A method of deep learning combined with specific mr data processing for qsm analysis

Abstract

The application discloses a method for QSM analysis by combining deep learning with specific MR data processing, and comprises the following steps: collecting T2* and Phase magnetic resonance data by using an MGE sequence, and manually marking the biological structure outer contour by a researcher; combining the MGE characteristics to realize data enhancement and the marking of the researcher to form a training data set of deep learning; constructing a deep learning segmentation network based on a pre-training backbone improvement; training a model to segment a boundary fuzzy biological tissue T2* image and generate a phase mask; performing batch processing on the phase mask data and registering the phase mask data with Phase magnetic resonance data, and finally using the phase mask data for QSM analysis. The improved deep learning method is used to generate the phase mask required in QSM analysis, and specific MR data processing is combined to realize batch image registration, so that the QSM analysis method is applied to animal MRI image research and assists scientific research analysis, and the analysis accuracy is improved.

Description

Technical Field

[0001] This invention relates to magnetic resonance imaging data processing, and more particularly to a method for QSM analysis using deep learning combined with specific MR data processing. Background Technology

[0002] Magnetic resonance imaging (MRI) technology has advanced rapidly over the past decade. Susceptibility-weighted imaging and quantitative susceptibility imaging (QSM), developed from phase data in MRI images, have made it possible to visualize and quantify iron content in biological organisms through rapid iteration (A. Deistung, F. Schweser, and J.R. R. Eichenbach, “Overview of quantitative susceptibility mapping,” NMR Biomed, vol. 30, no. 4, pp. e3569, Apr, 2017.). Specifically, the combination of multigradient echo (MGE) sequences with post-processing techniques and modeling calculations significantly improves the contrast between iron-containing regions and other regions in the image, enabling quantitative calculation of iron content. The distribution of magnetic susceptibility in local tissues of a biological organism is reflected in QSM, making previously difficult-to-distinguish regions show clear boundaries. In recent years, QSM has not only been applied to the analysis of MRI data in biological tissues, but has also been gradually extended to the tracing of exogenous magnetic materials in vivo (Y. Wang, and T. Liu, “Quantitative susceptibility mapping (QSM): Decoding MRI data for a tissue magnetic biomarker,” MagnReson Med, vol. 73, no. 1, pp. 82-10l, Jan, 2015.). Quantitative relative magnetic susceptibility analysis can reflect the residence time of exogenous objects and their distribution in vivo, providing new perspectives and schemes for the in vivo tracing of novel materials, and therefore has attracted increasing attention. However, existing QSM analysis software (such as STI suite, Medi, etc.) is mainly designed for applications involving biological brain tissue MRI images (W. Li, B. Wu, and C. Liu, "Quantitative susceptibility mapping of human brain reflects spatial variation in tissue composition," Neuroimage, vol. 55, no. 4, pp. 1645-56, Apr 15, 2011.). When QSM is applied to other tissues (such as kidneys, neck, etc.) and exogenous substances are added, the final QSM analysis results may show signal loss. This is because in the initial stage of QSM processing, a phase mask needs to be generated to mask the biological tissue and exogenous materials as a whole. However, the mask generation methods built into the aforementioned QSM software are relatively weak and cannot generate the correct mask.Based on biological tissue structure, the efficient generation of accurate phase mask images enables rapid QSM processing of MRI data, which has significant application and research value.

