Cine MRI scanning plane automatic classification method and system based on deep learning

Abstract

The invention discloses a deep learning-based cine MRI (Magnetic Resonance Imaging) scanning plane automatic classification method and system, and relates to the technical field of tumor classification. According to the invention, for a multi-scanning plane cine MRI magnetic resonance guided radiotherapy system, on the basis of cracking a preorder encrypted binary file, the output of timing sequence cine MRI in each scanning plane visual angle mode can be recovered, and a foundation is laid for comprehensive and effective analysis of a subsequent tumor target timing sequence motion track; automatic classification and time sequence numbering of a cine MRI scanning plane are realized by using a time sequence deep learning prediction model, and the global space-time correlation characteristics of the cine MRI can be accurately captured; and hysteresis and errors of manual judgment are greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of tumor classification, and in particular to a method and system for automatic classification of cine MRI scan planes based on deep learning. Background Art

[0002] Thoracic and abdominal malignancies such as lung cancer, liver cancer, and rectal cancer have high morbidity and mortality rates, posing a serious threat to the lives and health of the nation. However, due to the influence of organ motion, precise radiotherapy for thoracic and abdominal tumors is fraught with challenges. Magnetic resonance-guided radiotherapy (MRgRT) offers the advantages of no ionizing radiation, high-contrast soft tissue imaging, and integrated radiotherapy equipment. It uses cine MRI technology to monitor the movement of tumors and surrounding tissues in real time and adaptively adjust radiotherapy plans. Existing MRgRT systems (such as Elekta Unity) can acquire time-series cine MRI images of single or multiple orthogonal 2D scanning planes in real time at a fixed sampling frequency during MRgRT treatment. This provides the imaging information foundation for subsequent real-time tracking of tumor targets and adaptive adjustment of radiotherapy plans.

[0003] Compared to single-plane Cine MRI, Cine MRI in multiple orthogonal planes can comprehensively reflect the motion trajectory of the tumor target during treatment, facilitating the implementation of precise adaptive radiotherapy. However, existing MRI-guided radiotherapy systems can only output encrypted binary Cine MRI files, whose timing and scanning plane information must be cracked and restored. Existing processing solutions cannot further distinguish the scanning planes of Cine MRI (i.e., Cine MRI in the axial, sagittal, and coronal scanning planes is cracked and output in a chaotic order), ultimately resulting in the problem of discontinuous temporal sequence of the output Cine MRI in each scanning direction. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for automatic classification of cine MRI scan planes based on deep learning to improve the above technical problems.

[0005] In order to achieve the above-mentioned object of the invention, the embodiment of the present invention provides the following technical solutions:

[0006] A deep learning-based automatic classification method for cine MRI scan planes includes:

[0007] Obtain and crack the binary encrypted cine MRI timing files of the magnetic resonance-guided radiotherapy system under different scanning angles to obtain the cine MRI timing cracked files; the scanning angles include sagittal, coronal, and axial.

[0008] Numbering and preprocessing the cine MRI time sequence cracking file to obtain the preprocessed cine MRI time sequence cracking file;

[0009] Build a time series deep learning prediction model; input the preprocessed cine MRI time series cracking file into the time series deep learning prediction model, and output the time series prediction result;

[0010] Based on the timing prediction results, the preprocessed cine MRI timing cracked file is time-recovered to complete the automatic classification of the cine MRI scanning plane.

[0011] Furthermore, the cine MRI timing cracking file is numbered and preprocessed to obtain the preprocessed cine MRI timing cracking file, including:

[0012] Number the cine MRI time series cracking files and get the corresponding numbering results;

[0013] Perform linear interpolation and normalization on the cine MRI timing crack file to obtain the normalized cine MRI timing crack file;

[0014] The normalized cine MRI time sequence cracking file and its corresponding encoding results are superimposed and collected to obtain the preprocessed cine MRI time sequence cracking file.

