A medical image processing method and system based on deep learning

Abstract

The present invention belongs to the field of medical image processing technology, and specifically relates to a medical image processing method and system based on deep learning, which creates a key frame extraction model and uses labeled medical image data to train the key frame extraction model, obtains medical video stream data of the patient's brain and performs data preprocessing, and inputs it into the key frame extraction model; creates an image processing model, and inputs the extracted key frame image containing specified waveform data into the image processing model for key feature recognition; the image processing model performs image segmentation on the key feature area in the key frame image based on the recognized key features; the image processing model extracts and outputs the specified features of the target area separated from the image, formats the output specified feature data, converts it into a specified data format, and exports it. The above process can obtain continuous blood flow data based on the image processing process of TCCD medical imaging data.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a medical image processing method and system based on deep learning. Background Art

[0002] In clinical diagnosis and treatment, with the continuous advancement of science and technology and the continuous development of the medical field, personalized medicine is moving towards a more efficient and precise direction, and is expected to improve the quality and effectiveness of medical services in the future. With the development of medical imaging equipment, the quality of tissue imaging data obtained by medical imaging equipment is getting higher and higher, and it can image the finer structures of patient tissues and lesions. Organ anatomical images obtained through advanced medical imaging technology can objectively reflect tissue structure and pathological changes, providing important assistance to doctors in diagnosis and treatment, from positioning evidence to guiding treatment.

[0003] As the quality of medical imaging improves, a matching medical image processing method is needed. The existing medical image processing and analysis process involves directly extracting lesion features from the medical image data and performing lesion analysis based on the extracted lesion feature data to obtain the corresponding disease condition data.

[0004] However, the accuracy of lesion and condition analysis based solely on feature extraction from medical imaging data is insufficient, and it also fails to capture dynamic changes in patient tissues. For example, in clinical diagnosis and treatment, cerebral blood flow signals are often acquired using transcranial Doppler ultrasound (TCD) and transcranial color Doppler ultrasound (TCCD). TCD utilizes the Doppler effect, emitting high-frequency sound waves through an ultrasound probe. These waves change frequency as they pass through the bloodstream, and this change can be used to calculate blood flow velocity. TCCD combines traditional Doppler ultrasound with color coding technology, detecting blood flow by emitting high-frequency sound waves. Color coding indicates the direction and velocity of blood flow, allowing medical personnel to intuitively identify and assess vascular status. Both TCD and TCCD have their own advantages and disadvantages in acquiring cerebral blood flow. TCD only measures blood flow velocity and lacks detailed structural information, limiting it to velocity data. TCCD, on the other hand, not only provides blood flow velocity information but also displays the anatomical structure of the blood vessels (via color coding), facilitating a more comprehensive assessment. However, it lacks continuous, real-time cerebral blood flow signals.

[0005] Since most clinical diagnosis and treatment departments, such as intensive care units and clinical anesthesia departments, are only equipped with TCCD devices but lack TCD devices, the ability to obtain continuous cerebral blood flow signal monitoring is very limited. How to perform image processing based on TCCD medical imaging data to obtain continuous blood flow data is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of the present invention is to provide a medical image processing method and system based on deep learning, which is used to perform image processing on medical image data based on TCCD to obtain continuous blood flow data.

[0007] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0008] A medical image processing method based on deep learning, comprising the following steps:

[0009] S1: creating a key frame extraction model, training the key frame extraction model using labeled medical image data, and validating the key frame extraction model using a validation set to determine whether specified performance parameters meet specified requirements;

[0010] S2: obtaining medical video stream data of the patient's brain and performing data preprocessing, inputting the preprocessed medical video stream data into the trained key frame extraction model, wherein the key frame extraction model extracts key frame images containing specified waveform data from the medical video stream data of the patient's brain;

[0011] S3: creating an image processing model, and inputting the extracted key frame image containing the specified waveform data into the image processing model, wherein the image processing model recognizes key features in the pre-processed key frame image;

[0012] S4: the image processing model performs image segmentation on the key feature area in the key frame image based on the identified key features, that is, separating the target area in the key frame image from the image;

[0013] S5: The image processing model extracts and outputs designated features from the target area separated from the image, formats the output designated feature data, converts it into a designated data format, and exports it according to the designated data format to obtain continuous tissue data.