[0003] Currently, a series of studies on phase mask generation have been published, including threshold segmentation methods (K. J. Shanthi, and MS. Kumar, “Skull stripping and automatic segmentation of brain MRI using seed growth and threshold techniques,” pp. 422-426%@1424413559, 2007.), patch-based segmentation algorithms (E. Roura, A. Oliver, M. Cabezas et al., “MARGA: multispectral adaptive region growing algorithm for brain extraction on axial MRI,” Comput Methods Programs Biomed, vol. 113, no. 2, pp. 655-73, Feb, 2014.), and watershed algorithms (DERex, D. W. Hattuck, RP. Woods et al., “A meta-algorithm for brain extraction in…”). MRI, “Neuroimage, vol.23, no.2, pp.625-37, Oct, 2004.”, Bayesian methods, etc. (G. Zhao, F. Liu, JAOler et al., “Bayesian convolutional neural network-based MRI brain extraction on nonhuman primates,” Neuroimage, vol.175, pp.32-44, Jul 15, 2018.). Phase mask image generation methods belong to image segmentation algorithms. The main problem with these traditional segmentation methods is insufficient segmentation accuracy and low performance in biological structures outside of brain tissue, resulting in signal loss in QSM processing results, which is difficult to meet the needs of practical scientific research applications.Deep learning methods have solved many problems in various fields in recent years. By automatically learning useful feature representations from big data, convolutional neural networks have shown significant advantages in segmenting biological tissues (S. Minaee, YY Boykov, F. Porikli et al., “Image Segmentation Using Deep Learning: A Survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1-1, 2021.). In addition, analyzing and manipulating MRI data formats, batch assembling automatically generated mask images and registering them with phase images before re-entering them into QSM analysis software is a necessary process before QSM analysis and after mask generation (M. Larobina, and L. Murino, “Medical Image File Formats,” Journal of Digital Imaging, vol. 27, no. 2, pp. 200-206, 2014.). Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method for QSM analysis that combines deep learning with specific MR data processing to improve the accuracy of analysis.

[0005] Technical solution: The method for improving analysis accuracy by combining deep learning with specific MR data processing for QSM analysis includes the following steps:

[0006] Step 1: Raw T2* and Phase magnetic resonance imaging data were acquired using MGE sequences. Researchers manually annotated the outer contours of the organisms based on experience and the actual physiological structure of the animals, obtaining the initial dataset and some corresponding annotations.

[0007] Step 2: Utilize the characteristics of MGE multi-TE time, and combine manually labeled data by researchers to perform data augmentation, thereby expanding the dataset and forming a complete training dataset;

[0008] Step 3: Construct a deep learning segmentation network based on pre-trained backbone improvements;

[0009] Step 4: Input the training data obtained in Step 2 into the constructed segmentation network model for learning, and use the Dice loss function to correct the network to obtain a segmentation network with higher accuracy;

[0010] Step 5: Input the data to be segmented into the segmentation network trained in Step 4 to obtain the mask results, then automatically combine them into phase mask data and register them with the Phase magnetic resonance data. Finally, input the registration results into the QSM analysis software for batch analysis.

[0011] Further, step 1, which involves acquiring raw T2* and Phase magnetic resonance imaging data, includes: using a 9.4T small animal magnetic resonance imaging device to perform neck scans on the animal model under test (rat), acquiring neck magnetic resonance data including T2* and Phase modalities, and classifying, correcting image grayscale, and preprocessing the T2* and Phase modal data; the researchers manually annotating the biological outer contour, specifically: the researchers manually draw and fill in the outer contour region of the rat's neck on the T2* magnetic resonance image; the magnetic resonance data and the region marked by the researchers constitute the basic dataset.

[0012] Furthermore, step 2, which utilizes the characteristics of multiple TE times in MGE, specifically refers to the following: a single magnetic resonance image data of a rat under an MGE sequence consists of multiple TE times, with a neck slice location corresponding to multiple T2* images and Phase images at different TE times; the step of combining manually annotated data by researchers to perform data augmentation and expand the dataset specifically includes: expanding the magnetic resonance image annotations of a single TE time to magnetic resonance images at different TE times at the same slice location, forming a multi-TE time T2* image and annotated image matching outer contour annotation dataset.

[0013] Further, step 3 specifically involves: constructing a deep learning segmentation network model based on a pre-trained backbone. The segmentation network model is based on the UNet architecture and is divided into a feature extraction branch and a decoding branch. The feature extraction branch mainly consists of a VGG16 pre-trained network model, whose input is a T2* image. It performs multiple convolution calculations on the input data to extract the effective parts of the data. The decoding branch upsamples and expands the results of the feature extraction branch, and uses a skip connection method to merge the results of the corresponding stage of the feature extraction branch into the image segmentation range in each upsampling process, thereby obtaining the final segmentation probability map.

