A method and system for intelligent recognition of cardiac CT images based on convolutional neural networks

Abstract

This invention discloses an intelligent recognition method and system for cardiac CT images based on convolutional neural networks, belonging to the field of medical image intelligent recognition technology. The method includes the following steps: constructing a fusion model of an improved ResNet-12 and a residual hybrid attention module; extracting deep local features from cardiac CT images through the improved ResNet-12 and enhancing the channel and spatial features of lesion areas through the residual hybrid attention module, thereby achieving accurate recognition of multiple types of cardiac diseases; preprocessing the input cardiac CT images, including CT value normalization, size unification, and data augmentation, outputting standardized feature data; and embedding the residual hybrid attention module after the residual module of the improved ResNet-12. This invention ensures generalization ability in scenarios with limited samples in primary healthcare institutions, ultimately improving the efficiency and accuracy of cardiac disease diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image recognition technology, specifically to an intelligent recognition method and system for cardiac CT images based on convolutional neural networks. Background Technology

[0002] Heart diseases, such as coronary artery disease and acute myocardial infarction, have become among the leading causes of death from chronic diseases worldwide. Cardiac CT, due to its non-invasive nature, high resolution, and ability to clearly visualize the coronary artery tree and myocardial structure, has become a core imaging method for screening heart diseases. However, current cardiac CT diagnosis faces three major challenges: 1. Manual diagnosis is difficult and inefficient: Cardiac CT images contain complex structures such as coronary artery branches, such as the left anterior descending artery, circumflex artery, myocardium, and heart chambers. Subtle lesions, such as early coronary soft plaques and myocardial micro-infarction foci, have little gray difference from normal tissue, making them easy to miss if relying on the doctor's experience. A single cardiac CT scan requires reviewing 500-800 slices, with an average diagnosis time of 15-20 minutes, which is difficult to meet the needs of rapid diagnosis in emergency departments. 2. Existing deep learning models lack adaptability: Traditional CNN models such as ResNet and VGGNet have limitations in cardiac CT recognition: they cannot fully extract deep features such as myocardial density gradient and coronary artery stenosis; single-channel / spatial attention modules such as CBAM use pooling operations, which easily lose key diagnostic information such as the jagged texture of coronary artery edges and differences in cardiac chamber wall thickness; moreover, existing models are mostly designed for single diseases, such as only recognizing coronary heart disease, and cannot cover multiple categories of diseases such as cardiac tumors and pericardial effusion, resulting in narrow clinical applicability. Large models require large amounts of data and computational resources as support, and their performance is poor when the dataset is small, thus they are not suitable for the current situation of small datasets.

[0003] 3. Weak integration and generalization ability of methods and devices: Most cardiac CT recognition tools on the market are standalone software that cannot be linked with hospital PACS systems. Doctors need to manually upload images and download reports, which disrupts the clinical workflow. In addition, model training depends on large-scale labeled datasets, usually requiring tens of thousands of samples. However, primary healthcare institutions have insufficient labeled cardiac CT samples, mostly less than a thousand, which leads to a sharp drop in generalization ability and an increase in misdiagnosis rate after model deployment. In summary, there is an urgent need for an innovative model that takes into account "subtle feature capture, global structural correlation, and medical system integration" to resolve the core contradiction between accurate cardiac CT identification and clinical application. Summary of the Invention

[0004] The technical objective of this invention is to address the above-mentioned shortcomings by providing a method and system for intelligent recognition of cardiac CT images based on convolutional neural networks. This method and system can ensure generalization ability in scenarios with limited samples in primary healthcare institutions, ultimately improving the efficiency and accuracy of cardiac disease diagnosis.

[0005] The technical solution adopted by this invention to solve its technical problem is: A method for intelligent recognition of cardiac CT images based on convolutional neural networks, the implementation of which includes the following steps: We constructed an improved ResNet-12 and Residual Hybrid Attention Module (RHAM) fusion model. By improving ResNet-12, we extracted deep local features (such as coronary artery morphology and myocardial density) from cardiac CT images. By using the Residual Hybrid Attention Module (RHAM), we enhanced the channel and spatial features of the lesion area, thus achieving accurate identification of multiple types of cardiac diseases. The input cardiac CT images are preprocessed, including CT value normalization, size uniformity, and data augmentation (random flipping and brightness perturbation), and the output is standardized feature data. The improved ResNet-12 includes residual modules, each of which consists of multiple residual units. It enhances feature expression through BatchNormalization and LeakyReLU activation, and gradually extracts deep local features from cardiac CT. After improving the residual module of ResNet-12, a residual hybrid attention module RHAM is embedded. The residual hybrid attention module RHAM is stacked in the order of "residual channel attention → residual spatial attention". It is connected to the residual through convolution operation to preserve the channel and spatial key information of cardiac lesions. The residual hybrid attention module RHAM outputs a feature map that is processed by global average pooling and Dropout, and then inputs it into a fully connected layer to map it into a low-dimensional feature vector. The Softmax module outputs the category probabilities of eight corresponding heart diseases, such as coronary heart disease and myocardial infarction.