[0014] Preferably, the key frame extraction model in step S1 is a convolutional neural network model, and the convolutional neural network model is provided with an input layer, a combination layer of a convolution layer + a pooling layer, a fully connected layer, and an output layer, wherein the combination layer of the convolution layer + the pooling layer includes a convolution layer with a convolution kernel size of 5×5 and a step size of 2, a pooling layer with a pooling kernel size of 3×3 and a step size of 1, and the input of the fully connected layer is 120 nodes and the output is 84 nodes;

[0015] The input layer inputs raw image data of a medical video stream;

[0016] The convolution layer is provided with multiple layers, and a series of convolution kernels in the combination layer of the convolution layer + the pooling layer are used to perform convolution operations on the input image data to obtain local features in the image, and each convolution kernel generates a feature map;

[0017] There are multiple pooling layers, each of which downsamples the feature map output by the convolutional layer above it, reducing the spatial dimension and number of parameters of the feature map while retaining the specified features;

[0018] The fully connected layer globally integrates the features extracted by the combination of the convolution layer and the pooling layer, maps them to the sample label space, and transmits the global integration results to the output layer for output.

[0019] Preferably, the specific process of step S2 is as follows:

[0020] S21: Obtaining medical video stream data of the patient's brain and performing image denoising processing, and inputting the denoised medical video stream data into the trained key frame extraction model;

[0021] S22: The key frame extraction model extracts key frame images containing designated waveform data from the patient's brain medical video stream data, and performs image enhancement processing on the key frame images.

[0022] Preferably, the specific process of denoising the patient's brain medical video stream data in step S21 is:

[0023] S211: decomposing the observation signal in the patient's brain medical video stream image data into wavelet coefficients of different scales by discrete wavelet transform, and selecting an appropriate wavelet basis and decomposition level number;

[0024] S212: Setting a specified threshold, and obtaining a denoising coefficient for the wavelet coefficient based on a soft threshold function;

[0025] S213: reconstructing the denoising coefficient into a denoised signal using inverse wavelet transform;

[0026] S214: De-noising the patient's brain medical video stream data based on the reconstructed denoising signal.

[0027] Preferably, the specific process of performing image enhancement processing on the key frame image in step S22 is as follows:

[0028] S221: Select multiple different Gaussian function standard deviations for the key frame image at multiple scales, and perform Gaussian blur on the key frame image to obtain multiple blurred images L 1( x , y ), L 2( x , y),···, L n ( x , y ), n is the scale quantity;

[0029] S222: For images at each scale L 1( x , y ), L 2( x , y ),···, L n ( x , y ) performs logarithmic operation to obtain the reflection component of each scale, and obtains Log [ R 1( x , y )], Log [ R 2( x , y )],···, Log [ R n ( x , y )];

[0030] S223: Perform weighted averaging on the reflection components of each scale to obtain a weighted average reflection component Log [ R ( x , y )] = Σ Weight ( i )* Log [ Ri ( x , y )];

[0031] S224: The weighted average reflection component Log [ R ( x , y )] is quantized to pixel values ranging from 0 to 255 as the final output key frame image.

[0032] Preferably, the specific process of identifying key features in the pre-processed key frame image in step S3 is as follows:

[0033] S31: Processing continuous key frame images through Gaussian function to construct a continuous Gaussian scale space;

[0034] S32: Calculate the difference between two adjacent Gaussian scale spaces in the continuous Gaussian scale space in step S31, and construct a Gaussian difference pyramid;

[0035] S33: In the constructed Gaussian difference pyramid, a specified pixel point in the key frame image is compared with a specified number of neighboring pixels around it and a specified number of pixels at two scales to determine whether the specified pixel point is an extreme point. If so, the specified pixel point is marked as a candidate key point.