[0015] Furthermore, the time series deep learning prediction model includes an input layer, an encoder, and a decoder connected in series; the encoder module includes an STB-Transformer1 submodule, an STB-Transformer2 submodule, and an STB-Transformer3 submodule connected in series; the STB-Transformer1 submodule includes a patch partition layer, a linear embedding layer, and an STB-Transformer unit connected in series; the STB-Transformer2 submodule includes a patch merge1 layer and STB-Transformer unit; the STB-Transformer3 submodule includes a patchmerge2 layer and 3 STB-Transformer units connected in series; each STB-Transformer unit includes a window-self-attention block and a moving window-self-attention block connected in series; the window-self-attention block includes a normalization layer LN1, a W-MSA layer, a normalization layer LN2 and a MLP1 layer; the moving window-self-attention block includes a normalization layer LN3, a SW-MSA layer, a normalization layer LN4 and a MLP2 layer; the encoder includes a first convolution module, a second convolution module and a fully connected module connected in series; the first convolution module includes a first maximum pooling layer and a DWConv1 layer connected in series; the second convolution module includes a second maximum pooling layer and a DWConv2 layer connected in series; the fully connected module includes a flatten layer, a drop out layer and a fully connected layer connected in series.

[0016] Furthermore, the training process of the time series deep learning prediction model includes:

[0017] Get the cine MRI time series cracking training file;

[0018] Input the cine MRI time series cracking training file into the input layer, and output the corresponding cine MRI time series training data;

[0019] Input the cine MRI time series training data into the encoder, and output the corresponding cine MRI time series training encoded data;

[0020] Input the cine MRI time series training coded data into the decoder, and output the corresponding time series prediction training results;

[0021] Based on the time series prediction training results, the weight parameters of the time series deep learning prediction model are adjusted to complete the training of the time series deep learning prediction model.

[0022] Furthermore, the step of inputting the cineMRI time series training data into the encoder and outputting the corresponding cineMRI time series training coded data comprises:

[0023] The MRI temporal training encoding data is input into the STB-Transformer1 submodule, and the global-local MRI temporal training initial features are output;

[0024] The initial features of global-local MRI temporal training are input into the STB-Transformer2 submodule, and the deep features of global-local MRI temporal training are output;

[0025] The global-local MRI temporal training deep features are input into the STB-Transformer3 submodule, and the cine MRI temporal training encoding data is output.

[0026] Furthermore, the training process of the STB-Transformer1 submodule includes:

[0027] Input the MRI temporal training encoding data into the patch partition layer, and output the MRI temporal training segmentation block data;

[0028] Input the MRI time series training segmentation block data into the linear embedding layer, and output the MRI time series training linear data;

[0029] The MRI time series training linear data is input into the STB-Transformer unit, and the global-local MRI time series training initial features are output.

[0030] Furthermore, the MRI time series training linear data is input into the STB-Transformer unit, and the global-local MRI time series training initial features are output, including:

[0031] Input the MRI time series training linear data into the normalization layer LN1, and output the normalized MRI time series training data;

[0032] The normalized MRI time series training data is input into the W-MSA layer, and the local-MRI time series key features are output;

[0033] Add the local-MRI time series key features and the MRI time series training linear data to obtain the local-MRI time series superposition features;

[0034] The local-MRI time series superposition features are input into the normalization layer LN2, and the normalized MRI time series key features are output;

[0035] The normalized MRI time series key features are input into the MLP1 layer, and the MRI time series nonlinear features are output;

[0036] The MRI time series nonlinear features and the local-MRI time series superposition features are added to obtain the key features of local MRI time series training;

[0037] The local MRI temporal training key features are input into the normalization layer LN3, and the normalized local MRI temporal training key features are output;

[0038] The normalized local MRI time series training key features are input into the SW-MSA layer, and the global MRI time series key features are output;

[0039] The global-MRI time series key features and the local MRI time series training key features are added to obtain the global-MRI time series superposition features;

[0040] The global-MRI temporal superposition feature is input to the normalization layer LN4, and the normalized global-MRI temporal superposition feature is output;

[0041] The normalized global-MRI temporal superposition features are input into the MLP2 layer, and the MRI temporal nonlinear key features are output;

[0042] The MRI time series nonlinear key features and the global-MRI time series superposition features are added to obtain the global-local MRI time series training initial features.