[0014] Further, step 4, which involves inputting the training data into the constructed segmentation network model for learning and using the Dice loss function to correct the network, specifically involves: firstly, randomly dividing the dataset into a training set and a test set. The training set is used to initialize training, and the test set is used to improve the model's generalization ability. To further prevent overfitting, online data augmentation is performed on the input data, including methods such as axis transformation, mirror transformation, slant transformation, arbitrary angle rotation, gamma transformation, and affine transformation. The step of using the Dice loss function to correct the network specifically involves using the Dice loss function as the training loss function to improve the similarity between the segmentation result and the labeled image, thereby achieving the training objective and obtaining a deep network with high prediction accuracy.

[0015] Further, in step 5, the data to be segmented is input into the segmentation network trained in step 4 to obtain the mask result, which is then automatically combined into phase mask data and registered with the Phase magnetic resonance data. Specifically, the data to be segmented is input into the trained network, the segmentation result is output as the data result, and then registered with the corresponding Phase magnetic resonance data. The DICOM file header of the Phase data is reused to assemble the segmentation result into a DICOM file. A marker is introduced at the initial position of the data matrix to distinguish different mask images at the same slice position. Finally, the DICOM files of the segmentation result are batched into NII data to generate the final phase mask data that can be recognized by QSM software.

[0016] Beneficial Effects: Compared with existing technologies, this invention has the following advantages: This invention aims to solve the problem of signal loss in existing QSM algorithms when processing regions with unclear boundaries, thereby improving the accuracy of analysis. Based on T2* and Phase MRI images acquired using MGE sequences, this invention establishes a method for QSM analysis using deep learning combined with specific MR data processing. It utilizes the multi-TE time characteristics of MGE sequences to expand the segmentation dataset through a single MRI capture; it improves the UNet deep neural network using the VGG16 skeleton; it corrects the segmentation results using the Dice loss function, achieving mask generation for MRI image data of biological tissues with blurred boundaries; simultaneously, it performs batch processing on the generated DICOM data, introducing markers at the initial position of the data matrix to annotate and recombine the mask images, and then matching and assembling them with Phase images to form NII data; ultimately, it achieves the generation of usable data sets for the QSM algorithm, realizing the purpose of QSM processing for MRI image data in regions with unclear boundaries. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention;

[0018] Figure 2 These are the original T2*, Phase, and labeled images;

[0019] Figure 3 This is a schematic diagram of the expanded basic dataset of this invention;

[0020] Figure 4 This is a diagram of the mask generation network structure of the present invention;

[0021] Figure 5 This is a schematic diagram illustrating how the present invention introduces flag bits to distinguish mask DICOMs; Detailed Implementation

[0022] See Figure 1 The present invention includes the following specific steps:

[0023] S101. Using MGE sequences to acquire T2* and Phase magnetic resonance imaging data, researchers manually annotated the outer contours of organisms to form a basic dataset.

[0024] S102. Based on the basic data, expand the dataset according to the characteristics of the MGE sequence;

[0025] S103. Construct a deep learning segmentation network based on pre-trained backbone improvements;

[0026] S104. Input the training data obtained in S102 into the constructed segmentation network model for learning, and use the Dice loss function to correct the network to obtain a segmentation network with high accuracy.

[0027] S105. Input the data to be segmented into the segmentation network trained in S104 to obtain the mask results, then automatically combine them into phase mask data and register them with the Phase magnetic resonance data. Finally, input the registration results into the QSM analysis software for batch analysis.

[0028] The specific measures are as follows:

[0029] Step S101, which involves acquiring raw T2* and Phase magnetic resonance imaging data, includes: using a 9.4T small animal magnetic resonance imaging device to scan the neck of the animal model (rat), acquiring neck magnetic resonance data including T2* and Phase modalities, and classifying, correcting image grayscale, and preprocessing the T2* and Phase modal data; the researchers manually annotating the outer contour of the organism, specifically: the researchers manually draw and fill in the outer contour region of the rat's neck on the T2* magnetic resonance image; the magnetic resonance data and the region marked by the researchers constitute the basic dataset.