[0036] S34: Positioning the candidate key points, removing low-contrast points and edge response points therefrom, and assigning multiple main directions to each candidate key point after removing the low-contrast points and edge response points therefrom;

[0037] S35: construct a vector of a specified dimension around each candidate key point after removing low-contrast points and edge response points, and calculate the gradient direction histogram of the area around the key point to describe the local image structure.

[0038] Preferably, the specific process of performing image segmentation on the key feature area in the key frame image in step S4 is as follows:

[0039] S41: describing the local image structure based on the result of the gradient direction histogram of the area around the key point in step S35, and generating an initial region or superpixel by using the similarity of the pixels of the image;

[0040] S42: merging adjacent regions to obtain a cut region according to a similarity criterion;

[0041] S43: performing image segmentation based on the acquired segmentation area, and performing thinning, denoising, and hole filling processing on the boundary of the segmented image.

[0042] In a second aspect, a deep learning-based medical image processing system is provided for implementing any one of the deep learning-based medical image processing methods described above, comprising a model creation module, a data acquisition module, a data preprocessing module, a key frame extraction model, an image processing model, and a data formatting module, wherein the data acquisition module is connected to the preprocessing module, the data preprocessing module is connected to the key frame extraction model, the key frame extraction model is connected to the image processing model, and the image processing model is connected to the data formatting module;

[0043] The model creation module is used to create the key frame extraction model and the image processing model;

[0044] The data acquisition module is used to obtain medical video stream data of the patient's brain;

[0045] The data preprocessing module is used to perform data preprocessing on the patient's brain medical video stream data;

[0046] The key frame extraction model is used to extract key frame images containing specified waveform data from the patient's brain medical video stream data;

[0047] The image processing model identifies key features in the pre-processed key frame image and performs image segmentation on the key feature area in the key frame image, that is, separating the target area in the key frame image from the image;

[0048] The data formatting module is used to perform data formatting processing on the output specified feature data, convert it into a specified data format, and export it according to the specified data format.

[0049] Preferably, the data acquisition module is a transcranial color Doppler ultrasound TCCD, which acquires vascular structure image data of the patient's brain.

[0050] The beneficial effects of the present invention include:

[0051] The deep learning-based medical image processing method and system provided by the present invention create a key frame extraction model and train the key frame extraction model using labeled medical image data. Medical video stream data of the patient's brain is obtained and preprocessed before being input into the key frame extraction model. An image processing model is created and the extracted key frame image containing specified waveform data is input into the image processing model for key feature recognition. The image processing model performs image segmentation on the key feature area in the key frame image based on the recognized key features. The image processing model extracts and outputs the specified features of the target area separated from the image, formats the output specified feature data, converts it into a specified data format, and exports it. The above process can obtain continuous blood flow data based on the image processing process of TCCD medical imaging data.

[0052] First, the constructed key frame extraction model sets the input layer, the combination layer of convolution layer + pooling layer, the fully connected layer and the output layer. The combination layer of convolution layer + pooling layer includes a convolution layer with a convolution kernel size of 5×5 and a stride of 2, a pooling layer with a pooling kernel size of 3×3 and a stride of 1, the input of the fully connected layer is 120 nodes, and the output is 84 nodes. Through this key frame extraction model, the key frame images of the patient's brain medical video stream data are accurately extracted, which provides a data basis for the image processing of the subsequent image processing model and ensures the effectiveness of the subsequent image processing process.

[0053] Secondly, the observation signal in the patient's brain medical video stream image data is decomposed into wavelet coefficients of different scales through discrete wavelet transform, and the appropriate wavelet basis and decomposition layer number are selected; the specified threshold is set, and the denoising coefficient is obtained for the wavelet coefficient based on the soft threshold function. The denoising coefficient is reconstructed into a denoised signal using the inverse wavelet transform. The patient's brain medical video stream data is denoised based on the reconstructed denoised signal. This process achieves the effective removal of noise data in the medical video stream image data and improves the accuracy of the results of the subsequent image processing process.