[0043] Furthermore, the method of performing time sequence recovery on the pre-processed cine MRI time sequence cracking file based on the time sequence prediction result includes:

[0044] Based on the time series prediction results, the pre-processed cine MRI time series cracking file is divided into data sets of three scanning angles: sagittal, coronal and axial;

[0045] Based on the numbering results, the three data sets are sorted to obtain the sorting results;

[0046] Based on the sorting results, the cine MRI images of each data set are collected and the corresponding dynamic image gif files are generated.

[0047] A deep learning-based automatic classification system for cine MRI scan planes includes:

[0048] The cine MRI acquisition module is used to obtain the binary encrypted cine MRI timing files of the magnetic resonance guided radiotherapy system under different scanning angles and crack them to obtain the cine MRI timing cracked files;

[0049] A numbering preprocessing module is used to number and preprocess the cine MRI time sequence cracking file to obtain the preprocessed cine MRI time sequence cracking file;

[0050] The time series deep learning prediction module is used to input the pre-processed cine MRI time series cracking file into the time series deep learning prediction model and output the time series prediction result;

[0051] The timing recovery module is used to perform timing recovery on the pre-processed cine MRI timing cracking file based on the timing prediction results.

[0052] Furthermore, the time series deep learning prediction module includes:

[0053] A processing unit, for processing the pre-processed cine MRI time series cracking file using the input layer;

[0054] A feature extraction module for encoding the cine MRI time series training data using an encoder;

[0055] The prediction unit is used to predict the cine MRI time series training coded data using a decoder to obtain a time series prediction result.

[0056] The beneficial effects of the present invention are:

[0057] 1. For multi-scanning plane Cine MRI magnetic resonance guided radiotherapy systems, based on the cracking of the previous encrypted binary file, the output of the time-series Cine MRI under each scanning plane viewing mode can be restored, laying the foundation for the subsequent comprehensive and effective analysis of the time-series motion trajectory of the tumor target.

[0058] 2. The temporal deep learning prediction model is used to realize the automatic classification and temporal numbering of cine MRI scanning planes, which can accurately capture the spatiotemporal correlation characteristics of the entire cine MRI domain.

[0059] 3. Greatly reduce the lag and error of manual judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 A flow chart of a method in an embodiment of the present invention;

[0062] Figure 2 This is a structural diagram of a time series deep learning prediction model in an embodiment of the present invention;

[0063] Figure 3 This is a structural diagram of an encoder module in an embodiment of the present invention;

[0064] Figure 4 This is a structural diagram of the STB-Transformer unit in an embodiment of the present invention;

[0065] Figure 5 2 is a system structure diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and represented in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0067] See also Figure 1 This embodiment provides a method for automatically classifying cine MRI scan planes based on deep learning, which includes:

[0068] S1. Obtain and crack the binary encrypted Cine MRI timing files of the magnetic resonance-guided radiotherapy system under different scanning angles to obtain the Cine MRI timing cracking files; the scanning angles include sagittal scanning angle, coronal scanning angle, and axial scanning angle; the Cine MRI timing cracking files use Cine MRI images in the MHA format.

[0069] S2. Numbering and preprocessing the cine MRI time sequence cracking file to obtain the preprocessed cine MRI time sequence cracking file.

[0070] The S2 includes:

[0071] S2-1. Number the cine MRI time sequence cracking files and obtain the corresponding numbering results.

[0072] S2-2. Linear interpolation and normalization are performed on the cine MRI timing cracking file to obtain a normalized cineMRI timing cracking file.

[0073] The corresponding cine MRI image is interpolated to 2mm*2mm pixel interval points using linear interpolation to obtain the interpolated cine MRI image.

[0074] Based on the center of the cine MRI image, the interpolated cine MRI image is cropped to 256*256 or 512*512 pixels to obtain the normalized cine MRI cracked file.

[0075] S2-3. The normalized cine MRI time sequence cracking file and its corresponding encoding result are superimposed and assembled to obtain a preprocessed cine MRI time sequence cracking file.

