Myocardial tissue recognition method and system for myocardial echocardiography image

Abstract

The application discloses a myocardial acoustic contrast image myocardial tissue recognition method and system, which comprises the following steps: preprocessing an original MCE image; inputting the image into an Unet encoder to extract a multi-scale feature map; optimizing the features through a multi-scale channel cross fusion CCT module, converting the feature maps of various scales into feature tokens and splicing, calculating attention weighted features through a multi-head channel cross attention mechanism, and processing the features through a residual structure and a multi-layer perception machine to obtain an optimized feature map; fusing and reconstructing the optimized feature map and the Unet decoder up-sampling features to output a myocardial tissue segmentation mask; and finally generating auxiliary diagnosis information; the system comprises an MCE image acquisition module, a preprocessing module, an Unet-CCT model inference module, a result output terminal and a data management platform; the CCT module is used for realizing accurate fusion of multi-scale features, significantly improving myocardial tissue segmentation precision and model robustness, realizing automatic recognition, greatly improving diagnosis efficiency, and being suitable for auxiliary diagnosis of ischemic heart disease.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image processing and auxiliary diagnosis of ischemic heart disease, and particularly to a method and system for myocardial tissue recognition in myocardial acoustic contrast images. Background Technology

[0002] Myocardial contrast echocardiography (MCE) is an ultrasound imaging technique that observes the perfusion status of myocardial microcirculation by intravenously injecting contrast agents. It is of great significance for the diagnosis of ischemic heart disease. Currently, myocardial tissue analysis of MCE images mainly relies on the following two methods: Manual image interpretation: Clinicians determine the ischemic area by observing the echo intensity and morphological contour of myocardial tissue in MCE images and combining them with the patient's medical history. The main drawback of this method is that the diagnosis is highly subjective and relies heavily on the doctor's experience. In addition, the image interpretation efficiency is low and the diagnosis of a single patient takes a long time, making it difficult to meet the needs of large-scale physical examinations or rapid screening in primary hospitals. Furthermore, manual image interpretation has limited ability to identify early micro-ischemic lesions, which can easily lead to missed diagnoses or misdiagnoses. Deep learning segmentation techniques based on a single Unet model: This approach utilizes the encoder-decoder structure and skip connections of Unet to attempt automatic segmentation of MCE images. However, when directly splicing shallow and deep features using skip connections, the semantic differences between features at different scales are not considered, resulting in blurred segmentation boundaries. Furthermore, the Unet model lacks robustness to inherent noise, uneven contrast agent distribution, and artifacts in MCE images, exhibiting poor generalization ability. Additionally, the feature fusion method is simplistic, failing to establish dependencies between channels, leading to low feature utilization. The limitations of currently used technologies make it difficult to meet the accuracy and efficiency requirements for myocardial tissue identification using MCE images in clinical diagnosis. Summary of the Invention

[0003] In view of this, the present invention provides a method and system for myocardial tissue recognition in myocardial acoustic contrast images, which solves the problems of strong human subjectivity, low segmentation accuracy of deep learning models and insufficient feature fusion in existing myocardial tissue recognition technologies for MCE images, and achieves accurate segmentation and automated recognition of myocardial tissue, thereby improving the efficiency and accuracy of auxiliary diagnosis of ischemic heart disease.

