational efficiency and memory usage.

[0200] 5. Early stopping strategy: Stop training when the loss on the validation set has not improved for 10 consecutive epochs.

[0201] 6. Model ensemble: Five models with different random initializations are trained, and their average prediction results are used during inference.

[0202] The online diagnosis step (S4) applies the trained model to actual diagnosis. The specific process is as follows:

[0203] 1. Image preprocessing: The input angiography image is preprocessed in the same way as the training data.

[0204] 2. Model inference: The preprocessed image is input into the trained model.

[0205] 3. Post-processing: Post-process the model output, including threshold screening (confidence > 0.5) and non-maximum suppression (IoU threshold is 0.5).

[0206] 4. Result interpretation: Convert the model output into clinically understandable diagnostic results, including the stenosis location, degree, and confidence.

[0207] 5. Visualization: Generate a heat map to intuitively show the areas that the model focuses on and improve the interpretability of the diagnosis.

[0208] Through the above steps, the method of the present invention can achieve intelligent diagnosis of coronary artery stenosis and provide strong auxiliary decision-making support for clinicians.

[0209] To verify the superiority of the intelligent coronary artery stenosis diagnosis system and its method based on deep learning of the present invention, we designed a set of simulation experiments to simulate the clinical application scenarios. The experiments used a dataset of 10,000 coronary angiography images from multiple top-three hospitals, and all these data were annotated by senior cardiovascular experts. The experiments were divided into an embodiment of the present invention and two comparative examples, as follows:

[0210] Example 1: The complete system of the present invention was adopted, including a data acquisition module, a data preprocessing module, a model construction module based on the improved Inception-ResNet-v2, an evaluation module, and an acceleration module.

[0211] Comparative Example 1: Traditional machine learning methods were used, and a support vector machine (SVM) combined with manually designed features was used for coronary artery stenosis diagnosis.

[0212] Comparative Example 2: A basic convolutional neural network (CNN) model was adopted, which did not include the attention map mechanism and the adaptive loss function proposed by the present invention.

[0213] The experimental conditions were set as follows:

[0214] 1. Hardware environment: All experiments were carried out on a server equipped with an NVIDIA Tesla V100 GPU.

[0215] 2. Dataset division: 5-fold cross-validation was adopted, and each fold included 8,000 training data and 2,000 test data.

[0216] 3. Evaluation metrics: Diagnostic accuracy, sensitivity, specificity, F1 score, AUC value, diagnostic time, and model size.

[0217] The detection standards and methods are as follows:

[0218] 1. Diagnostic accuracy: Number of correctly diagnosed samples / Total number of samples;

[0219] 2. Sensitivity: True positives / (True positives + False negatives);

[0220] 3. Specificity: True negatives / (True negatives + False positives);

[0221] 4. F1 score: 2 * (Precision * Recall) / (Precision + Recall);

[0222] 5. AUC value: Calculate the area under the ROC curve through the scikit-learn library;

[0223] 6. Diagnostic time: Average diagnostic time for a single image, measured using the time module of Python;

[0224] 7. Model size: the size of the model parameter file in MB;

[0225] Table 1. Comparison of experimental results of Example 1, Comparative Example 1 and Comparative Example 2

[0226]

[0227]

[0228] It can be seen from the experimental results in Table 1 that Example 1 of the present invention is significantly better than the two comparative examples in all indicators. The specific analysis is as follows:

[0229] 1. Diagnostic accuracy: Example 1 achieved a high accuracy of 97.8%, which was 8.6% and 3.3% higher than Comparative Example 1 and Comparative Example 2, respectively. This shows that the system of the present invention can more accurately identify and diagnose coronary artery stenosis, greatly reducing the possibility of misdiagnosis and missed diagnosis.

[0230] 2. Sensitivity and specificity: Example 1 performed well in both indicators, reaching 96.9% and 98.4% respectively. High sensitivity means that the system can effectively detect cases with stenosis, while high specificity means that the system is also accurate in judging normal cases. This balanced performance is crucial for clinical applications.

