A method for detecting activation of mouse brain nuclei based on manganese-enhanced magnetic resonance imaging

Abstract

The application discloses a method for detecting mouse brain nucleus activation based on manganese-enhanced magnetic resonance imaging, and belongs to the field of image processing, which comprises the following steps: converting the manganese-enhanced magnetic resonance image of the mouse head collected into NIFTI format, keeping the direction and voxel size of the image consistent with the standard mouse brain atlas template; performing bias field correction on the image; loading the PLKA- nnUNet model, performing image segmentation on the image, outputting a binary brain mask, and extracting individual mouse brain images from the image after bias field correction; performing image registration and intensity normalization; calculating the mean relaxation rate of the four hippocampal subregions of the mouse brain R1 map for each individual mouse brain image after registration, which is used for quantitatively comparing the activation of the mouse brain; and performing statistical testing at the voxel level on the registered and normalized individual mouse brain images to identify the brain regions with intensity differences between different experimental conditions. The method reduces the dependence on professional operation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing and neural network learning, and particularly relates to a method for detecting activation of a mouse brain nucleus based on manganese-enhanced magnetic resonance imaging. BACKGROUND

[0002] Manganese-enhanced magnetic resonance imaging (MEMRI) uses divalent manganese ions (Mn²⁺) as a substitute marker for calcium ions, which can enter excitatory neurons through voltage-gated / ligand-gated channels, be taken up with the level of neural activity, and produce enhancement on T1-weighted MRI, thereby more directly reflecting the activation state of the neural nucleus. Compared with fMRI based on hemodynamics, MEMRI has higher spatial resolution and more direct physiological relevance.

[0003] However, the current MEMRI research still lacks a unified, automated and robust process in the image processing link, especially in the two key steps of skull removal and image registration. The traditional method is sensitive to parameters, has insufficient generalization, and is easily affected by the differences in mouse brain volume / morphology and acquisition artifacts, thereby reducing the sensitivity and consistency of subsequent inter-group statistical analysis.

[0004] In addition, the intensity non-uniformity (such as Bias Field caused by coils) and intensity standardization across individuals / multiple time points of MEMRI data also affect the quantitative analysis of regions of interest (ROIs) and voxel-level statistics (SPM). Therefore, it is of great significance to construct a standardized and automated processing flow for MEMRI mouse brain images and make methodological innovations in key steps. SUMMARY

[0005] The purpose of the present application is to provide a method for detecting activation of a mouse brain nucleus based on manganese-enhanced magnetic resonance imaging, so as to solve the problems of non-uniform standardization of the existing MEMRI image processing flow, low degree of automation in key steps, and insufficient robustness, and ultimately realize stable and objective detection of the activation of the mouse brain nucleus.

[0006] The technical solution for achieving the purpose of the present application is as follows:

[0007] A method for detecting activation of a mouse brain nucleus based on manganese-enhanced magnetic resonance imaging, comprising:

[0008] Converting the collected manganese-enhanced magnetic resonance image of the mouse head from original DICOM data to NIFTI format, and keeping the direction and voxel size of the NIFTI format image consistent with the standard mouse brain atlas template;

[0009] Performing bias field correction based on the N4 algorithm on the image to remove intensity drift caused by magnetic field inhomogeneity;

[0010] Load the PLKA-nnUNet model, perform image segmentation on the image, output a binary brain mask, and extract individual mouse brain images from the image after bias field correction by the N4 algorithm; the PLKA-nnUNet is introduced into the parallel large kernel convolution module, the parallel fusion attention module and the lightweight convolution design on the basis of nnUNet;

[0011] Align the extracted individual mouse brain images to the standard mouse brain atlas template to realize image registration;

[0012] Intensity normalization is performed on each registered individual mouse brain image;

[0013] The T1-mapping quantitative map of each registered individual mouse brain image is obtained by taking the reciprocal to obtain the corresponding relaxation rate map, the hippocampus atlas is superimposed, and the relaxation rate mean values of the DG, CA3, CA1 and CA2 four hippocampus subregions of the mouse brain are calculated, which are used for quantitatively comparing the activation of the mouse brain; statistical test is performed on the registered and normalized individual mouse brain images at the voxel level to identify the brain regions with intensity differences between different experimental conditions.