[0030] Step S102, utilizing the characteristics of multiple TE times in MGE, specifically refers to the following: a single magnetic resonance image of a rat under an MGE sequence consists of multiple TE times, with a neck slice location corresponding to multiple T2* images and Phase images at different TE times; the step of combining manually labeled data by researchers to perform data augmentation and expand the dataset specifically includes: extending the magnetic resonance image annotations of a single TE time by researchers to magnetic resonance images at different TE times at the same slice location, forming a multi-TE time T2* image and labeled image matching outer contour annotation dataset.

[0031] Step S103 constructs a deep learning segmentation network model based on a pre-trained backbone. The segmentation network model is based on the UNet architecture and is divided into a feature extraction branch and a decoding branch. The feature extraction branch is mainly composed of a VGG16 pre-trained network model. Its input is a T2* image, and it performs multiple convolution calculations on the input data to extract the effective part of the data. The decoding branch upsamples and expands the results of the feature extraction branch, and uses a skip connection method to merge the results of the corresponding stage of the feature extraction branch into the image segmentation range in each upsampling process, thereby obtaining the final segmentation probability map.

[0032] Step S104 inputs the training data from S102 into the constructed segmentation network model for learning. Specifically, the dataset is first randomly divided into a training set and a test set. The training set is used for initial training, and the test set is used to improve the generalization ability of the model. To further prevent overfitting, the input data is augmented online using random stretching, translation, flipping, and rotation transformations. The Dice loss function is used to correct the network. Specifically, the Dice loss function is used as the training loss function to improve the similarity between the segmentation results and the labeled images, thereby achieving the training objective and obtaining a segmentation network with higher accuracy.

[0033] Step S105 inputs the data to be segmented into the segmentation network trained in S104. The segmentation results are automatically combined into phase mask data and registered with the Phase magnetic resonance data. Finally, the registration results are input into the QSM analysis software for batch analysis. Specifically, the data to be segmented is input into the trained network, the segmentation results are output as data results, and then registered with the corresponding Phase magnetic resonance data. The DICOM file header of the Phase data is reused to assemble the segmentation results into a DICOM file. A marker is introduced at the initial position of the data matrix to distinguish different mask images at the same slice position. Finally, the DICOM files of the segmentation results are batched into NII data to generate the final phase mask data that can be recognized by the QSM software.

[0034] Example 1

[0035] See Figure 1 This embodiment specifically includes:

[0036] Step S101: Acquire T2* and Phase magnetic resonance imaging data. Researchers manually annotated the biological outer contour, specifically by installing a surface coil on a 9.4T small animal MRI scanner, using an MGE sequence to scan the neck of the animal model (rat), and acquiring neck MRI data including T2* and Phase modalities. The T2* and Phase modal data were then classified, image grayscale corrected, and format preprocessed. Based on this, researchers manually drew and delineated the outer contour region of the rat's neck on the T2* MRI image, obtaining a complete mask marker with clear boundaries and no holes. The original T2* and Phase images and the manually annotated images are shown below. Figure 2 As shown, the raw magnetic resonance data and the researchers' labels constitute the basic dataset.

[0037] Step S102: Utilizing the characteristics of MGE scan sequences, where a single slice location corresponds to multiple MR images with different TE times (e.g., six MR images corresponding to one slice location with TE times of 4ms, 8ms, 12ms, 16ms, 20ms, and 24ms), and combining the results of researchers manually labeling slices at single TE times (e.g., labeling on the 4ms image), the remaining five MR images at TE times do not need to be manually drawn; the annotations at the same slice location can be directly reused. A diagram illustrating the expansion of the basic dataset is shown below. Figure 3 As shown, by labeling only a portion of the data, the labels can be expanded to all data, achieving rapid and efficient expansion of the basic dataset.

[0038] In step S102, before the training data is input into the deep learning network model, the data format needs to be converted and the resolution unified: First, the original DICOM data is converted into image data (e.g., JPG format) in batches to facilitate the deep neural network's reading of the data; then, all image data is cropped to 230×230 pixels. 2 The data blocks are sized to ensure the entire outer contour is included in the data, and some invalid background areas are removed. In addition to the data augmentation method in S102, further data augmentation is performed to prevent overfitting. Methods include axis transformation, mirror transformation, shear transformation, arbitrary angle rotation, gamma transformation, and affine transformation. These transformations need to be applied simultaneously to the labeled data to ensure consistency between the training and labeled data.