[0076] The normalized Cine MRI time sequence cracked files are grouped into groups of 256 frames and stacked in a global order. If the number of frames in the last subset is less than 256, the images starting from the first frame are cyclically filled in according to the global order of the images to replenish the number of subset frames to 256. Taking the first, second, and last groups as examples, the global numbers of the images within the groups are: {1, 2, ..., 256}; {257, ..., 512}; {n, n+1, ..., N}. According to the global order of the images, the Python SimpleITK library is used to splice the two-dimensional Cine MRI images of each subset (normalized Cine MRI time sequence cracked files) to obtain the corresponding multi-layer three-dimensional MHA images (preprocessed Cine MRI time sequence cracked files). The size of a single three-dimensional MHA image is 256 (width) * 256 (height) * 256 (number of layers)

[0077] S3. Construct a time series deep learning prediction model; input the preprocessed cine MRI time series cracking file into the time series deep learning prediction model, and output the corresponding time series prediction results. Each data point in the preprocessed cine MRI time series cracking file is sequentially input into the time series deep learning prediction model, and outputs the corresponding time series prediction results for each data point.

[0078] like Figure 2 As shown, the time series deep learning prediction model includes an input layer, an encoder, and a decoder connected in series; the encoder module includes an STB-Transformer1 submodule, an STB-Transformer2 submodule, and an STB-Transformer3 submodule connected in series; Figure 3As shown in Figure 1, the STB-Transformer1 submodule includes a patchpartition layer, a linear embedding layer, and an STB-Transformer unit connected in series; the STB-Transformer2 submodule includes a patch merge1 layer and an STB-Transformer unit connected in series; the STB-Transformer3 submodule includes a patch merge2 layer and three STB-Transformer units connected in series; Figure 4 As shown, each STB-Transformer unit consists of a series of window-self-attention blocks and moving-window-self-attention blocks. The window-self-attention block includes a normalization layer LN1, a W-MSA layer, a normalization layer LN2, and an MLP1 layer. The moving-window-self-attention block includes a normalization layer LN3, a SW-MSA layer, a normalization layer LN4, and an MLP2 layer. The encoder includes a first convolutional module, a second convolutional module, and a fully connected module, connected in sequence. The first convolutional module includes a first max pooling layer and a DWConv1 layer connected in sequence. The second convolutional module includes a second max pooling layer and a DWConv2 layer connected in sequence. The fully connected module includes a flatten layer, a dropout layer, and a fully connected layer connected in sequence. The W-MSA layer uses a window-multi-head attention mechanism, while the SW-MSA layer uses a moving-window-multi-head attention mechanism. Both the MLP1 and MLP2 layers are multilayer perceptrons. The STB-Transformer unit is the Swin Transformer Block, which uses a shifting window mechanism to encode data blocks and capture the local and global spatiotemporal correlations between blocks.

[0079] The input layer has one channel, which inputs the preprocessed, shuffled cine MRI subset images. The dimension is 256*256*256@the number of channels in the input layer. The patch partition layer has a size of 4 and 96 channels, resulting in an output feature dimension of 64*64*64@96 channels. The patch merge1 layer has a size of 4 and 96 channels, resulting in an output feature dimension of 32*32*32@96 channels. The parameters of the first and second max pooling layers are both 4*4*4. The convolution kernel of the DWConv1 layer is 2*2*2@192 channels, resulting in an output feature dimension of 8*8*8@192 channels. The convolution kernel of the DWConv2 layer is 2*2*2@96 channels, resulting in an output feature dimension of 4*4*4@96 channels. The threshold of the dropout layer is 0.5. The output of the fully connected layer is a 256*4 vector.