[0014] Compared with the prior art, the present application has the following advantages:

[0015] (1) The processing flow is highly standardized and automated: the present application provides an end-to-end complete scheme from data preprocessing to statistical analysis, greatly reduces the tedious manual operation and subjective intervention, significantly improves the processing efficiency and the repeatability of the results, and provides a reliable technical foundation for large-scale, multi-time point brain function research.

[0016] (2) Breakthrough in skull removal key step in precision and robustness: the proposed PLKA-nnUNet model, through the cooperative design of parallel large kernel convolution and attention mechanism, is superior to traditional methods and other mainstream deep learning models in multiple objective evaluation indicators. It can effectively deal with complex situations such as fuzzy brain tissue edge, existence of small damage or uneven signal, realize more complete and accurate brain tissue extraction, and the model is designed to be lightweight, easy to deploy and apply.

[0017] (3) Improve the accuracy and applicability of image registration: the innovative SyNRACC registration strategy replaces the traditional mutual information metric with a normalized cross-correlation metric that is more sensitive to MEMRI signal enhancement characteristics, which performs better in multiple registration accuracy indicators. This ensures that brain images of different individuals and different time points can be more accurately aligned in the standard space, laying a solid spatial foundation for subsequent quantitative comparison.

[0018] (4) Effective support for the extraction and analysis of deep neural activity information: This method eliminates technical variations through standardized processes, allowing small MEMRI signal differences derived from neuronal activity to be stable and objective. Experiments have shown that the application of this process can reliably detect significant signal changes in mouse hippocampal DG, CA3 and other nuclei under specific intervention, providing a powerful analysis tool for brain function and neural circuit research. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 For the contrast of mouse brain magnetic resonance images before and after modifying header file information.

[0020] Figure 2 For the contrast of mouse brain magnetic resonance images after N4 bias field correction.

[0021] Figure 3 For the structure diagram of parallel multi-scale large kernel convolution attention module.

[0022] Figure 4 For the structure diagram of grouped large kernel convolution attention gate module.

[0023] Figure 5 For the overall architecture diagram of the PLKA-nnUNet model.

[0024] Figure 6 For the skull removal effect comparison chart of PLKA-nnUNet and other methods under different data sets.

[0025] Figure 7 For the quantitative index comparison chart of skull removal effect of PLKA-nnUNet and other methods.

[0026] Figure 8 For the ablation experiment results of the PLKA-nnUNet model.

[0027] Figure 9 For the overall effect chart of the SyNRACC image registration method.

[0028] Figure 10 For the image registration effect comparison chart of SyNRACC registration and other registration methods under the same data.

[0029] Figure 11 For the processing flowchart of standard space template and brain atlas.

[0030] Figure 12 For the processing flowchart of the manganese-enhanced magnetic resonance images collected.

[0031] Figure 13 For the mouse brain manganese ion metabolism experiment flowchart.

[0032] Figure 14 R1 visualization contrast map of manganese ion in mouse brain over time after tail vein injection.

[0033] Figure 15 R1 value change curve and significant difference map of each hippocampal nuclear group of mice compared to before manganese ion injection.

[0034] Figure 16 Schematic diagram of manganese-enhanced magnetic resonance imaging experiment process based on magnetic stimulation of iron-based magnetic nanoparticles.

[0035] Figure 17 Schematic diagram of injected magnetic nanoparticles and mouse brain injection site.

[0036] Figure 18 Visualization contrast map of each hippocampal nuclear group of mice before and after magnetic stimulation.

[0037] Figure 19 Significant difference contrast map between magnetic stimulation group and non-magnetic stimulation group in different hippocampal nuclear groups of interest.

[0038] Figure 20 Axial position hippocampal different layer voxel level comparison significant difference map. DETAILED DESCRIPTION

[0039] The application will be further described below in conjunction with the drawings and specific examples.

[0040] Example 1: Standardized process of MEMRI data processing and key algorithm implementation

[0041] One method for detecting mouse brain nuclear activation based on manganese-enhanced magnetic resonance imaging in this example includes the following processes:

[0042] 1.1 Data format conversion and header file information standardization: convert the collected manganese-enhanced magnetic resonance images of mouse head from original DICOM data to NIFTI format, adjust the qform and sform matrices in the NIFTI header file through a script, so that the direction and voxel size of the NIFTI image are consistent with the standard mouse brain atlas template published by the Australian Mouse Brain Mapping Consortium (AMBMC, Australian Mouse Brain Mapping Consortium), and modify the header file information before and after the image contrast as shown in Figure 1 By adjusting the above script, the problem of rotation or flipping deviation between the original image and the standard template is solved, and the alignment of the image direction and the template is achieved.