[0054] Again, by selecting multiple different Gaussian function standard deviations for the key frame image at multiple scales, and performing Gaussian blur on the key frame image to obtain multiple blurred images, logarithmic operation is performed on the image at each scale to obtain the reflection component of each scale, and weighted averaging is performed on the reflection component of each scale to obtain the weighted average reflection component. The weighted average reflection component is quantized into pixel values ranging from 0 to 255, thereby achieving enhancement of the local features of the image and further improving the accuracy of the results of the subsequent image processing process.

[0055] Thirdly, continuous key frame images are processed by Gaussian function to construct a continuous Gaussian scale space to construct a Gaussian difference pyramid; the specified pixel point in the key frame image is compared with the specified number of neighboring pixels around it and the specified number of pixels on two scales, the extreme points are marked as candidate key points, the low contrast points and edge response points are removed, and multiple main directions are assigned to each candidate key point; a vector of specified dimension is constructed around each candidate key point, and the gradient direction histogram of the surrounding area of the key point is statistically analyzed to describe the local image structure, thereby achieving accurate identification of key features in the key frame image and improving the accuracy of subsequent image segmentation.

[0056] Finally, the initial regions or superpixels are generated based on the similarity of the image pixels; according to the similarity criterion, the adjacent regions are merged to obtain the cut regions; the image is cut based on the obtained cut regions, and the boundaries of the cut images are refined, denoised and the holes are filled, thus achieving accurate segmentation of medical imaging data, and the specified features of the target areas separated from the image are extracted and continuously output to obtain continuous data. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the process of medical image processing based on deep learning of the present invention.

[0058] Figure 2 Schematic diagram of the process of identifying key features of the present invention.

[0059] Figure 3Schematic diagram of the architecture of the deep learning-based medical image processing system of the present invention. DETAILED DESCRIPTION

[0060] The following is combined with Figures 1 to 3 The present invention is described in further detail:

[0061] Example 1

[0062] See attached Figure 1 As shown, a medical image processing method based on deep learning includes the following steps:

[0063] S1: creating a key frame extraction model, training the key frame extraction model using labeled medical image data, and validating the key frame extraction model using a validation set to determine whether specified performance parameters meet specified requirements;

[0064] S2: Obtaining the patient's brain medical video stream data and performing data preprocessing, inputting the preprocessed medical video stream data into the trained key frame extraction model, and the key frame extraction model extracting the key frame image containing the specified waveform data from the patient's brain medical video stream data;

[0065] S3: creating an image processing model, and inputting the extracted key frame image containing the specified waveform data into the image processing model, so that the image processing model recognizes key features in the pre-processed key frame image;

[0066] S4: The image processing model performs image segmentation on the key feature area in the key frame image based on the identified key features, that is, separating the target area in the key frame image from the image;

[0067] S5: The image processing model extracts and outputs the specified features of the target area separated from the image, formats the output specified feature data, converts it into a specified data format, and exports it according to the specified data format to obtain continuous tissue data.

[0068] By creating a key frame extraction model and training the key frame extraction model with labeled medical image data, the patient's brain medical video stream data is obtained and pre-processed, and then input into the key frame extraction model; an image processing model is created, and the extracted key frame image containing the specified waveform data is input into the image processing model for key feature recognition; the image processing model performs image segmentation on the key feature area in the key frame image based on the recognized key features; the image processing model extracts and outputs the specified features of the target area separated from the image, formats the output specified feature data, converts it into a specified data format, and exports it. The above process can obtain continuous blood flow data based on the image processing process of medical imaging data such as TCD and TCCD.

[0069] In this embodiment, the key frame extraction model in step S1 is a convolutional neural network model, which is provided with an input layer, a combination layer of convolution layer + pooling layer, a fully connected layer and an output layer. The combination layer of convolution layer + pooling layer includes a convolution layer with a convolution kernel size of 5×5 and a step size of 2, a pooling layer with a pooling kernel size of 3×3 and a step size of 1, the input of the fully connected layer is 120 nodes, and the output is 84 nodes.