[0006] This method is based on a fusion architecture of improved ResNet-12 and residual hybrid attention mechanism, which is used for the accurate identification of structural abnormalities such as coronary arteries, myocardium, and heart chambers in cardiac CT images. It also supports efficient integration with existing hospital medical information systems (HIS / LIS / PACS) and is suitable for clinical auxiliary diagnosis scenarios in cardiovascular specialties and primary healthcare institutions.

[0007] Furthermore, the preprocessing, CT value normalization, taking into account the HU value characteristics of cardiac tissue, distinguishes non-cardiac areas from cardiac areas, reducing background interference; the normalization formula is as follows: ; Where HU(x,y) represents the pixel values ​​of the original CT image, HU min HU maxThese are the minimum and maximum CT values ​​for the region of interest (ROI) of the heart, respectively, where [L,U] represents the target interval. Then scale it to a uniform size; Data augmentation is achieved by random horizontal flipping and CT value brightness perturbation to avoid model overfitting; CT values ​​are normalized and matched with the grayscale features of myocardium and blood to improve the accuracy of feature extraction.

[0008] Furthermore, the feature extraction of ResNet-12 is improved by optimizing the residual block structure on the basis of the original ResNet-12 and adjusting the residual module so that each group contains residual units. Each residual unit contains convolution, and after convolution, batch normalization and activation function activation are performed to avoid gradient vanishing of low gray-scale features of myocardium. Subsequently, the downsampling is controlled by the convolution stride within the residual block, and the final output high-dimensional feature map is used to capture basic features such as coronary artery branch morphology and myocardial layer thickness.

[0009] Furthermore, Residual Channel Attention Block (RCA) includes: first performing global average pooling on the input feature map to obtain channel identifiers; then capturing the non-linear relationships between channels through convolutional layers; finally activating the channels using an activation function to obtain channel weights; and finally weighting the channel identifiers with the original feature map and fusing them through residual skip connections. The formula for the channel identifiers is as follows: , Where C: number of channels, H: height, W: width, reflecting the characteristic contribution of each channel to "myocardial density" and "coronary artery calcification", the channel weight and residual fusion formula is as follows: ; Residual Spatial Attention Block (RSA) involves: fusing channel information through convolution on the residual channel attention output feature map to obtain a spatial identifier; activating this identifier using an activation function to obtain spatial weights; weighting these weights with the original feature map; and then fusing them through residual skip connections to enhance the spatial localization of the cardiac lesion region. The formula for the spatial identifier is as follows: , Locate areas of coronary artery stenosis and myocardial abnormalities.

[0010] Furthermore, the model is trained using a label-smoothed cross-entropy loss function. This reduces the model's dependence on hard labels and improves generalization ability when the number of cardiac CT samples is limited. The formula for the cross-entropy loss function is as follows: , Where K represents the corresponding type of heart disease. For real labels, To predict probabilities, This is a smoothing coefficient. When the sample size is limited, such as 500-800 slices per disease category, it reduces the model's dependence on labels and improves generalization ability. Furthermore, by using a RESTful interface to interact with the medical information system (PACS / HIS), the system can retrieve cardiac CT images and transmit the identification results (including disease probability and heat map). Regularly use confirmed case data from the medical information system to incrementally train the model, freeze the backbone network weights, fine-tune only the fully connected layers, and continuously optimize the recognition accuracy. Furthermore, interaction with the medical information system (PACS / HIS) is achieved through a RESTful interface, executed in three phases: Data retrieval phase: The trigger condition is that the doctor marks a case of "suspected heart disease" in the PACS system and initiates an identification request; the system retrieves patient information, including age, gender, past history of heart disease, and cardiac CT images, and automatically converts them into JPG format; dual access control using hospital intranet IP whitelist + token dynamic authentication is adopted to prevent medical data leakage; Identification and Result Generation Stage: The preprocessed CT images are input into the improved ResNet-12 model, which outputs a probability distribution containing the disease, such as "coronary heart disease: 94%, normal: 5%, pericardial effusion: 1%". At the same time, a heat map of the lesion area is generated, marking the location of coronary artery stenosis and the range of myocardial abnormalities. The results are encapsulated in structured JSON format, including patient ID, image ID, disease category, probability, and heat map link. Results feedback and model iteration phase: The recognition results are written to the PACS system through a RESTful interface, allowing doctors to view heatmaps and manually correct diagnoses; Confirmed case data (including pathology reports and surgical results) are synchronized from the HIS system monthly for incremental model training (freezing the backbone network and only fine-tuning the fully connected layers) to achieve continuous optimization of model performance.