[0043] 1.2 N4 bias field correction: Perform bias field correction on the image based on the N4 algorithm to remove slow intensity drift caused by magnetic field inhomogeneity. The specific implementation uses the N4BiasFieldCorrection function in the ANTs (Advanced Normalization Tools) toolbox, with the number of iterations set to 4, the convergence threshold set to 0.001, and the spline distance parameter set to 200. As shown in FIG. 1, after N4 bias field correction, the intensity distribution uniformity of the image is significantly improved. Figure 2

[0044] 1.3 Skull removal: Load the PLKA-nnUNet model to perform image segmentation on the three-dimensional image, output a binary brain mask, and extract brain tissue from the image after N4 bias field correction. PLKA-nnUNet is a lightweight segmentation network improved based on nnUNet, which maintains the advantages of nnUNet's adaptive configuration training process while enhancing the ability to extract global context information by introducing parallel multi-scale large kernel convolution attention modules, grouped large kernel convolution attention gate modules, and depth separable convolution design. The following is a description of the core structure of the model:

[0045] 1.3.1 Network training and inference process

[0046] Load the pre-trained PLKA-nnUNet model weights and configuration file. In the inference stage, input the three-dimensional image volume that has been preprocessed by nnUNet (such as resampling and normalization). After forward propagation of the network, output the probability map of each voxel belonging to brain tissue, and obtain the binary segmentation mask by thresholding (0.5). Multiply this mask with the image that has been corrected by the N4 algorithm bias field, and the brain tissue extraction is completed.

[0047] 1.3.2 Encoder and decoder basic structure

[0048] The nnUNet model adopts the classic U-Net encoder-decoder architecture and has the following key improvements:

[0049] Parallel multi-scale large kernel convolution attention module (PMSLKA): This module is embedded after each standard convolution layer in the decoder and consists of a parallel large kernel convolution module (PMSLK) and a parallel fusion attention module (PMA) connected in turn, aiming to enhance the model's understanding and fusion of global brain scale context information.

[0050] Grouped large kernel convolution attention gate module (GLAG): This module is set at the jump connection and is used to dynamically calibrate the feature stream when the decoder performs feature fusion, i.e., fusing the encoder features from the jump connection and the upsampled decoder features, focusing on significant features with less computational cost.​

[0051] Depthwise separable convolution (DSC): In the place of traditional 3D convolution, depthwise separable convolution (DSC) is used in each standard convolutional layer of the encoder (down-sampling path) and decoder (up-sampling path). DSC first performs a depthwise convolution (Depthwise Conv3D) on each input channel independently, and then combines the channel information through a 1x1x1 pointwise convolution (Pointwise Conv3D). This operation significantly reduces the computational complexity and parameter amount, achieving lightweight design of the model while maintaining the effectiveness of feature extraction.

[0052] 1.3.3 Parallel multi-scale kernel convolution attention module

[0053] It is composed of two core parts connected in sequence: the parallel kernel convolution module is responsible for extracting multi-scale features and generating attention weights, and the parallel fusion attention module further refines and calibrates the features from the three dimensions of channel, pixel, and space. Figure 3 Structure diagram of parallel multi-scale kernel convolution attention module.

[0054] 1.3.3.1 Parallel kernel convolution module

[0055] This module adopts a multi-scale feature extraction strategy combining serialization and parallelization. First, the input features are batch normalized (BatchNorm) to speed up convergence; the normalized features are sequentially transformed by two basic convolution layers: first, a 1x1x1 3D pointwise convolution (PWConv3D) is used for lightweight channel feature mapping and interaction; then a 5x5x5 3D convolution (Conv3D) is used to extract basic local neighborhood features.

[0056] After that, the feature stream is divided into three parallel branches, each using a depthwise dilated convolution (DWDConv3D) with different kernel sizes but the same dilation rate (dilation rate=3) to capture context information in different receptive field ranges. Specifically: branch one uses a depthwise dilated convolution with a kernel size of 19x19x19 to capture long-range dependencies in the image; branch two uses a depthwise dilated convolution with a kernel size of 13x13x13 to balance context information and structural details at a medium scale; branch three uses a depthwise dilated convolution with a kernel size of 7x7x7 to focus on extracting more fine-grained local detail features.