[0070] The input layer inputs the original image data of the medical video stream. The convolution layer is provided with multiple layers, and a series of convolution kernels in the combination layer of the convolution layer + the pooling layer are used to perform convolution operations on the input image data to obtain local features in the image. Each convolution kernel generates a feature map. There are multiple pooling layers, and each pooling layer downsamples the feature map output by the convolution layer above it, reducing the spatial dimension and number of parameters of the feature map while retaining the specified features. The fully connected layer globally integrates the features extracted by the combination layer of the convolution layer + the pooling layer, maps them to the sample label space, and transmits the global integration results to the output layer for output. Through this key frame extraction model, the key frame images of the patient's brain medical video stream data are accurately extracted, which provides a data basis for the image processing of the subsequent image processing model and ensures the effectiveness of the subsequent image processing process.

[0071] Example 2

[0072] Based on Example 1, the specific process of step S2 is as follows:

[0073] S21: Obtaining medical video stream data of the patient's brain and performing image denoising processing, and inputting the denoised medical video stream data into the trained key frame extraction model;

[0074] S22: The key frame extraction model extracts key frame images containing designated waveform data from the patient's brain medical video stream data, and performs image enhancement processing on the key frame images.

[0075] In this embodiment, the specific process of performing denoising on the patient's brain medical video stream data in step S21 is as follows:

[0076] S211: decomposing the observation signal in the patient's brain medical video stream image data into wavelet coefficients of different scales by discrete wavelet transform, and selecting an appropriate wavelet basis and decomposition level number;

[0077] S212: Setting a specified threshold, and obtaining a denoising coefficient for the wavelet coefficient based on a soft threshold function;

[0078] S213: reconstructing the denoising coefficient into a denoised signal using inverse wavelet transform;

[0079] S214: De-noising the patient's brain medical video stream data based on the reconstructed denoising signal.

[0080] The above-mentioned discrete wavelet transform is used to decompose the observation signal in the patient's brain medical video stream image data into wavelet coefficients of different scales, and an appropriate wavelet basis and decomposition layer number are selected; a specified threshold is set, and a denoising coefficient is obtained for the wavelet coefficient based on a soft threshold function, and the denoising coefficient is reconstructed into a denoised signal using an inverse wavelet transform. The patient's brain medical video stream data is denoised based on the reconstructed denoised signal, thereby achieving effective removal of noise data in the medical video stream image data and improving the accuracy of the results of subsequent image processing processes.

[0081] In this embodiment, the specific process of performing image enhancement processing on the key frame image in step S22 is as follows:

[0082] S221: Select multiple different Gaussian function standard deviations for the key frame image at multiple scales, and perform Gaussian blur on the key frame image to obtain multiple blurred images L 1( x , y ), L 2( x , y ),···, L n ( x , y ), n is the scale quantity;

[0083] S222: For images at each scale L 1( x , y ), L 2( x , y ),···, L n ( x ,y ) performs logarithmic operation to obtain the reflection component of each scale, and obtains Log [ R 1( x , y )], Log [ R 2( x , y )],···, Log [ R n ( x , y )];

[0084] S223: Perform weighted averaging on the reflection components of each scale to obtain a weighted average reflection component Log [ R ( x , y )] = Σ Weight ( i )* Log [ Ri ( x , y )];

[0085] S224: The weighted average reflection component Log [ R ( x , y )] is quantized to pixel values ranging from 0 to 255 as the final output key frame image.