[0080] The training process of the time series deep learning prediction model includes:

[0081] S3-1. Obtain the cine MRI timing cracking training file;

[0082] S3-2, input the cine MRI time series cracking training file into the input layer, and output the corresponding cine MRI time series training data with a dimension of 256*256*256@1;

[0083] S3-3, inputting the cine MRI time series training data into the encoder, and outputting corresponding cine MRI time series training encoded data;

[0084] The S3-3 includes:

[0085] S3-3-1. Input the MRI temporal training encoded data into the STB-Transformer1 submodule, and output the global-local MRI temporal training initial features;

[0086] The S3-3-1 includes:

[0087] S3-3-11. Input the MRI temporal training encoded data into the patch partition layer, and output the MRI temporal training segmentation block data;

[0088] S3-3-12. Input the MRI time series training segmentation block data into the linear embedding layer, and output the MRI time series training linear data;

[0089] S3-3-13. Input the MRI time series training linear data into the STB-Transformer unit, and output the global-local MRI time series training initial features. The corresponding process includes:

[0090] The MRI time series training linear data is input into the window-self-attention block, and the output is the local MRI time series training key features. The corresponding process is:

[0091] P11, input the MRI time series training linear data into the normalization layer LN1, and output the normalized MRI time series training data;

[0092] P12, input the normalized MRI time series training data into the W-MSA layer, and output the local-MRI time series key features;

[0093] P13, adding the local-MRI time series key features and the MRI time series training linear data to obtain the local-MRI time series superposition features;

[0094] P14, input the local-MRI time series superposition features into the normalization layer LN2, and output the normalized MRI time series key features;

[0095] P15, input the normalized MRI time series key features into the MLP1 layer, and output the MRI time series nonlinear features;

[0096] P16. Add the MRI time series nonlinear features and the local-MRI time series superposition features to obtain the key features of local MRI time series training.

[0097] The key features of local MRI temporal training are input into the moving window-self-attention block, and the initial features of global-local MRI temporal training are output. The corresponding process is:

[0098] P21, input the local MRI temporal training key features into the normalization layer LN3, and output the normalized local MRI temporal training key features;

[0099] P22, input the normalized local MRI time series training key features into the SW-MSA layer, and output the global-MRI time series key features;

[0100] P23, add the global-MRI time series key features and the local MRI time series training key features to obtain the global-MRI time series superposition features;

[0101] P24, input the global-MRI time series superposition feature to the normalization layer LN4, and output the normalized global-MRI time series superposition feature;

[0102] P25, input the normalized global-MRI time series superposition features into the MLP2 layer, and output the MRI time series nonlinear key features;

[0103] P26. Add the MRI time series nonlinear key features and the global-MRI time series superposition features to obtain the global-local MRI time series training initial features.

[0104] Therefore, the formula corresponding to the multi-layer cascade STB-Transformer unit is (taking two-layer cascade as an example, l=1):

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] in, 、 Respectively represent the local-MRI time series superposition feature and the global-MRI time series superposition feature, 、 、 Respectively represent the linear data of MRI time series training, the key features of local MRI time series training, and the initial features of global-local MRI time series training, represents the normalization operation, represents a multilayer perceptron, 、 They represent the window-multi-head attention mechanism and the moving window-multi-head attention mechanism respectively.

[0110] S3-3-2, input the initial features of global-local MRI temporal training into the STB-Transformer2 submodule, and output the deep features of global-local MRI temporal training;

[0111] The S3-3-2 includes:

[0112] S3-3-21. Input the initial features of global-local MRI temporal training into the patch merge1 layer for downsampling, and output the downsampled initial features of global-local MRI temporal training;

[0113] S3-3-22. Input the downsampled global-local MRI temporal training initial features into the STB-Transformer unit, that is, process them using the same method as S3-3-13, and output the global-local MRI temporal training deep features.

[0114] S3-3-3. Input the global-local MRI temporal training deep features into the STB-Transformer3 submodule, and output the cine MRI temporal training encoding data.

[0115] The S3-3-3 includes:

[0116] S3-3-31. Input the global-local MRI temporal training deep features into the patch merge2 layer for downsampling, and output the downsampled global-local MRI temporal training deep features;

[0117] S3-3-32. Input the downsampled global-local MRI temporal training deep features into three STB-Transformer units in series in sequence. Each STB-Transformer unit uses the same method as S3-3-13 to output the cine MRI temporal training encoding data.