[0057] The output feature maps of the three branches are then concatenated along the channel dimension, and the fused features are sent to a multi-layer perceptron (MLP) for feature transformation and integration. The MLP is composed of two 3D point convolution layers in series, with LeakyReLU activation functions used between layers. The final output of the MLP is passed through a Sigmoid activation function and then connected in residual to the original input features (before batch normalization) of this module to complete the first stage of attention guidance.

[0058] 1.3.3.2 Parallel fusion attention module

[0059] This module receives the features enhanced by the aforementioned parallel kernel convolution module. First, the input features are batch normalized (BatchNorm), and then the features are input in parallel to three independent attention branches: channel attention module (CA), pixel attention module (PA), and spatial attention module (SA). All three attention modules use standard implementations.

[0060] The channel attention module first performs global average pooling on each channel in the spatial dimension, then inputs these channel-level features into a subnetwork composed of two point convolution layers, and finally generates the final channel attention weight vector through a Sigmoid activation function. The weight vector is finally multiplied with the original input features of the channel attention module on a channel-by-channel basis to enhance important feature channels.

[0061] The pixel attention module learns the importance weight of each spatial position (voxel) through a subnetwork composed of two point convolution layers, with ReLU activation functions used between layers. The output is passed through a Sigmoid activation function to generate a pixel-level attention weight map, which is then multiplied with the original input features of the pixel attention module on an element-by-element basis to strengthen the features of key spatial positions.

[0062] The spatial attention module is used to capture important regions in the spatial dimension. First, the feature maps are separately subjected to global average pooling and global maximum pooling in the channel dimension to obtain two single-channel spatial feature maps. Then, these two spatial feature maps are concatenated in the channel dimension and input into a 7x7 convolution layer to learn the importance patterns of different spatial positions. Finally, the output is multiplied with the original feature map input to the spatial attention module on a point-by-point basis through a Sigmoid activation function, so that the network pays more attention to key regions.

[0063] The output feature maps of the three branches are spliced together along the channel dimension, and the spliced features are fused and reduced in dimension by two consecutive 1x1x1 3D point convolution layers. LeakyReLU activation functions are used between the two 3D point convolution layers, and finally a Sigmoid activation function is used. The input features of the parallel fusion attention module are added element by element to form a residual connection, and finally a feature calibrated by multiple attention is output, providing more accurate feature representation for subsequent processing.

[0064] 1.3.4 Group Large Kernel Attention Gate (GLAG):

[0065] This module is used to capture significant features with less computational cost during feature fusion. The input of this module is the path feature from the model encoder jump connection and the up-sampled feature from the previous layer of the decoder. First, 3x3x3 group 3D convolution (Group Conv3D) is applied to the jump connection feature and the up-sampled feature, with a group number of 8. Then, the outputs of the two group 3D convolutions are batch normalized and added element by element. Finally, the feature passes through a multi-layer perceptron (MLP) composed of two layers of 3D point convolution. The LeakyReLU function is activated in the middle, and the Sigmoid function is used to output the attention coefficient map. The coefficient map is multiplied element by element with the up-sampled feature to obtain the gated feature. Figure 4 The structure diagram of the group large kernel attention gate module is shown in

[0066] The overall architecture of the PLKA-nnUNet model is shown in Figure 5 The skull segmentation results of the PLKA-nnUNet model are compared with six models, namely RATS, SHERM, 3DPCNN, V-Net, SwinUNet, and UNETR. As shown in Figure 6 In the skull removal effect comparison chart, the red rectangle represents the failure area, i.e., the skull tissue is not removed or over-segmented. RATS incorrectly classifies non-brain tissue as brain tissue, while 3DPCNN and SHERM perform better, but the poor image quality and complex brain boundary lead to unstable segmentation. The results of the PLKA-nnUNet proposed in the present application are closest to the true segmentation value, with smoother boundaries and no over-segmentation. In Figure 7 Compared with Six the segmentation accuracy of six models, Dice score, Hausdorff distance, and Jaccard index are used as evaluation indicators, and the results show that the PLKA-nnUNet obtains the optimal score in the three indicators. Figure 8The results of ablation studies show that after adding the PMSLKA module, the scores of Dice and Jaccard are improved by 0.78% and 0.73% respectively. After adding the GLAG module, these data are improved by 0.29% and 0.26% respectively. By combining the two modules, the highest scores of Dice and Jaccard are improved by 1.06% and 1.55% respectively. This confirms that the introduction of the two mechanisms can improve the ability of the model to capture global dependencies and learn important features. In addition, replacing standard convolution with depthwise separable convolution reduces the complexity of the model, although the scores of Dice and Jaccard decrease slightly, but the performance is still an acceptable trade-off. The data show that PLKA-nnUNet extracts the best brain tissue boundary integrity and the least residual non-brain tissue in the skull removal task through the synergistic effect of the above modules, and is suitable for processing magnetic resonance images of rodents such as mice, providing a reliable foundation for subsequent MEMRI data analysis.