[0004] To achieve the above objectives, the present invention provides a method for myocardial tissue identification in acoustic contrast images of the myocardium, comprising the following steps: S1. Preprocess the acquired raw myocardial acoustic contrast (MCE) images, including size normalization and image enhancement. S2. Input the preprocessed MCE image into the Unet encoder, and generate multiple skip connection layer feature maps of different scales through multi-layer convolutional kernel pooling operations. S3. Optimize the feature maps of skip connection layers at different scales by cross-fusion of CCT with multi-scale channels; S301, Multi-scale feature embedding: Convolutional pooling is performed on the feature maps of each scale to convert them into feature tokens, and the feature tokens of all scales are concatenated into a total feature sequence. S302, Multi-head channel cross-attention mechanism calculation: Construct the query matrix corresponding to each scale feature, as well as the key matrix and value matrix based on the total feature sequence. Calculate the similarity between each scale feature and the total feature sequence through the multi-head attention mechanism, obtain the attention-weighted features, and fuse them. S303. The attention-weighted features are fused through a residual structure and a multilayer perceptron to obtain optimized features at each scale. S4. The optimized feature maps at each scale are fused with the upsampled features of the Unet decoder. The image resolution is restored through layer-by-layer convolution, and the myocardial tissue segmentation mask is output. S5. Threshold the segmentation mask, calculate the area and mean echo intensity of the myocardial tissue region, and generate an auxiliary diagnostic report.

[0005] Preferably, the preprocessing includes: normalizing the size of the original myocardial acoustic contrast imaging (MCE) image to 512×512 pixels, removing noise through Gaussian filtering, and enhancing the contrast between myocardial tissue and background using histogram equalization.

[0006] Preferably, the Unet encoder generates feature maps of skip connection layers at four scales. The number of channels are respectively , , , .

[0007] Preferably, the query matrix, key matrix, and value matrix are used to calculate the similarity between features at each scale and the total feature sequence through a multi-head attention mechanism to obtain a similarity matrix. The output of the multi-head channel cross-attention is then calculated, expressed as: in, Indicates the first Cross-attention output of features at each scale. Represents a similarity matrix. This represents the value matrix corresponding to the total feature sequence. Indicates matrix transpose. This represents the Softmax function. This indicates a normalization operation. Indicates the first A query matrix for each scale feature. This represents the key matrix corresponding to the total feature sequence. This indicates the total number of channels.

[0008] Preferably, the multi-channel cross-attention mechanism calculates and generates attention-weighted features, expressed as: Wherein, it represents the first The multi-head attention fusion output at each scale represents the number of attention heads. Indicates the first Cross-attention output of each attention head.

[0009] Preferably, the attention-weighted features are fused using a residual structure and a multilayer perceptron to obtain optimized features at each scale, expressed as: in, Indicates the first Features optimized at each scale This indicates multilayer perceptron processing.

[0010] The present invention also provides a myocardial tissue recognition system based on acoustic contrast imaging of myocardium, comprising: The MCE image acquisition module uses clinical ultrasound contrast imaging equipment to acquire contrast echo images of raw myocardial tissue. The preprocessing module is used to preprocess the acquired raw myocardial tissue angiography echo images, completing the size normalization, grayscale correction and noise removal of the raw images; The Unet-CCT model inference module adopts a CPU+GPU heterogeneous architecture and has a built-in pre-trained Unet-CCT fusion model for performing feature extraction, feature optimization and image reconstruction, and outputting a myocardial tissue segmentation mask. The results output terminal is used to visualize the myocardial tissue segmentation mask and the generated auxiliary diagnostic report. The data management platform is used to store raw MCE images, segmentation results, and patient information, and supports data querying and backtracking.

[0011] Preferably, the Unet-CCT model is constructed by embedding a multi-scale channel cross-fusion CCT module in the skip connection layer of the Unet encoder. The CCT module includes a multi-scale feature embedding unit, a multi-head channel cross-attention mechanism loss unit, and a residual output unit.