[0086] Example 3

[0087] On the basis of Example 1 or Example 2, see Figure 2 As shown, the specific process of identifying the key features in the pre-processed key frame image in step S3 is as follows:

[0088] S31: Processing continuous key frame images through Gaussian function to construct a continuous Gaussian scale space;

[0089] S32: Calculate the difference between two adjacent Gaussian scale spaces in the continuous Gaussian scale space in step S31, and construct a Gaussian difference pyramid;

[0090] S33: In the constructed Gaussian difference pyramid, a specified pixel point in the key frame image is compared with a specified number of neighboring pixels around it and a specified number of pixels at two scales to determine whether the specified pixel point is an extreme point. If so, the specified pixel point is marked as a candidate key point.

[0091] S34: Positioning the candidate key points, removing low-contrast points and edge response points therefrom, and assigning multiple main directions to each candidate key point after removing the low-contrast points and edge response points therefrom;

[0092] S35: construct a vector of a specified dimension around each candidate key point after removing low-contrast points and edge response points, and calculate the gradient direction histogram of the area around the key point to describe the local image structure.

[0093] In this embodiment, the specific process of performing image segmentation on the key feature area in the key frame image in step S4 is as follows:

[0094] S41: describing the local image structure based on the result of the gradient direction histogram of the area around the key point in step S35, and generating an initial region or superpixel by using the similarity of the pixels of the image;

[0095] S42: merging adjacent regions to obtain a cut region according to a similarity criterion;

[0096] S43: performing image segmentation based on the acquired segmentation area, and performing thinning, denoising, and hole filling processing on the boundary of the segmented image.

[0097] In another implementation of this embodiment, the imported video stream data is extracted frame by frame and the key information area is automatically cut out, thereby ensuring the speed of key frame image extraction and the accuracy of cropping, and providing high-quality image data for the subsequent text recognition module.

[0098] A medical image processing system based on deep learning, used to implement any one of the medical image processing methods based on deep learning, see Figure 3 As shown, it includes a model creation module, a data acquisition module, a data preprocessing module, a key frame extraction model, an image processing model and a data formatting module. The data acquisition module is connected to the preprocessing module, the data preprocessing module is connected to the key frame extraction model, the key frame extraction model is connected to the image processing model, and the image processing model is connected to the data formatting module.

[0099] The model creation module is used to create the key frame extraction model and the image processing model, the data acquisition module is used to obtain the patient's brain medical video stream data, and the data preprocessing module is used to preprocess the patient's brain medical video stream data. The key frame extraction model is used to extract the key frame image containing the specified waveform data in the patient's brain medical video stream data. The image processing model identifies the key features in the preprocessed key frame image and performs image cutting on the key feature area in the key frame image, that is, separates the target area in the key frame image from the image. The data formatting module is used to format the output specified feature data, convert it into a specified data format, and export it according to the specified data format. The data acquisition module is a transcranial color Doppler ultrasound TCCD, which obtains the vascular structure image data of the patient's brain.

[0100] The system also features an optical character recognition module, which, based on a deep learning model, achieves high-precision character recognition and can extract text content and location information, providing the data foundation for subsequent structured data generation. After obtaining the text content and coordinate location, the system splits the data according to specific rules and organizes it into a table format to facilitate subsequent statistics and analysis. After the statistical analysis is complete, the system automatically generates a structured table based on the extracted text and location information. The table supports export in .xlsx format, facilitating further analysis and storage of the data.

[0101] In summary, the medical image processing method and system based on deep learning provided by the present invention constructs a key frame extraction model that sets an input layer, a combination layer of a convolution layer + a pooling layer, a fully connected layer, and an output layer. Through the key frame extraction model, accurate extraction of key frame images of the patient's brain medical video stream data is achieved, providing a data basis for the image processing of the subsequent image processing model, ensuring the effectiveness of the subsequent image processing process. The observation signal in the patient's brain medical video stream image data is decomposed into wavelet coefficients of different scales through discrete wavelet transform; a specified threshold is set, and a denoising coefficient is obtained for the wavelet coefficient based on a soft threshold function, and the denoising coefficient is reconstructed into a denoising signal using an inverse wavelet transform. The patient's brain medical video stream data is denoised based on the reconstructed denoising signal. The process of denoising the patient's brain medical video stream data based on the reconstructed denoising signal achieves effective removal of noise data in the medical video stream image data and improves the accuracy of the results of the subsequent image processing process.