[0067] 1.4 Image registration based on mouse standard brain space: The SyNRACC registration method improved based on the ANTs toolbox is used to perform rigid, affine and nonlinear registration in turn, and the individual mouse brain image extracted in step 1.3 is aligned to the AMBMC standard mouse brain atlas template. In the rigid registration stage, the individual mouse brain image is subjected to a six-degree-of-freedom (X, Y, Z translation, rotation around X, Y, Z) spatial transformation by a gradient descent optimizer, with mutual information (Mutual Information, MI) as the similarity measure of the algorithm, to preliminarily eliminate the overall position and direction differences of the image to be registered relative to the template image; In the affine registration stage, scaling and shearing transformations are further introduced based on the rigid registration (a total of 12 degrees of freedom: scaling and shearing transformation for each of the aforementioned six degrees of freedom), a regularized gradient descent optimizer is used, and MI is still used as the similarity measure to correct the global geometric deformation (such as size, tilt difference) between images; In the nonlinear registration stage, the MI similarity measure in the original SyNRA registration method is replaced by normalized cross-correlation (Normalized Cross-Correlation, NCC), with the following specific parameter settings: gradient step size 0.1, regularization Gaussian smoothing parameter σ=3, random coordinate descent method for optimization, and iteration number 100 times. The NCC measure is invariant to intensity linear changes, which can effectively improve the registration accuracy. As shown in the overall effect diagram of the registration method of Figure 9 As shown in the overall effect diagram of the registration method of Figure 10SyNRACC registration and other registration methods in the same data image registration effect comparison, can be seen compared with the registration results based on the traditional SyNRA method, the method has obvious improvement in calibration accuracy.

[0068] 1.5 Intensity normalization: The normalization of intensity values between images is particularly important for the signal intensity analysis of manganese-enhanced magnetic resonance non-quantitative images after registration, but it can be ignored for quantitative imaging methods. The most common and suitable normalization methods in manganese-enhanced magnetic resonance are mean normalization and reference region normalization. Mean normalization refers to calculating the average signal intensity of all voxels (i.e. the average signal intensity of the whole image) for each individual mouse brain image after registration, and then dividing the original intensity value of each voxel in the image by the average signal intensity of the whole image to obtain the normalized intensity value. Reference region normalization is to select a known anatomical region with stable signal intensity under experimental conditions as a reference region, calculate the average intensity value of all voxels in the reference region (i.e. the reference average intensity value), and then divide the intensity value of each voxel in the whole image by the reference average intensity value to obtain the normalized intensity value. This method can unify the brain signal intensity of different individuals to the same level.

[0069] 1.6 Statistical analysis of manganese-enhanced magnetic resonance images: For the T1-mapping quantitative images of manganese-enhanced magnetic resonance images after registration in step 1.4, the corresponding R1 map (relaxation rate map) is obtained by taking the reciprocal; then, in the standard mouse brain atlas template space published by AMBMC, the hippocampal atlas provided by AMBMC is superimposed, and the average relaxation rate R1 of the four hippocampal subregions DG, CA3, CA1, and CA2 of the mouse brain is calculated for quantitative comparison of the activation of the mouse brain. In addition, by using the SPM12 toolkit, specific experimental design matrices (e.g. comparison of treatment groups and control groups, or comparison of multiple time points before and after treatment of the same individual in longitudinal studies) are used for analysis. According to the design matrix, longitudinal comparison between time points and comparison between groups (unpaired t-test) can be performed, and SPM performs voxel-level statistical test on the whole brain image registered and normalized to the standard space to generate a statistical parameter map to identify brain regions with significant signal intensity differences between different experimental conditions. Unpaired t-test is to calculate the difference between groups in the average R1 value of each hippocampal subregion for two independent samples; SPM voxel-level statistics refers to using the SPM12 toolkit to first smooth all R1 maps, and then construct a generalized linear model (GLM) for group comparison.