[0012] Compared with the prior art, the beneficial effects of the present invention are: This invention effectively solves the semantic difference problem caused by the simple splicing of features at different levels in the traditional Unet model by introducing a multi-scale channel cross-fusion (CCT) module. The CCT module uses a multi-head channel cross-attention mechanism to establish the dependency relationship between channels, so that the shallow edge features and deep semantic features can be accurately fused, the boundary segmentation is clearer, and the early small ischemic lesions can be accurately identified. This invention achieves fully automated processing of myocardial tissue identification in MCE images without manual intervention. With the support of a CPU+GPU heterogeneous architecture inference server, the processing time for a single MCE image is reduced to 0.8 seconds, which is about 1,000 times more efficient than traditional manual image reading (about 10-15 minutes / case). This advantage enables it to meet the needs of large-scale physical examination centers and primary hospitals for high-throughput and rapid diagnosis. This invention innovatively introduces instance normalization into the multi-head channel cross-attention mechanism, which effectively eliminates the interference of uneven contrast agent distribution, noise and artifacts on feature extraction, enabling the model to maintain stable recognition performance on image data of different ultrasound devices and different patient groups, and has strong generalization ability. The method provided by this invention can automatically identify and significantly reduce the reliance on the experience level of clinicians, enabling non-senior doctors or medical staff in primary healthcare institutions to obtain high-quality auxiliary diagnostic results. This helps to alleviate the uneven distribution of medical resources and is of great significance for the early screening and early intervention of ischemic heart disease, facilitating the promotion and application of the technology in primary healthcare institutions. Attached Figure Description

[0013] Figure 1 is a flowchart of the overall process of the present invention; Figure 2 This is a system framework diagram of the present invention; Figure 3 This is a schematic diagram of clinical MCE image recognition acquired in an embodiment of the present invention. Detailed Implementation

[0014] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0015] Example 1 This embodiment provides a method for myocardial tissue identification in acoustic contrast imaging of myocardium, including the following steps: S1. Preprocess the acquired raw myocardial acoustic contrast (MCE) images, including size normalization and image enhancement: The original myocardial acoustic contrast (MCE) images were normalized to 512×512 pixels, noise was removed by Gaussian filtering, and histogram equalization was used to enhance the contrast between myocardial tissue and the background.

[0016] S2. Input the preprocessed MCE image into the Unet encoder, and pass it through 4 layers of convolution and pooling operations to generate feature maps of four scales of skip connection layers. The number of channels are respectively , , , .

[0017] S3. Optimize the feature maps of skip connection layers at different scales by cross-fusion of CCT with multi-scale channels; S301, Multi-scale Feature Embedding: Convolutional pooling is performed on feature maps at four different scales, generating images with sizes of [missing data]. The two-dimensional sequence is transformed into a feature token. , , , The feature tokens from all scales are concatenated into a total feature sequence. ; S302, Multi-head channel cross-attention mechanism calculation: A multi-head attention model is used to handle dependencies between channels, thereby optimizing features in each Unet encoding layer: constructing query matrices corresponding to features at each scale. and the key matrix based on the total feature sequence. Sum matrix The expression is: in, Indicates the sequence length. Indicates the number of channels in the hop connection. , , , ; The similarity between features at each scale and the total feature sequence is calculated using a multi-head attention mechanism to obtain a similarity matrix. The output of the multi-channel cross-attention is then calculated, expressed as follows: in, Indicates the first Cross-attention output of features at each scale. Represents a similarity matrix. This represents the value matrix corresponding to the total feature sequence. Indicates matrix transpose. This represents the Softmax function. This indicates a normalization operation. Indicates the first A query matrix for each scale feature. This represents the key matrix corresponding to the total feature sequence. Indicates the total number of channels; In multi-head attention scenarios, the output is fused from the results of multiple heads, expressed as: Wherein, it represents the first The multi-head attention fusion output at each scale represents the number of attention heads. Indicates the first Cross-attention output of each attention head; S303. The attention-weighted features are fused using a residual structure and a multilayer perceptron to obtain optimized features at each scale, expressed as: in, Indicates the first Features optimized at each scale This indicates multilayer perceptron processing.

[0018] S4. Optimize the feature maps at each scale. The image resolution is restored by fusing the upsampled features of the Unet decoder and then outputting a myocardial tissue segmentation mask through layer-by-layer convolution. S5. Threshold the segmentation mask, calculate the area and mean echo intensity of the myocardial tissue region, and generate an auxiliary diagnostic report.