[0102] By applying Gaussian blur to a keyframe image at multiple scales using different standard deviations of Gaussian functions, the keyframe image is Gaussian blurred to generate multiple blurred images. A logarithmic operation is performed on the image at each scale to obtain the reflectance component at each scale, which is then weighted averaged. The weighted average reflectance component is then quantized to a specified range of pixel values. This enhances the image's local features and further improves the accuracy of subsequent image processing. Successive keyframe images are processed using Gaussian functions to construct a Gaussian difference pyramid in a continuous Gaussian scale space. A specified pixel in the keyframe image is compared with a specified number of its surrounding pixels and a specified number of pixels at two scales. Extreme points are marked as candidate keypoints, low-contrast points and edge-response points are removed, and multiple principal directions are assigned to each candidate keypoint. A vector of a specified dimension is constructed around each candidate keypoint, and the gradient direction histogram of the area surrounding the keypoint is used to describe the local image structure. This allows for precise identification of key features in the keyframe image and improves the accuracy of subsequent image segmentation. The initial region or superpixel is generated based on the similarity of the pixels of the image; the adjacent regions are merged according to the similarity criterion to obtain the cutting region; the image is cut based on the obtained cutting region, and the boundaries of the cut image are refined, denoised and the holes are filled, thereby achieving accurate segmentation of medical imaging data, and the specified features of the target area separated from the image are extracted and continuously output to obtain continuous data.