[0118] S3-4, inputting the cine MRI time series training coded data into the decoder, and outputting the corresponding time series prediction training results;

[0119] The S3-4 includes:

[0120] S3-4-1. Input the cineMRI temporal training encoded data into the first convolution module, and output the first cineMRI temporal training decoded data. The corresponding process is:

[0121] Input the cine MRI temporal training coded data into the first maximum pooling layer, and output the pooled cine MRI temporal training coded data;

[0122] The pooled cine MRI temporal training encoded data is input into the DWConv1 layer, and the output is the cine MRI temporal training decoded data.

[0123] S3-4-2. Input the first cine MRI time series training decoded data into the second convolution module, and output the second cine MRI time series training decoded data. The corresponding process is:

[0124] The first cine MRI temporal training decoded data is input into the second maximum pooling layer, and the output is the second pooled cine MRI temporal training encoded data;

[0125] The cine MRI temporal training encoded data after secondary pooling is input into the DWConv2 layer, and the secondary cine MRI temporal training decoded data is output.

[0126] S3-4-3. Input the secondary cine MRI time series training decoding data into the fully connected module, and output the time series prediction training results. The corresponding process is:

[0127] Input the secondary cine MRI temporal training decoding data into the flatten layer, and output the flattened cineMRI temporal training decoding data;

[0128] The flattened cine MRI temporal training decoding data is input into the drop-out layer, and the regularized cine MRI temporal training decoding data is output;

[0129] The regularized cine MRI time series training decoding data is input into the fully connected layer, and the time series prediction training results are output.

[0130] S3-5. Based on the time series prediction training results, adjust the weight parameters of the time series deep learning prediction model to complete the training of the time series deep learning prediction model.

[0131] The temporal deep learning prediction model can perform deep feature extraction on cine MRI data, effectively capturing global and local spatiotemporal correlation characteristics. It can process cine MRI data from different scanning perspectives and adapt to images from multiple perspectives, such as sagittal, coronal, and axial, providing comprehensive data support for subsequent tumor motion trajectory analysis. It allows processing and prediction of data of different scales and complexities, and can be adjusted and optimized according to actual needs, with good flexibility and scalability.

[0132] S4. Based on the time series prediction results, the pre-processed Cine MRI time series cracking file is time-series restored to complete the automatic classification of the Cine MRI scanning plane. The time series prediction results are the 2-4 bit column vector results of each data in the pre-processed Cine MRI time series cracking file.

[0133] The S4 includes:

[0134] S4-1. Based on the time series prediction results, the pre-processed cine MRI time series cracking file is divided into data sets of three scanning angles: sagittal, coronal, and axial;

[0135] S4-2. Sort the three data sets based on the numbering results to obtain a sorting result;

[0136] S4-3. Based on the sorting results, the cine MRI images of each data set are collected and corresponding dynamic image gif files are generated.

[0137] This method, targeting multi-plane Cine MRI-guided radiotherapy systems, decrypts encrypted binary files and recovers the time-series Cine MRI output for each scan plane's viewing angle, laying the foundation for comprehensive and effective analysis of the temporal motion trajectory of tumor targets. Leveraging a temporal deep learning prediction model, it automatically classifies and sequentially numbers Cine MRI scan planes, accurately capturing the global spatiotemporal correlation characteristics of Cine MRI and significantly reducing the lag and errors associated with manual judgment, thereby improving the accuracy and efficiency of tumor radiotherapy.

[0138] like Figure 5 As shown, a cine MRI scan plane automatic classification system based on deep learning

[0139] In summary, the present invention comprises:

[0140] The cine MRI acquisition module is used to obtain the binary encrypted cine MRI timing files of the magnetic resonance guided radiotherapy system under different scanning angles and crack them to obtain the cine MRI timing cracked files;

[0141] A numbering preprocessing module is used to number and preprocess the cine MRI time sequence cracking file to obtain the preprocessed cine MRI time sequence cracking file;

[0142] The time series deep learning prediction module is used to input the pre-processed cine MRI time series cracking file into the time series deep learning prediction model and output the time series prediction result;

[0143] The timing recovery module is used to perform timing recovery on the pre-processed cine MRI timing cracking file based on the timing prediction results.