[0019] Example 2 This embodiment provides a myocardial tissue recognition system based on acoustic contrast imaging of the myocardium, including: The MCE image acquisition module uses clinical ultrasound contrast imaging equipment to acquire contrast echo images of raw myocardial tissue. The preprocessing module is used to preprocess the acquired raw myocardial tissue angiography echo images, completing the size normalization, grayscale correction and noise removal of the raw images; The Unet-CCT model inference module adopts a CPU+GPU heterogeneous architecture and has a pre-trained Unet-CCT fusion model built in, which is used to perform feature extraction, feature optimization and image reconstruction, and output a myocardial tissue segmentation mask. The Unet-CCT model is constructed by embedding a multi-scale channel cross-fusion CCT module in the skip connection layer of the Unet encoder. The CCT module includes a multi-scale feature embedding unit, a multi-head channel cross-attention mechanism loss unit and a residual output unit. The results output terminal is used to visualize the myocardial tissue segmentation mask and the generated auxiliary diagnostic report. The data management platform is used to store raw MCE images, segmentation results, and patient information, and supports data querying and backtracking.

[0020] Example 3 This embodiment conducts an experiment based on the method and system for myocardial tissue recognition from acoustic contrast images of myocardium provided in Embodiments 1 and 2. I. Experimental Environment and Parameter Settings (1) Hardware settings: MCE image acquisition equipment: Ultrasonic contrast imaging system; Preprocessing module: Intel Core i7-12700K CPU; Unet-CCT model inference module: Intel Core i9-13900K CPU + NVIDIA RTX 4090 GPU (24G VRAM); Results output terminal: Doctor's workstation monitor.

[0021] (2) Software settings: Operating system: Ubuntu 22.04 LTS; Deep learning framework: PyTorch 2.0; Image processing libraries: OpenCV 4.8.0, PIL 9.5.0; Data management platform: MySQL 8.0 + Django 4.2.

[0022] (3) Model parameter settings: Input image size: 512×512 pixels; Kernel size: 3×3, stride 1, padding 1; Pooling kernel size: 2×2, step size 2; Multi-head attention count N=8; Learning rate: 1e-4, optimizer: AdamW; Training batch size: 16, number of training rounds: 100.

[0023] II. Experimental Procedure Dataset preparation: 1000 clinical MCE image data were collected, of which 800 were used as the training set and 200 were used as the test set; all images were annotated by doctors to represent the myocardial tissue regions as the gold standard.

[0024] Image preprocessing: The training and test set images are subjected to size normalization, Gaussian filtering (kernel size 5×5, standard deviation 1.5), and histogram equalization.

[0025] Model training: (1) Construct a Unet-CCT fusion model and embed the CCT module into the four skip connection layers of Unet; (2) Use the cross-entropy loss function as the training objective; (3) An early stopping strategy is adopted. Training is stopped when the validation set loss does not decrease for 10 consecutive rounds, and the optimal model weights are saved.

[0026] Model testing: Input the test set images into the trained model, output the myocardial tissue segmentation mask, and calculate evaluation metrics such as intersection-over-union (IoU), accuracy, and recall.

[0027] III. Application Validation MCE images of 50 cardiac patients were selected, and myocardial tissue identification was performed using manual image reading, a single Unet model, and the method of this invention. The results showed that the diagnostic accuracy of the method provided by this invention was 8% higher than that of manual image reading, verifying the clinical applicability of the technology.