[0231] 3. F1 score and AUC value: The F1 score of Example 1 is 0.976, and the AUC value reaches 0.993. These two comprehensive indicators further prove the superiority of the system of the present invention. In particular, the AUC value close to 1 shows that the system can maintain excellent diagnostic performance at different decision thresholds.

[0232] 4. Diagnosis time: The average diagnosis time of Example 1 is only 0.3 seconds, which is 4 times faster than that of Comparative Example 1 and nearly 2 times faster than that of Comparative Example 2. This high efficiency is of great significance for improving the work efficiency of medical institutions and alleviating the pressure on medical resources.

[0233] 5. Model size: Although the model size of Example 1 (85 MB) is larger than that of Comparative Example 1 (20 MB), it is smaller than that of Comparative Example 2 (110 MB). Considering that the performance of Example 1 is much better than that of Comparative Example 1, and that the performance is better than that of Comparative Example 2 while maintaining a smaller model size, it indicates that the present invention has achieved a good balance in model design.

[0234] These results fully demonstrate the superiority of the present invention. The high accuracy and balanced sensitivity and specificity indicate that the system can provide reliable diagnostic results and effectively assist doctors in clinical decision-making. The fast diagnosis speed can significantly improve medical efficiency, especially in emergency situations, to save precious treatment time. The reasonable model size makes the system easy to deploy and update, and is suitable for various medical scenarios.

[0235] The best embodiment is Embodiment 1 that adopts the complete system of the present invention. It not only performs best in various indicators, but also embodies the core innovation points of the present invention: the combination of the improved Inception-ResNet-v2 model structure, the attention map mechanism and the adaptive loss function enables the model to better capture the characteristics of coronary artery stenosis; the data preprocessing and enhancement technology improves the generalization ability of the model; the GPU acceleration and the optimized inference process greatly shorten the diagnosis time.

[0236] The comprehensive application of these innovation points enables the system of the present invention to achieve rapid diagnosis while maintaining high accuracy, providing a powerful support tool for clinical practice. In the future, with the accumulation of more clinical data and the further optimization of algorithms, this system is expected to play a greater role in scenarios such as early screening of coronary heart disease and rapid diagnosis in emergency, making important contributions to improving the diagnosis and treatment level of cardiovascular diseases.

[0237] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. The intelligent diagnosis system for coronary artery stenosis based on deep learning is characterized by: include: Data acquisition module, used to: Acquiring angiographic image data of coronary arteries; Preprocessing the angiography image data, including grayscale conversion, filtering and normalization steps, to obtain a sample image to be detected; A data annotation module is connected to the data acquisition module and is used to: Receiving the sample image to be detected sent by the data acquisition module; Annotating the position and degree of stenosis of the sample image to be detected; A model building module, connected to the data annotation module, is used to: Based on the labeled sample images, constructing a convolutional neural network model for diagnosing coronary artery stenosis; Selecting a loss function, wherein the loss function is constructed by combining an attention graph mechanism and a cross entropy loss function; An evaluation module, connected to the model building module, is used to: Evaluating the diagnostic accuracy of the convolutional neural network model; Based on the evaluation results, optimizing the network parameters of the convolutional neural network model; The acceleration module is connected to the evaluation module and is used for: By executing the diagnosis method of the convolutional neural network model in parallel, the diagnosis time is reduced.

2. The system according to claim 1, characterized in that The preprocessing steps in the data acquisition module include: (1) Grayscale: Perform weighted averaging of the three RGB color channels of each pixel of the acquired RGB format coronary angiography image, and take the average value as the pixel value after grayscale processing; (2) Filtering: By performing mean filtering on the image, the average grayscale value of each pixel in the image is obtained to represent its grayscale value, where the domain is selected as a 5×5 square area; (3) Normalization: Subtract 256 from the grayscale value of each pixel in the coronary angiography image, divide it by 512, and finally round it to form an image ranging from 0 to 1.

3. The system according to claim 1, characterized in that The stenosis degree categories include: no stenosis, mild stenosis, moderate stenosis and severe stenosis, and each stenosis degree category includes two types of data.