[0144] The time series deep learning prediction module includes:

[0145] A processing unit, for processing the pre-processed cine MRI time series cracking file using the input layer;

[0146] A feature extraction module for encoding the cine MRI time series training data using an encoder;

[0147] The prediction unit is used to predict the cine MRI time series training coded data using a decoder to obtain a time series prediction result.

[0148] Through modular design, this system automates the entire process, from cine MRI data acquisition, preprocessing, deep learning prediction, to time series recovery. The cine MRI acquisition module effectively decrypts encrypted files and acquires raw data from different scanning angles. The numbering preprocessing module standardizes data to improve data quality. The time series deep learning prediction module utilizes advanced deep learning architecture to accurately predict time series features. The time series recovery module converts the prediction results into intuitive dynamic images for subsequent analysis and application. The entire system is efficient, accurate, and stable, significantly improving the intelligence and clinical practicality of MRI-guided radiotherapy and providing strong support for tumor radiotherapy.

[0149] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A deep learning-based automatic classification method for cine MRI scan planes, characterized in that: include: Obtain and crack the binary encrypted cine MRI timing files of the magnetic resonance-guided radiotherapy system under different scanning angles to obtain the cine MRI timing cracked files; the scanning angles include sagittal, coronal, and axial. Numbering and preprocessing the cine MRI time sequence cracking file to obtain the preprocessed cine MRI time sequence cracking file; Build a time series deep learning prediction model; input the preprocessed cine MRI time series cracking file into the time series deep learning prediction model, and output the time series prediction result; Based on the timing prediction results, the preprocessed cine MRI timing cracked file is time-recovered to complete the automatic classification of the cineMRI scanning plane.

2. The method for automatic classification of cine MRI scan planes based on deep learning according to claim 1, characterized in that: The cine MRI time sequence cracking file is numbered and preprocessed to obtain the preprocessed cine MRI time sequence cracking file, including: Number the cine MRI time series cracking files and get the corresponding numbering results; Perform linear interpolation and normalization on the cine MRI timing crack file to obtain the normalized cine MRI timing crack file; The normalized cine MRI time sequence cracking file and its corresponding encoding results are superimposed and collected to obtain the preprocessed cine MRI time sequence cracking file.

3. The method for automatic classification of cine MRI scan planes based on deep learning according to claim 1, characterized in that: The temporal deep learning prediction model includes an input layer, an encoder, and a decoder connected in series; the encoder module includes an STB-Transformer1 submodule, an STB-Transformer2 submodule, and an STB-Transformer3 submodule connected in series; the STB-Transformer1 submodule includes a patch partition layer, a linear embedding layer, and an STB-Transformer unit connected in series; the STB-Transformer2 submodule includes a patch merge1 layer and an STB-Transformer unit connected in series; the STB-Transformer3 submodule includes a patch merge2 layer and 3 STB-Transformer units; each STB-Transformer unit includes a window-self-attention block and a moving window-self-attention block in series; the window-self-attention block includes a normalization layer LN1, a W-MSA layer, a normalization layer LN2 and a MLP1 layer; the moving window-self-attention block includes a normalization layer LN3, a SW-MSA layer, a normalization layer LN4 and a MLP2 layer; the encoder includes a first convolution module, a second convolution module and a fully connected module in series; the first convolution module includes a first maximum pooling layer and a DWConv1 layer in series; the second convolution module includes a second maximum pooling layer and a DWConv2 layer in series; the fully connected module includes a flatten layer, a drop out layer and a fully connected layer in series.

4. The method for automatic classification of cine MRI scan planes based on deep learning according to claim 3, characterized in that: The training process of the time series deep learning prediction model includes: Get the cine MRI time series cracking training file; Input the cine MRI time series cracking training file into the input layer, and output the corresponding cine MRI time series training data; Input the cine MRI time series training data into the encoder, and output the corresponding cine MRI time series training encoded data; Input the cine MRI time series training coded data into the decoder, and output the corresponding time series prediction training results; Based on the time series prediction training results, the weight parameters of the time series deep learning prediction model are adjusted to complete the training of the time series deep learning prediction model.