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for myocardial tissue identification in acoustic contrast imaging of myocardium, characterized in that, Includes the following steps: S1. Preprocess the acquired raw myocardial acoustic contrast (MCE) images, including size normalization and image enhancement. S2. Input the preprocessed MCE image into the Unet encoder, and generate multiple skip connection layer feature maps of different scales through multi-layer convolutional kernel pooling operations. S3. Optimize the feature maps of skip connection layers at different scales by cross-fusion of CCT with multi-scale channels; S301, Multi-scale feature embedding: Convolutional pooling is performed on the feature maps of each scale to convert them into feature tokens, and the feature tokens of all scales are concatenated into a total feature sequence. S302, Multi-head channel cross-attention mechanism calculation: Construct the query matrix corresponding to each scale feature, as well as the key matrix and value matrix based on the total feature sequence. Calculate the similarity between each scale feature and the total feature sequence through the multi-head attention mechanism, obtain the attention-weighted features, and fuse them. S303. The attention-weighted features are fused through a residual structure and a multilayer perceptron to obtain optimized features at each scale. S4. The optimized feature maps at each scale are fused with the upsampled features of the Unet decoder. The image resolution is restored through layer-by-layer convolution, and the myocardial tissue segmentation mask is output. S5. Threshold the segmentation mask, calculate the area of ​​the myocardial tissue region, and generate an auxiliary diagnostic report.

2. The method for myocardial tissue identification in acoustic contrast imaging of myocardium according to claim 1, characterized in that, The preprocessing includes: normalizing the size of the original myocardial acoustic contrast imaging (MCE) image to 512×512 pixels, removing noise through Gaussian filtering, and enhancing the contrast between myocardial tissue and the background through histogram equalization.

3. The method for myocardial tissue identification in acoustic contrast imaging of myocardium according to claim 1, characterized in that, The Unet encoder generates feature maps of skip connection layers at four scales. The number of channels are respectively , , , .

4. The method for myocardial tissue identification in acoustic contrast imaging of myocardium according to claim 1, characterized in that, The query matrix, key matrix, and value matrix are used to calculate the similarity between features at each scale and the total feature sequence through a multi-head attention mechanism to obtain a similarity matrix. The output of the multi-head channel cross-attention is then calculated, expressed as: in, Indicates the first Cross-attention output of features at each scale. Represents a similarity matrix. This represents the value matrix corresponding to the total feature sequence. Indicates matrix transpose. This represents the Softmax function. This indicates a normalization operation. Indicates the first A query matrix for each scale feature. This represents the key matrix corresponding to the total feature sequence. This indicates the total number of channels.

5. The method for myocardial tissue identification in acoustic contrast imaging of myocardium according to claim 4, characterized in that, The multi-channel cross-attention mechanism calculates and generates attention-weighted features, expressed as follows: Wherein, it represents the first The multi-head attention fusion output at each scale represents the number of attention heads. Indicates the first Cross-attention output of each attention head.

6. The method for myocardial tissue recognition in acoustic contrast imaging of myocardium according to claim 5, characterized in that, Attention-weighted features are fused using a residual structure and a multilayer perceptron to obtain optimized features at each scale, expressed as: in, Indicates the first Features optimized at each scale This indicates multilayer perceptron processing.

7. A myocardial tissue recognition system based on acoustic contrast imaging of the myocardium, using the method described in any one of claims 1-6, characterized in that, include: The MCE image acquisition module uses clinical ultrasound contrast imaging equipment to acquire contrast echo images of raw myocardial tissue. The preprocessing module is used to preprocess the acquired raw myocardial tissue angiography echo images, completing the size normalization, grayscale correction and noise removal of the raw images; The Unet-CCT model inference module adopts a CPU+GPU heterogeneous architecture and has a built-in pre-trained Unet-CCT fusion model for performing feature extraction, feature optimization and image reconstruction, and outputting a myocardial tissue segmentation mask. The results output terminal is used to visualize the myocardial tissue segmentation mask and the generated auxiliary diagnostic report. The data management platform is used to store raw MCE images, segmentation results, and patient information, and supports data querying and backtracking.

8. A myocardial tissue recognition system based on acoustic contrast imaging of myocardium according to claim 7, characterized in that, The Unet-CCT model is constructed by embedding a multi-scale channel cross-fusion CCT module in the skip connection layer of the Unet encoder. The CCT module includes a multi-scale feature embedding unit, a multi-head channel cross-attention mechanism loss unit, and a residual output unit.

Images