4. The system according to claim 1, characterized in that The data annotation module specifically annotates the stenosis position and degree of the preprocessed sample image to be detected as follows: Machine or manual labeling is used, and the labeling results are saved in a standard format for storage as training data. The labeling content includes the coordinates of the starting point of the stenosis, the coordinates of the end point, and the category of the degree of stenosis.

5. The system according to claim 1, characterized in that The model building module specifically includes: The first submodule is used for building a convolutional neural network model, obtaining data of coronary angiography images and stenosis degree categories, obtaining information of human body parts corresponding to the coronary angiography images to determine the corresponding stenosis degree categories, and using the coronary angiography images as training sets, respectively preparing data training sets for no stenosis, mild stenosis, moderate stenosis and severe stenosis, and using them as inputs of the convolutional neural network of the coronary angiography images, and training to obtain coronary angiography image recognition models of various stenosis degree categories; The second submodule is used for loss function selection, which selects a suitable loss function of the convolutional neural network model to obtain a convolutional neural network model for diagnosing coronary artery stenosis.

6. The system according to claim 5, characterized in that The convolutional neural network model for diagnosing coronary artery stenosis in the first submodule is the Inception-ResNet-v2 convolutional neural network model; the second submodule constructs a loss function by combining the attention graph mechanism and the cross entropy loss function.

7. The system according to claim 1, characterized in that The evaluation module is specifically: Obtain the diagnostic results of the coronary angiography image and compare them with the preprocessed coronary angiography image. If the diagnostic result is consistent with the preprocessing result, the convolutional neural network model is evaluated to be accurate in diagnosis. If the diagnostic result is inconsistent with the preprocessing result, the convolutional neural network model is evaluated to be wrong in diagnosis. The accuracy is evaluated and the network parameters of the convolutional neural network model are continuously optimized according to the accuracy.

8. The system according to claim 1, characterized in that Also includes: A user interaction interface is used for interaction between the user and the system. Through the operation interface, the user can input coronary angiography images and obtain diagnosis results; A data preprocessing module, used for preprocessing the coronary angiography image to reduce noise and improve image quality; The intelligent diagnosis module for coronary artery stenosis is used to use a deep learning algorithm to perform intelligent diagnosis on the input coronary angiography images to be diagnosed and output the diagnosis results; The diagnosis result output display module is used to display the obtained diagnosis results.

9. The system according to claim 8, characterized in that The coronary artery stenosis intelligent diagnosis module comprises: The model intelligent diagnosis unit is used to send the angiography image to be diagnosed into the trained coronary artery stenosis intelligent diagnosis model, and obtain the severity of coronary artery stenosis through the diagnosis results.

10. An intelligent diagnosis method for coronary artery stenosis based on deep learning, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Dataset acquisition: Obtain the public coronary angiography dataset, which contains a large number of angiography images; S2. Dataset screening and preprocessing: Process the images in the dataset, remove noise, and remove non-lesion areas in the images to obtain angiography data; S3. Model design and training: Design of deep learning-based intelligent diagnosis model for coronary artery stenosis, model training data, loss function design, model hyperparameter setting and model evaluation index design; S4. Online diagnosis: The angiography image to be diagnosed is input into the trained intelligent diagnosis model, and the diagnosis result is obtained through the operation result of the intelligent diagnosis model, thereby obtaining the severity of coronary artery stenosis; Wherein, the step S3 comprises: S31. Design of intelligent diagnosis model for coronary artery stenosis: The intelligent diagnosis model for coronary artery stenosis is constructed using a convolutional neural network, which includes the following steps: 1) Send the input image to the convolution layer for feature extraction; 2) After the convolution layer operation, the feature map extracted is globally averaged and pooled to convert the feature map into a channel descriptor; 3) Use a multi-layer perceptron to map channel descriptors to category space; 4) Finally, an activation function is used to generate the classification probability of the diagnosis result; S32. Model training data, loss function design, model hyperparameter setting and model evaluation indicator design: The training data of the model includes training data set, validation data set and test data set. The model hyperparameter setting includes neural network structure design and gradient descent optimizer selection. The model evaluation indicator design includes testing the model and obtaining the diagnostic accuracy through testing, where the diagnostic accuracy is calculated by accuracy, recall rate and specificity.

Images