5. The method for automatic classification of cine MRI scan planes based on deep learning according to claim 4, characterized in that: The step of inputting the cine MRI time series training data into the encoder and outputting the corresponding cine MRI time series training coded data comprises: The MRI temporal training encoding data is input into the STB-Transformer1 submodule, and the global-local MRI temporal training initial features are output; The initial features of global-local MRI temporal training are input into the STB-Transformer2 submodule, and the deep features of global-local MRI temporal training are output; The global-local MRI temporal training deep features are input into the STB-Transformer3 submodule, and the cineMRI temporal training encoding data is output.

6. The method for automatic classification of cine MRI scan planes based on deep learning according to claim 5, characterized in that: The training process of the STB-Transformer1 submodule includes: Input the MRI temporal training encoding data into the patch partition layer, and output the MRI temporal training segmentation block data; Input the MRI time series training segmentation block data into the linear embedding layer, and output the MRI time series training linear data; The MRI time series training linear data is input into the STB-Transformer unit, and the global-local MRI time series training initial features are output.

7. The method for automatic classification of cine MRI scan planes based on deep learning according to claim 6, characterized in that: The MRI time series training linear data is input into the STB-Transformer unit, and the global-local MRI time series training initial features are output, including: Input the MRI time series training linear data into the normalization layer LN1, and output the normalized MRI time series training data; The normalized MRI time series training data is input into the W-MSA layer, and the local-MRI time series key features are output; Add the local-MRI time series key features and the MRI time series training linear data to obtain the local-MRI time series superposition features; The local-MRI time series superposition features are input into the normalization layer LN2, and the normalized MRI time series key features are output; The normalized MRI time series key features are input into the MLP1 layer, and the MRI time series nonlinear features are output; The MRI time series nonlinear features and the local-MRI time series superposition features are added to obtain the key features of local MRI time series training; The local MRI temporal training key features are input into the normalization layer LN3, and the normalized local MRI temporal training key features are output; The normalized local MRI time series training key features are input into the SW-MSA layer, and the global MRI time series key features are output; The global-MRI time series key features and the local MRI time series training key features are added to obtain the global-MRI time series superposition features; The global-MRI temporal superposition feature is input to the normalization layer LN4, and the normalized global-MRI temporal superposition feature is output; The normalized global-MRI temporal superposition features are input into the MLP2 layer, and the MRI temporal nonlinear key features are output; The MRI time series nonlinear key features and the global-MRI time series superposition features are added to obtain the global-local MRI time series training initial features.

8. The method for automatic classification of cine MRI scan planes based on deep learning according to claim 1, characterized in that: The method of performing time sequence recovery on the pre-processed cine MRI time sequence cracking file based on the time sequence prediction result includes: Based on the time series prediction results, the pre-processed cine MRI time series cracking file is divided into data sets of three scanning angles: sagittal, coronal and axial; Based on the numbering results, the three data sets are sorted to obtain the sorting results; Based on the sorting results, the cine MRI images of each data set are collected and the corresponding dynamic image gif files are generated.

9. A deep learning-based automatic classification system for cine MRI scan planes, for implementing the deep learning-based automatic classification system for cine MRI scan planes according to any one of claims 1 to 8, characterized in that: include: The cine MRI acquisition module is used to obtain the binary encrypted cine MRI timing files of the magnetic resonance guided radiotherapy system under different scanning angles and crack them to obtain the cine MRI timing cracked files; A numbering preprocessing module is used to number and preprocess the cine MRI time sequence cracking file to obtain the preprocessed cine MRI time sequence cracking file; The time series deep learning prediction module is used to input the pre-processed cine MRI time series cracking file into the time series deep learning prediction model and output the time series prediction result; The timing recovery module is used to perform timing recovery on the pre-processed cine MRI timing cracking file based on the timing prediction results.

10. The deep learning-based cine MRI scan plane automatic classification system according to claim 9, characterized in that: The time series deep learning prediction module includes: A processing unit, for processing the pre-processed cine MRI time series cracking file using the input layer; A feature extraction module for encoding the cine MRI time series training data using an encoder; The prediction unit is used to predict the cine MRI time series training coded data using a decoder to obtain a time series prediction result.

Images