[0011] This invention also claims a convolutional neural network-based intelligent recognition system for cardiac CT images, comprising: Image preprocessing module: responsible for CT value normalization, size adjustment, format conversion and data enhancement of cardiac CT images, and outputting input data with uniform specifications; Improved ResNet-12 feature extraction module: It contains multiple residual groups and extracts deep local features of cardiac CT such as coronary artery branches and myocardial layer thickness through multi-layer residual convolution and LeakyReLU activation. It also alleviates gradient vanishing in deep networks by using residual skip connections. The Residual Hybrid Attention (RHAM) module is composed of Residual Channel Attention (RCA) and Residual Spatial Attention (RSA). It enhances the channel and spatial features of the cardiac lesion region by fusing convolution operations with residuals, thus avoiding the loss of key information. Classification output module: Through global average pooling, Dropout and fully connected layers, the feature vector is mapped to high dimension, and the probability of multiple types of heart disease is output through Softmax, while generating heat maps of lesion areas; Medical information system integration module: Adopting the RESTful protocol, it realizes the retrieval of cardiac CT images and the return of recognition results, and supports access control with IP whitelist + token authentication; Model iteration module: Supports dataset partitioning, incremental training (freezing the backbone network) and canary release, and regularly optimizes model performance using confirmed case data; This system can implement the methods described above.

[0012] The present invention also claims a cardiac CT image intelligent recognition device based on a convolutional neural network, comprising: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to implement the above method.

[0013] The present invention also claims a computer-readable medium storing computer instructions that, when executed by a processor, enable the implementation of the above-described method.

[0014] Compared with existing technologies, the intelligent recognition method and system for cardiac CT images based on convolutional neural networks of the present invention have the following advantages: 1. Significantly improved recognition accuracy: The improved ResNet-12 network has stronger deep feature extraction capabilities; the residual hybrid attention module avoids the loss of subtle cardiac features, such as coronary soft plaques and myocardial microinfarctions, through convolution + residual skip connections. On the test set, thousands of cardiac CT slices containing multiple diseases achieved a high classification accuracy.