[0070] Based on the above six steps, the specific and complete manganese-enhanced magnetic resonance image processing flow and operation process are shown in Figure 11 and Figure 12 .

[0071] Example 2: Longitudinal evaluation of mouse brain hippocampus nuclei based on manganese ion metabolic kinetics

[0072] 2.1 Scheme overview: After tail vein injection of MnCl2, T1-mapping is collected at multiple time points; R1 is processed and calculated through the process of Example 1; R1 is extracted within the range of hippocampus atlas (DG, CA3, CA1, CA2) and plotted against time, and the mouse brain manganese ion metabolism experiment process is shown in Figure 13 .

[0073] 2.2 Key phenomena: 2-3 hours after injection, hippocampus starts to absorb manganese ions, and reaches peak at about 4 hours; thereafter, the whole gradually metabolizes, but the DG and CA3 regions remain longer (significant difference signals can still be seen on a scale of up to a week), suggesting that there is heterogeneity in uptake and metabolic kinetics of different nuclei, and the R1 visualization graph of mouse brain Mn2+ changes over time after tail vein injection, and the R1 value change curve and significant difference graph of each hippocampus nucleus region compared to before injection of Mn2+ are shown in Figure 14 and Figure 15 .

[0074] 2.3 Application value: Through the experimental results, about 168 hours after injection can be selected as the best time point for subsequent magnetic stimulation research on mouse brain nuclei activation, reducing the interference of residual manganese ions in non-target brain regions on the experiment. In addition, this experiment can also effectively quantitatively analyze the activity of mouse brain nerves in each nucleus after injection of MnCl2, and explore the manganese ion-related kinetics of each nucleus region in the mouse brain.

[0075] Example 3: Activation of mouse brain hippocampus nuclei under magnetic stimulation based on iron-based magnetic nanomaterials

[0076] 3.1 Scheme overview: The mice are divided into two groups, one group is injected with magnetic nanoparticles and MnCl2 without magnetic stimulation (MS−), and the other group is injected with magnetic nanoparticles and MnCl2 with magnetic stimulation (MS+), each with 7 mice. Within seven days after injection of MnCl2, the magnetic stimulation group of mice is subjected to three days of magnetic stimulation in time periods, each for about 60 minutes, twice a day. The magnetic stimulation device used is a conical rotating magnetic field with a rotating frequency of 20 Hz and a magnetic field size of 100 mT. The entire experimental process is shown in Figure 16 , and the injected magnetic nanoparticles and the mouse brain injection site are shown in Figure 17The magnetic resonance images were collected starting from the 168th hour after injection of MnCl2, and the two sets of data were processed according to the manganese-enhanced magnetic resonance image processing procedure in Example 1 to finally generate manganese-enhanced magnetic resonance images registered to the standard mouse brain template space; then, first, the region of interest (ROI) quantitative measurement was performed: according to the method in Example 2, the R1 average values of the hippocampal DG, CA3, CA1, and CA2 nuclear regions of the two groups of mice were calculated, and a two-group unpaired t test was performed to calculate whether there was a significant difference between the R1 values in the hippocampal nuclear regions of interest and to draw an R1 value comparison bar chart of the hippocampal nuclear regions of interest between the magnetic stimulation group and the non-magnetic stimulation group. In addition, voxel-level statistical comparison and analysis were also performed. Statistical parametric mapping (SPM) was used to perform statistical difference analysis and comparison between individual voxels or corresponding voxels of different groups at the global voxel level to further verify the difference distribution of the hippocampus-related layers;

[0077] 3.2 Key phenomena: visual display shows that the R1 value of the mouse hippocampal DG and CA3 nuclear regions of the magnetic stimulation group (MS+) is significantly higher than that of the non-magnetic stimulation group (MS-), and the ROI significant difference analysis shows that there is a more significant difference in the hippocampal nuclear regions DG and CA3, but no significant difference in the hippocampal nuclear regions CA1 and CA2, and the visual comparison of the mouse hippocampal nuclear regions before and after magnetic stimulation is as shown in Figure 18 , and the significant difference comparison of different hippocampal nuclear regions ROIs is as shown in Figure 19 , and the SPM voxel-level statistical analysis shows that there are more voxels with significant differences in the hippocampal nuclear regions DG and CA3, and the corresponding significant difference map is as shown in Figure 20 .