[0039] Step S103: Construct a deep learning segmentation network based on pre-trained backbone improvement. See the structure diagram below. Figure 4The network model includes a feature extraction branch and a decoding branch. The original T2* image is input into the feature extraction branch of the network model, and the corresponding feature extraction result is then input into the decoding branch.

[0040] In step S103, the feature extraction branch includes convolutional layers (3×3 kernel size, stride 1, and channel numbers of 64, 128, 256, and 512 respectively), activation layers (ReLU function), and max pooling layers (2×2 pooling region size, stride 2). The feature extraction branch consists of five layers in total, with each layer performing two convolutional operations, one activation operation, and one pooling operation. Deeper networks have a larger field of view and pay more attention to texture features.

[0041] f ReLU (z) = max{0, z}

[0042] Where z represents the feature value on the feature layer.

[0043] In step S103, the decoding branch includes a convolutional layer (3×3 kernel size, stride 1), an activation layer (ReLU function), and an upsampling layer (3×3 deconvolution size, stride 2), as well as a step to connect the feature extraction branch to the decoding branch. The decoding branch combines data from the feature extraction branch, enabling better data integration and recognition, and preventing information loss during data upsampling. This significantly improves feature extraction capabilities. The formula for the ReLU function is:

[0044] In step S103, the final part of the decoding stage, the feature output is input into the argmax layer. In the argmax layer, the sigmoid function is used to evaluate the results of all channels, ultimately reducing the number of channels to 1 to obtain the weight value for each pixel in the space. The formula for the sigmoid function is:

[0045]

[0046] Where x represents the eigenvalues ​​on the feature map.

[0047] In step 103, in the final result layer where the number of channels is reduced to 1, each pixel is classified as either background or biological tissue, represented by 0 and 1 respectively. The final mask image is obtained by increasing the number of channels back to 3 according to the corresponding flags and filling the background and biological tissue with black (0,0,0) and white (256,256,256) respectively.

[0048] In step S104, the training data obtained in S102 is input into the constructed segmentation network model for learning, and the Dice loss function is used to correct the network, thereby training the deep learning network and obtaining a segmentation network with high accuracy. The optimizer used during training is the Adam algorithm, with its initial rate set to 10. -4 The training process is divided into two stages: first, the backbone feature extraction network is trained, and then the entire network is trained. Each stage involves 50 and 150 training iterations, respectively. The input batch size is 8, and the loss function is Diceloss, with the following formula:

[0049]

[0050] Where pi is the predicted value of the pixel at that point, and gi is the true labeled value of the pixel at that point. This loss function makes the predicted and true values ​​closer, accelerating the fitting of the training.

[0051] In step S105, the data to be segmented is input into the network trained in step S104, and the segmentation result is output as the data result. The network automatically processes the data to generate a mask image, which is then registered with the corresponding Phase MRI data. The DICOM file header of the Phase data is reused to assemble the mask image into a DICOM file. A flag is introduced at the initial position of the data matrix to distinguish different mask images at the same slice position (e.g., the flag bit for an image with a TE time of 4ms is assigned a value of 1, the flag bit for an image with a TE time of 8ms is assigned a value of 2, and so on). A schematic diagram illustrating the introduction of flag bits to distinguish between mask DICOM files is shown below. Figure 5 (As shown), this prevents the situation where, during the next processing step, when combining DICOM type data into NII data, the identical mask data is discarded, disrupting the correspondence and losing the registration relationship. Finally, the T2*, Phase, and mask DICOM files are batched into NII data respectively, generating the final phase mask data group that can be recognized by QSM software, which is then input into QSM software for data analysis.