[0015] 2. Strong clinical adaptability: Supports simultaneous identification of multiple common heart diseases, covering mainstream clinical needs; after integration with the PACS system, the diagnosis time for a single case is greatly shortened, and doctors do not need to switch systems, which fits the clinical workflow; heat map visualization helps doctors quickly locate lesions and reduces the difficulty of diagnosis. 3. Excellent generalization ability with small samples: The combination of label smoothing loss and Dropout ensures that the model's accuracy on the test set remains high even with thousands of labeled samples, meeting the deployment needs of primary healthcare institutions with insufficient samples. 4. Data security and compliance: IP whitelist + token authentication complies with the "National Health and Medical Big Data Standards and Security Specifications". Data transmission is completed entirely within the hospital's intranet without external data interaction, ensuring patient privacy and medical data security. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the RHAM-ResNet-12 architecture provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the integration process of the medical information system RHAM-ResNet-12 provided in this embodiment of the invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0018] A method for intelligent recognition of cardiac CT images based on convolutional neural networks, the implementation of which includes the following steps: We construct an improved ResNet-12 and residual hybrid attention module RHAM fusion model. By improving ResNet-12, we extract deep local features (such as coronary artery morphology and myocardial density) from cardiac CT images. By using the residual hybrid attention module RHAM, we enhance the channel and spatial features of the lesion area, and work together to achieve accurate identification of multiple types of heart diseases. The input cardiac CT images are preprocessed, including CT value normalization, size uniformity, and data augmentation (random flipping and brightness perturbation), and the output is standardized feature data. The improved ResNet-12 includes residual modules, each of which consists of multiple residual units. It enhances feature expression through BatchNormalization and LeakyReLU activation, and gradually extracts deep local features from cardiac CT. After improving the residual module of ResNet-12, a residual hybrid attention module RHAM is embedded. The residual hybrid attention module RHAM is stacked in the order of "residual channel attention → residual spatial attention". It is connected to the residual through convolution operation to preserve the channel and spatial key information of cardiac lesions. The residual hybrid attention module RHAM outputs a feature map that is processed by global average pooling and Dropout, and then inputs it into a fully connected layer to map it into a low-dimensional feature vector. The Softmax module outputs the category probability of eight corresponding heart diseases, such as coronary heart disease and myocardial infarction. The model is trained using the label smooth cross-entropy loss function, which reduces the model's dependence on hard labels and improves its generalization ability when the cardiac CT sample is limited. By using a RESTful interface to interact with medical information systems (PACS / HIS), the system can retrieve cardiac CT images and transmit the identification results (including disease probability and heat map). Regularly use confirmed case data from the medical information system to incrementally train the model, freeze the backbone network weights, fine-tune only the fully connected layers, and continuously optimize the recognition accuracy. The improved ResNet-12 uses convolutional kernels for its residual units, performing initial convolution and max pooling downsampling, followed by downsampling controlled by the convolution stride within the residual block, ultimately outputting a high-dimensional feature map. The residual channel attention (RCA) includes: performing global average pooling on the input feature map to obtain channel identifiers, capturing nonlinear relationships between channels through convolutional layers, obtaining channel weights through Sigmoid activation, and then fusing them with the original feature map through residual skip connections. The residual spatial attention (RSA) includes: using convolution to fuse channel information in the RCA output feature map to obtain spatial identifiers, activating them with Sigmoid to obtain spatial weights, weighting them with the original feature map, and then fusing them through residual skip connections to enhance the spatial localization of the cardiac lesion region. The CT value normalization in the preprocessing step matches the grayscale features of myocardium and blood, improving the accuracy of feature extraction. By improving the fusion of ResNet-12 and the residual hybrid attention module, we can simultaneously capture subtle local cardiac lesions, such as coronary plaques and abnormal myocardial density, and correlate them with global structures, such as the correspondence between coronary artery branches and myocardial blood supply areas. We can also automate the entire diagnostic process through integration with medical information systems, shortening the diagnostic cycle. Furthermore, through small sample optimization strategies, we can ensure generalization ability in scenarios with limited samples in primary healthcare institutions, ultimately improving the efficiency and accuracy of cardiac disease diagnosis.

[0019] The technical solution of this method consists of two parts: a cardiac CT recognition network architecture and integration with a medical information system. Specifically: The cardiac CT recognition network architecture's core is an improved ResNet-12 fusion residual hybrid attention module. The overall network architecture process is as follows: image preprocessing, feature extraction, feature extraction enhancement, global average pooling and Dropout, and classification. Detailed steps: First, input preprocessing is performed. Cardiac CT images are obtained from the medical information system. The CT values ​​are then normalized using the following formula: ; Where HU(x,y) represents the pixel values ​​of the original CT image, HU min HU max These are the minimum and maximum CT values ​​for the region of interest (ROI) of the heart, respectively, where [L,U] represents the target interval. Then scale it to a uniform size; Data augmentation is achieved by random horizontal flipping and CT value brightness perturbation to avoid model overfitting.

[0020] The ResNet-12 feature extraction method is improved by optimizing the residual block structure based on the original ResNet-12 and adjusting the residual module so that each group contains residual units. Each residual unit contains convolution, and batch normalization and activation function activation are performed after convolution to avoid gradient vanishing of low grayscale features of myocardium. Subsequently, downsampling is controlled by the convolution stride within the residual block. The final output high-dimensional feature map is used to capture basic features such as coronary artery branch morphology and myocardial layer thickness.

[0021] The Residual Hybrid Attention Module (RHAM) is embedded into the residual connections of ResNet. The module consists of two parts: Residual Channel Attention Block (RCA) and Residual Spatial Attention Block (RSA), stacked in the order "RCA→RSA". The Residual Channel Attention Block (RCA) first performs global average pooling on the input feature map to obtain channel identifiers. The formula for the channel identifiers is as follows: , Where C: number of channels, H: height, W: width, reflecting the characteristic contribution of each channel to "myocardial density" and "coronary artery calcification", the channel weight and residual fusion formula is as follows: .

[0022] Residual Spatial Attention (RSA): The residual channel attention (RCA) output feature map is processed by convolution instead of traditional pooling, fusing channel information to obtain a spatial identifier; the formula for the spatial identifier is as follows: , Locate areas of coronary artery stenosis and myocardial abnormalities.