Claims

1. A medical image processing method based on deep learning, characterized in that: The following steps are involved: S1: creating a key frame extraction model, training the key frame extraction model using labeled medical image data, and validating the key frame extraction model using a validation set to determine whether specified performance parameters meet specified requirements; S2: obtaining medical video stream data of the patient's brain and performing data preprocessing, inputting the preprocessed medical video stream data into the trained key frame extraction model, wherein the key frame extraction model extracts key frame images containing specified waveform data from the medical video stream data of the patient's brain; S3: creating an image processing model, and inputting the extracted key frame image containing the specified waveform data into the image processing model, wherein the image processing model recognizes key features in the pre-processed key frame image; S4: the image processing model performs image segmentation on the key feature area in the key frame image based on the identified key features, that is, separating the target area in the key frame image from the image; S5: The image processing model extracts and continuously outputs designated features from the target area separated from the image, formats the output designated feature data, converts it into a designated data format, and exports it according to the designated data format to obtain continuous tissue data; The key frame extraction model in step S1 is a convolutional neural network model, which includes an input layer, a combination layer of convolutional layer + pooling layer, a fully connected layer, and an output layer. The combination layer of convolutional layer + pooling layer includes a convolution layer with a convolution kernel size of 5×5 and a stride of 2, a pooling layer with a pooling kernel size of 3×3 and a stride of 1, and the input of the fully connected layer is 120 nodes and the output is 84 nodes. The input layer inputs raw image data of a medical video stream; The convolution layer is provided with multiple layers, and a series of convolution kernels in the combination layer of the convolution layer + the pooling layer are used to perform convolution operations on the input image data to obtain local features in the image, and each convolution kernel generates a feature map; There are multiple pooling layers, each of which downsamples the feature map output by the convolutional layer above it, reducing the spatial dimension and number of parameters of the feature map while retaining the specified features; The fully connected layer globally integrates the features extracted by the combination of the convolution layer and the pooling layer, maps them to the sample label space, and transmits the global integration results to the output layer for output; The specific process of step S2 is as follows: S21: Obtaining medical video stream data of the patient's brain and performing image denoising processing, and inputting the denoised medical video stream data into the trained key frame extraction model; S22: The key frame extraction model extracts key frame images containing specified waveform data from the patient's brain medical video stream data, and performs image enhancement processing on the key frame images; The specific process of denoising the patient's brain medical video stream data in step S21 is as follows: S211: decomposing the observation signal in the patient's brain medical video stream image data into wavelet coefficients of different scales by discrete wavelet transform, and selecting an appropriate wavelet basis and decomposition level number; S212: Setting a specified threshold, and obtaining a denoising coefficient for the wavelet coefficient based on a soft threshold function; S213: reconstructing the denoising coefficient into a denoised signal using inverse wavelet transform; S214: performing denoising processing on the patient's brain medical video stream data based on the reconstructed denoising signal; The specific process of performing image enhancement processing on the key frame image in step S22 is as follows: S221: Select multiple different Gaussian function standard deviations for the key frame image at multiple scales, and perform Gaussian blur on the key frame image to obtain multiple blurred images L 1( x , y ), L 2( x , y ),···, L n ( x , y ), n is the scale quantity; S222: For images at each scale L 1( x , y ), L 2( x , y ),···, L n ( x , y ) performs logarithmic operation to obtain the reflection component of each scale, and obtains Log [ R 1( x , y )], Log [ R 2( x , y )],···, Log [ R n ( x , y )]; S223: Perform weighted averaging on the reflection components of each scale to obtain a weighted average reflection component Log [ R ( x , y )] =Σ Weight ( i )* Log [ Ri ( x , y )]; S224: The weighted average reflection component Log [ R ( x , y )] is quantized into pixel values ranging from 0 to 255 as the final output key frame image; The specific process of identifying key features in the pre-processed key frame image in step S3 is as follows: S31: Processing continuous key frame images through Gaussian function to construct a continuous Gaussian scale space; S32: Calculate the difference between two adjacent Gaussian scale spaces in the continuous Gaussian scale space in step S31, and construct a Gaussian difference pyramid; S33: In the constructed Gaussian difference pyramid, a specified pixel point in the key frame image is compared with a specified number of neighboring pixels around it and a specified number of pixels at two scales to determine whether the specified pixel point is an extreme point. If so, the specified pixel point is marked as a candidate key point. S34: Positioning the candidate key points, removing low-contrast points and edge response points therefrom, and assigning multiple main directions to each candidate key point after removing the low-contrast points and edge response points therefrom; S35: constructing a vector of a specified dimension around each candidate key point after removing low-contrast points and edge response points, and calculating a gradient direction histogram of the area around the key point to describe the local image structure; The specific process of performing image segmentation on the key feature area in the key frame image in step S4 is as follows: S41: describing the local image structure based on the result of the gradient direction histogram of the area around the key point in step S35, and generating an initial region or superpixel by using the similarity of the pixels of the image; S42: merging adjacent regions to obtain a cut region according to a similarity criterion; S43: performing image segmentation based on the acquired segmentation area, and performing thinning, denoising, and hole filling processing on the boundary of the segmented image.

2. A deep learning-based medical image processing system, used to implement the deep learning-based medical image processing method according to any one of claims 1, characterized in that: It includes a model creation module, a data acquisition module, a data preprocessing module, a key frame extraction model, an image processing model and a data formatting module, wherein the data acquisition module is connected to the preprocessing module, the data preprocessing module is connected to the key frame extraction model, the key frame extraction model is connected to the image processing model, and the image processing model is connected to the data formatting module; The model creation module is used to create the key frame extraction model and the image processing model; The data acquisition module is used to obtain medical video stream data of the patient's brain; The data preprocessing module is used to perform data preprocessing on the patient's brain medical video stream data; The key frame extraction model is used to extract key frame images containing specified waveform data from the patient's brain medical video stream data; The image processing model identifies key features in the pre-processed key frame image and performs image segmentation on the key feature area in the key frame image, that is, separating the target area in the key frame image from the image; The data formatting module is used to perform data formatting processing on the output specified feature data, convert it into a specified data format, and export it according to the specified data format.

3. A medical image processing system based on deep learning according to claim 2, characterized in that: The data acquisition module is a transcranial color Doppler ultrasound TCCD, which obtains vascular structure image data of the patient's brain.

Images