Claims

1. A method for QSM analysis using deep learning combined with specific MR data processing, characterized in that, Includes the following steps: Step 1: Use MGE sequences to acquire raw T2* and Phase magnetic resonance imaging data. Based on experience and the actual physiological structure of the animal, manually annotate the outer contour of the organism to obtain the initial dataset and some corresponding annotations. Step 2: Utilize the characteristics of MGE multi-TE time, and combine manually labeled data by researchers to perform data augmentation, thereby expanding the dataset and forming a complete training dataset; Step 3: Construct a deep learning segmentation network based on pre-trained backbone improvements; Step 4: Input the training data obtained in Step 2 into the constructed segmentation network model for learning, and use the Dice loss function to correct the network to obtain the trained segmentation network; Step 5: Input the data to be segmented into the segmentation network trained in Step 4 to obtain the mask results, then automatically combine them into phase mask data and register them with the Phase magnetic resonance data. Finally, input the registration results into the QSM analysis software for batch analysis. The process involves inputting the data to be segmented into the segmentation network trained in step 4 to obtain mask results, which are then automatically combined into phase mask data and registered with Phase magnetic resonance data. The method is as follows: the data to be segmented, i.e., the T2* image, is input into the trained network, the segmentation result is output as the data result, and then registered with the corresponding Phase magnetic resonance data. The DICOM file header of the Phase data is reused to assemble the segmentation results into a DICOM file. A marker is introduced at the initial position of the data matrix to distinguish different mask images at the same slice position. Finally, the DICOM files of the segmentation results are batched into NII data, ultimately generating phase mask data that can be recognized by QSM software.

2. The method for QSM analysis using deep learning combined with specific MR data processing according to claim 1, characterized in that, Step 1 involves acquiring raw T2* and Phase magnetic resonance imaging data: A 9.4T small animal magnetic resonance imaging device is used to scan the neck of the animal model under test, and neck magnetic resonance data including T2* and Phase modes are acquired. The T2* and Phase mode data are then classified, image grayscale corrected, and format preprocessed.

3. The method for QSM analysis using deep learning combined with specific MR data processing according to claim 1, characterized in that, Manual annotation of biological outer contours involves researchers manually drawing and filling in the outer contour region of the rat's neck on T2* magnetic resonance images. The magnetic resonance data and the regions marked by the researchers constitute the basic dataset.

4. The method for QSM analysis using deep learning combined with specific MR data processing according to claim 1, characterized in that, The feature of using multiple TE times in step 2 is that a single magnetic resonance image of a rat under an MGE sequence consists of multiple TE times, and a neck slice location corresponds to multiple T2* images and phase images at different TE times.

5. The method for QSM analysis using deep learning combined with specific MR data processing according to claim 1, characterized in that, Data augmentation by combining manually labeled data to expand the dataset involves extending the annotations of a single TE time magnetic resonance image to magnetic resonance images at different TE times at the same slice location, forming an outer contour annotation dataset that matches T2* images and labeled images one by one across multiple TE times.

6. The method for QSM analysis using deep learning combined with specific MR data processing according to claim 1, characterized in that, Step 3 constructs a deep learning segmentation network model based on a pre-trained backbone. The segmentation network model is based on the UNet architecture and is divided into a feature extraction branch and a decoding branch. The feature extraction branch is composed of a VGG16 pre-trained network model, which takes a T2* image as input and performs multiple convolution calculations on the input data to extract the effective part of the data. The decoding branch upsamples and expands the results of the feature extraction branch, and uses a skip connection method to merge the results of the corresponding stage of the feature extraction branch into the image segmentation range in each upsampling process, thereby obtaining the final segmentation probability map.

7. The method for QSM analysis using deep learning combined with specific MR data processing according to claim 1, characterized in that, Step 4, which involves inputting the training data into the constructed segmentation network model for learning and using the Dice loss function to correct the network, involves: firstly, randomly dividing the dataset into a training set and a test set. The training set is used for initial training, and the test set is used to improve the model's generalization ability. To further prevent overfitting, online data augmentation is performed on the input data using random stretching, translation, flipping, and rotation transformations. The step of using the Dice loss function to correct the network specifically involves using the Dice loss function as the training loss function to improve the similarity between the segmentation results and the labeled images, thereby achieving the training objective and obtaining a deep network with high prediction accuracy.

Images