[0023] Spatial weights are obtained after activation by an activation function. After being weighted with the original feature map, they are also fused using residual skip connections to enhance the spatial localization ability of cardiac lesion areas, such as the stenosis of the left anterior descending artery. The formula for fusion of spatial weights and residuals is as follows: . Classification and Loss Optimization: The RHAM output feature map is subjected to global average pooling to obtain a high-dimensional feature vector, and overfitting is suppressed by Dropout; the fully connected layer maps the feature vector to a low dimension, corresponding to the types of heart diseases to be classified, and outputs the class probability through Softmax; during the training phase, the label smoothed cross-entropy loss function is used, and the formula for the cross-entropy loss function is as follows: , Where K represents the corresponding type of heart disease. For real labels, To predict probabilities, This is a smoothing coefficient. When the sample size is limited, such as 500-800 slices per disease category, it reduces the model's dependence on labels and improves generalization ability. See the attached figure. Figure 1 The improved RHAM-ResNet-12 architecture is shown in the diagram.

[0024] Medical information system integration achieves bidirectional data interaction with hospital PACS / HIS systems through standardized interfaces, and is implemented in three phases: Data retrieval phase: The trigger condition is that the doctor marks a case of "suspected heart disease" in the PACS system and initiates an identification request; the system retrieves patient information, including age, gender, past history of heart disease, and cardiac CT images, and automatically converts it into JPG format; dual access control is adopted using a hospital intranet IP whitelist + token dynamic authentication to prevent medical data leakage.

[0025] Identification and Result Generation Stage: The preprocessed CT images are input into the improved ResNet-12 model, which outputs a probability distribution containing diseases, such as "coronary heart disease: 94%, normal: 5%, pericardial effusion: 1%". At the same time, a heat map of the lesion area is generated, marking the location of coronary artery stenosis and the range of myocardial abnormalities. The results are encapsulated in structured JSON format, including patient ID, image ID, disease category, probability, and heat map link.

[0026] Results feedback and model iteration phase: The recognition results are written to the PACS system through a RESTful interface, allowing doctors to view heatmaps and manually correct diagnoses; Confirmed case data (including pathology reports and surgical results) are synchronized from the HIS system monthly for incremental model training (freezing the backbone network and only fine-tuning the fully connected layers) to achieve continuous optimization of model performance.

[0027] Reference Appendix Figure 2 The flowchart shown is for the integration of the medical information system RHAM-ResNet-12.

[0028] The specific deployment and application methods of this approach are as follows: I. Model Deployment and Environment Configuration Hardware Deployment: A GPU server (configured with NVIDIA A100 24GB VRAM) is deployed within the hospital's intranet, connected to a dedicated storage device for image data storage. It connects to the medical information system server and doctor workstations via the hospital's LAN, ensuring data transmission is completed within a closed network.

[0029] Software configuration: Install the PyTorch deep learning framework, deploy the fusion model, and configure the interface service. Load the training model parameters, initialize modules such as image preprocessing, feature extraction, and classification inference, and establish a database connection with the medical information system.

[0030] Dataset preparation: Import cardiac CT images and divide them into training and validation sets according to a certain ratio. Perform data augmentation on the training set (random horizontal flipping, rotation, brightness adjustment, etc.), and unify the format to obtain usable images.

[0031] II. Model Training and Optimization Training parameter settings: Adam optimizer is used with early stopping enabled; learning rate is exponentially decaying; loss function is label smooth cross-entropy. Training process: In the first stage, the weights of the ResNet-12 backbone network are frozen, and only the RHAM module and fully connected layers are trained to enable the attention module to quickly adapt to cardiac CT features; in the second stage, all parameters are unfrozen and end-to-end training is performed; after each round of training, the accuracy is evaluated on the validation set, and the model weights with the highest accuracy on the validation set are saved. Performance optimization: Model quantization is used to reduce inference latency, thereby reducing the inference time per CT slice; a heatmap generation module is added to visualize lesion areas and enhance the interpretability of results.

[0032] III. Implementation of Medical Information System Integration Interface Development: Developing interfaces based on the Python Flask framework. Image retrieval interface: Receives the patient ID and CT image ID from the PACS system, and returns the converted JPG image and basic patient information; Inference interface: Receives JPG images, calls the model to output disease probabilities and links to heatmaps; Result feedback interface: Writes the recognition results in structured JSON format into the diagnostic report module of the PACS system. Function Embedded: Add a "Cardiac AI Recognition" button to the "Image Diagnosis" interface of the PACS system. After the doctor clicks it, a pop-up window will display the patient information and CT image preview. After confirmation, the recognition will be triggered. After the reasoning is completed, AI diagnostic suggestions will be automatically inserted in the report editing area (such as "It suggests the possibility of coronary heart disease. The lesion area is located in the proximal segment of the left anterior descending artery. It is recommended to combine coronary angiography for further diagnosis"). Access control configuration: Set three levels of access permissions: Doctors can only view the AI ​​results of the cases they have treated; Department administrators are responsible for reviewing the consistency of AI diagnoses (generating a "Conformity Rate between AI and Human Diagnosis" report every month); System administrators are responsible for auditing model updates and operation logs (logs are kept for ≥6 months, including the operator, time, and data ID).

[0033] IV. Clinical Application Process Case initiation: The doctor opens the patient's cardiac CT image in the PACS system, marks it as "suspected coronary heart disease / myocardial infarction", clicks the "Cardiac AI Recognition" button, and the system automatically verifies the doctor's permissions and initiates a request. Automatic reasoning: The method and its apparatus retrieve CT images and perform preprocessing (CT value normalization, size scaling), input the model, and output the disease probability distribution and heat map; if the disease with the highest probability is <80%, a pop-up window prompts "Manual review is recommended". Result Confirmation and Feedback: Doctors can view the AI ​​results and heat map, and manually correct the diagnosis (such as changing "coronary heart disease" to "normal heart"). After confirmation, they can click "Save Report". The system will write the final result to the PACS system and synchronize it to the electronic medical record module of the HIS system. Model iteration: At the end of each month, the method and its device automatically select ≥100 cases where "AI diagnosis and pathological diagnosis are consistent" for incremental model training. After training, the model is updated through gray release (first deployed to one department for testing for 2 weeks, and then fully deployed after no abnormalities) to ensure the stability of clinical use.

[0034] This method addresses the issues of low efficiency and high rate of missed diagnoses of subtle lesions in manual diagnosis of coronary artery lesions, myocardial abnormalities, and cardiac tumors in cardiac CT images. The method uses an improved ResNet-12 as its backbone network, embedding a residual hybrid attention module (RHAM) into its residual mapping structure. This module replaces traditional pooling with convolution operations, reducing the loss of information on key cardiac CT features such as coronary artery calcification edges, myocardial density gradients, and differences in cardiac chamber morphology. Simultaneously, it utilizes a residual skip connection structure to overcome the gradient vanishing limitation of deep networks, accurately capturing subtle cardiac lesion features. A global average pooling layer is designed to simplify model complexity, and a Dropout mechanism is subsequently introduced to alleviate overfitting with small samples. Finally, a label-smoothed cross-entropy loss function is employed to ensure strong generalization ability even with a limited number of cardiac CT samples. Furthermore, this method connects with medical information systems through a standardized interface to automatically retrieve cardiac CT images and identify lesions. It covers eight common heart diseases, including coronary heart disease, myocardial infarction, cardiac myxoma, and pericardial effusion, as well as generate diagnostic reports. This reduces the diagnosis time for a single case to within 30 seconds and increases the diagnostic accuracy rate of primary healthcare institutions to over 96.8%, providing an efficient auxiliary tool for early screening and accurate diagnosis of heart diseases.

[0035] This invention also provides an intelligent recognition system for cardiac CT images based on convolutional neural networks, comprising: Image preprocessing module: responsible for CT value normalization, size adjustment, format conversion and data enhancement of cardiac CT images, and outputting input data with uniform specifications; Improved ResNet-12 feature extraction module: It contains multiple residual groups and extracts deep local features of cardiac CT such as coronary artery branches and myocardial layer thickness through multi-layer residual convolution and LeakyReLU activation. It also alleviates gradient vanishing in deep networks by using residual skip connections. The Residual Hybrid Attention (RHAM) module is composed of Residual Channel Attention (RCA) and Residual Spatial Attention (RSA). It enhances the channel and spatial features of the cardiac lesion region by fusing convolution operations with residuals, thus avoiding the loss of key information. Classification output module: Through global average pooling, Dropout and fully connected layers, the feature vector is mapped to high dimension, and the probability of multiple types of heart disease is output through Softmax, while generating heat maps of lesion areas; Medical information system integration module: Adopting the RESTful protocol, it realizes the retrieval of cardiac CT images and the return of recognition results, and supports access control with IP whitelist + token authentication; Model Iteration Module: Supports dataset partitioning, incremental training (freezing the backbone network), and canary releases, and periodically optimizes model performance using confirmed case data.

[0036] The CT value normalization step of the image preprocessing module distinguishes non-heart areas from heart areas based on the HU value characteristics of heart tissue, thereby reducing background interference.

[0037] The RHAM module is embedded in the residual module of the improved ResNet-12, which captures mid-to-high-level features of cardiac CT, such as coronary artery stenosis and abnormal myocardial density, thereby improving the sensitivity of lesion identification. The classification output module supports the identification of multiple types of heart diseases, including coronary heart disease, myocardial infarction, cardiac myxoma, pericardial effusion, atrial septal defect, ventricular septal defect, aortic dissection, and normal heart. The results returned by the medical information system integration module are in structured JSON format, containing patient ID, image ID, disease category, probability, and heatmap link, and can be directly written into the diagnostic report module of the PACS system.

[0038] This system is capable of implementing the intelligent recognition method for cardiac CT images based on convolutional neural networks described in the above embodiments.

[0039] This invention also provides a cardiac CT image intelligent recognition device based on a convolutional neural network, comprising: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to implement the intelligent recognition method for cardiac CT images based on convolutional neural networks described in the above embodiments.

[0040] This invention also provides a computer-readable medium storing computer instructions. When executed by a processor, the computer instructions cause the processor to perform the intelligent cardiac CT image recognition method based on convolutional neural networks described in the above embodiments. Specifically, a system or apparatus equipped with a storage medium storing software program code that implements the functions of any of the above embodiments can be provided, and the computer (or CPU or MPU) of the system or apparatus can read and execute the program code stored in the storage medium.

[0041] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0042] Storage media embodiments for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0043] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0044] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0045] The present invention has been shown and described in detail above with reference to the accompanying drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above embodiments, those skilled in the art will know that more embodiments of the present invention can be obtained by combining the code review methods in the different embodiments. These embodiments are also within the protection scope of the present invention.

Claims

1. A method for intelligent recognition of cardiac CT images based on convolutional neural networks, characterized in that, The implementation of this method includes the following steps: We construct an improved ResNet-12 and residual hybrid attention module fusion model. By improving ResNet-12, we extract deep local features of cardiac CT images. By using the residual hybrid attention module, we enhance the channel and spatial features of the lesion area, and together we achieve accurate identification of multiple types of cardiac diseases. The input cardiac CT images are preprocessed, including CT value normalization, size uniformity, and data augmentation, and standardized feature data is output. The improved ResNet-12 includes residual modules, each of which consists of multiple residual units. It enhances feature expression through BatchNormalization and LeakyReLU activation, and gradually extracts deep local features from cardiac CT. After improving the residual module of ResNet-12, a residual hybrid attention module is embedded. The residual hybrid attention module is stacked in the order of residual channel attention → residual spatial attention. It is connected to the residual through convolution operation to preserve the channel and spatial key information of cardiac lesions. The residual hybrid attention module outputs a feature map that is processed by global average pooling and Dropout, and then inputs it into a fully connected layer to map it into a low-dimensional feature vector. The softmax function outputs the probability of the corresponding heart disease category.

2. The method for intelligent recognition of cardiac CT images based on convolutional neural networks according to claim 1, characterized in that, The preprocessing CT value normalization is performed to distinguish non-cardiac regions from cardiac regions based on the HU value characteristics of cardiac tissue; the normalization formula is as follows: ； Where HU(x,y) represents the pixel values ​​of the original CT image, HU min HU max These are the minimum and maximum CT values ​​of the region of interest in the heart, respectively, where [L,U] represents the target interval. Then scale it to a uniform size; Data augmentation is achieved by random horizontal flipping and CT value brightness perturbation to avoid model overfitting; CT values ​​are normalized and matched with the grayscale features of myocardium and blood to improve the accuracy of feature extraction.

3. A method for intelligent recognition of cardiac CT images based on convolutional neural networks according to claim 1 or 2, characterized in that, The ResNet-12 feature extraction is improved by optimizing the residual block structure based on the original ResNet-12 and adjusting the residual modules so that each group contains residual units. Each residual unit contains convolution, and batch normalization and activation function activation are performed after convolution to avoid gradient vanishing of low grayscale features of myocardium. Subsequently, downsampling is controlled by the convolution stride within the residual block, and the final output high-dimensional feature map is used to capture basic features including coronary artery branch morphology and myocardial layer thickness.

4. The method for intelligent recognition of cardiac CT images based on convolutional neural networks according to claim 1, characterized in that, Residual channel attention includes: first performing global average pooling on the input feature map to obtain channel identifiers; then capturing the non-linear relationships between channels through convolutional layers; finally activating the channels using an activation function to obtain channel weights; and finally weighting these weighted features with the original feature map and fusing them through residual skip connections. The formula for the channel identifiers is as follows: ， Where C represents the number of channels, H represents the height, and W represents the width, reflecting the characteristic contribution of each channel to myocardial density and coronary artery calcification. The formula for fusing channel weights with residuals is as follows: ； Residual spatial attention includes: fusing channel information through convolution on the residual channel attention output feature map to obtain a spatial identifier; activating the identifier using an activation function to obtain spatial weights; weighting these weights with the original feature map; and then fusing them through residual skip connections to enhance the spatial localization of the cardiac lesion region. The formula for the spatial identifier is as follows: ， Locate areas of coronary artery stenosis and myocardial abnormalities.

5. The method for intelligent recognition of cardiac CT images based on convolutional neural networks according to claim 1, characterized in that, The model is trained using a label-smoothed cross-entropy loss function, which reduces the model's dependence on hard labels and improves generalization ability when the cardiac CT sample is limited. The formula for the cross-entropy loss function is as follows: ， Where K represents the corresponding type of heart disease. For real labels, To predict probabilities, This is the smoothing coefficient.

6. The method for intelligent recognition of cardiac CT images based on convolutional neural networks according to claim 1, characterized in that, By using a RESTful interface to interact with the medical information system, the system can retrieve cardiac CT images and transmit recognition results back. Regularly use confirmed case data from the medical information system to incrementally train the model, freeze the backbone network weights, fine-tune only the fully connected layers, and continuously optimize the recognition accuracy.

7. The method for intelligent recognition of cardiac CT images based on convolutional neural networks according to claim 1, characterized in that, Interaction with the medical information system is achieved through a RESTful interface, executed in three phases: Data retrieval phase: The trigger condition is that the doctor marks a case as "suspected heart disease" in the PACS system and initiates an identification request; the system retrieves patient information, including age, gender, past history of heart disease, and cardiac CT images, and automatically converts it into JPG format; dual access control using a hospital intranet IP whitelist + token dynamic authentication is adopted to prevent medical data leakage; Identification and Result Generation Stage: The preprocessed CT images are input into the improved ResNet-12 model, which outputs a probability distribution containing the disease and generates a heat map of the lesion area, marking the location and abnormal range. The results are encapsulated in structured JSON format, including patient ID, image ID, disease category, probability, and heat map link. Results feedback and model iteration phase: The recognition results are written to the PACS system through a RESTful interface, allowing doctors to view heatmaps and manually correct diagnoses; Confirmed case data are synchronized from the HIS system monthly for incremental model training, enabling continuous optimization of model performance.

8. A cardiac CT image intelligent recognition system based on convolutional neural networks, characterized in that, include: Image preprocessing module: responsible for CT value normalization, size adjustment, format conversion and data enhancement of cardiac CT images, and outputting input data with uniform specifications; Improved ResNet-12 feature extraction module: Includes multiple residual groups, extracts deep local features from cardiac CT scans through multi-layer residual convolution and LeakyReLU activation, and alleviates gradient vanishing in deep networks through residual skip connections; Residual Hybrid Attention Module: Composed of residual channel attention and residual spatial attention stacked together, it enhances the channel and spatial features of the cardiac lesion area through convolution operation and residual fusion, and avoids the loss of key information; Classification output module: Through global average pooling, Dropout and fully connected layers, the feature vector is mapped to high dimension, and the probability of multiple types of heart disease is output through Softmax, while generating heat maps of lesion areas; Medical information system integration module: Adopting the RESTful protocol, it realizes the retrieval of cardiac CT images and the return of recognition results, and supports access control with IP whitelist + token authentication; Model iteration module: Supports dataset partitioning, incremental training and canary release, and regularly optimizes model performance using confirmed case data; The system is capable of implementing the method described in any one of claims 1 to 7.

9. A cardiac CT image intelligent recognition device based on a convolutional neural network, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to implement the method according to any one of claims 1 to 7.

10. A computer-readable medium, characterized in that, The computer-readable medium stores computer instructions that, when executed by a processor, enable the implementation of the method described in any one of claims 1 to 7.

Images