[0078] 3.3 Conclusion: the magnetic stimulation based on iron-based nanomaterials can enhance the uptake of manganese ions by the mouse hippocampal nuclear regions DG and CA3, slow down the metabolic decline of manganese ions in the hippocampus, and make the neurons in the nuclear region have stronger excitability, which further indicates that the magnetic stimulation based on iron-based nanomaterials can effectively activate the related nuclear regions of the hippocampus.

Claims

1. A method for detecting activation of brain nuclei in mice based on manganese-enhanced magnetic resonance imaging, characterized in that, include: The acquired manganese-enhanced magnetic resonance images of mouse heads were converted from raw DICOM data to NIFTI format, and the orientation and voxel size of the NIFTI format images were kept consistent with the standard mouse brain atlas template. The image is subjected to bias field correction based on the N4 algorithm to remove intensity drift caused by magnetic field inhomogeneity; Load the PLKA-nnUNet model, perform image segmentation on the image, output a binary brain mask, and extract individual mouse brain images from the image after bias field correction by the N4 algorithm. PLKA-nnUNet is based on nnUNet and introduces a parallel large kernel convolution module, a parallel fusion attention module, and a lightweight convolution design. Image registration is achieved by aligning extracted individual mouse brain images to a standard mouse brain atlas template. Intensity normalization was performed on each registered individual mouse brain image; The T1-mapping quantitative map of each registered individual mouse brain image was used to obtain the corresponding relaxation rate map by taking the reciprocal. The hippocampal atlas was then superimposed, and the mean relaxation rate of each of the four hippocampal subregions DG, CA3, CA1, and CA2 in the R1 map of the mouse brain was calculated to quantify and compare the activation status of the mouse brain. Voxel-level statistical tests were performed on the registered and normalized individual mouse brain images to identify brain regions with intensity differences between different experimental conditions. The PLKA-nnUNet model adopts a Net encoder-decoder architecture and has made the following improvements: Each standard convolutional layer of the decoder is followed by a parallel multi-scale large kernel convolutional attention module. The parallel multi-scale large kernel convolutional attention module consists of a sequentially connected parallel large kernel convolutional module and a parallel fusion attention module. The parallel large kernel convolutional module is responsible for extracting multi-scale features and generating attention weights, while the parallel fusion attention module further refines and calibrates features from three dimensions: channel, pixel, and space. A grouped large kernel convolutional attention gating module is set up to fuse encoder features from skip connections and upsampled decoder features during feature fusion in the decoder, and dynamically calibrate the feature flow. At each standard convolutional layer position in the encoder and decoder, depthwise separable convolution is used instead of 3D convolution. Depthwise separable convolution performs depthwise convolution on each input channel independently, and then combines the channel information through point convolution to achieve a lightweight convolution design. The parallel large kernel convolution module first performs batch normalization on the input features. The normalized features are then transformed by two basic convolutional layers: first, a 1×1×1 3D point convolution is used for channel feature mapping and interaction; then, a 5×5×5 3D convolution is used to extract local neighborhood features. The process is then divided into three parallel branches. Each branch uses deep dilated convolutions with different kernel sizes but the same dilation rate to capture contextual information within different receptive fields. The output feature maps of the three branches are then concatenated along the channel dimension. The fused features are then fed into a multilayer perceptron for feature transformation and integration. The output of the multilayer perceptron is then processed by a sigmoid activation function and residually connected to the original input features of this module to complete the first stage of attention guidance. Branch 1 uses a depth-dilated convolution with a kernel size of 19×19×19 to capture dependencies in the image; Branch 2 uses a depth-dilated convolution with a kernel size of 13×13×13 to balance contextual information and structural details; Branch 3 uses a depth-dilated convolution with a kernel size of 7×7×7 to extract local detail features.

2. The method for detecting activation of mouse brain nuclei based on manganese-enhanced magnetic resonance imaging according to claim 1, characterized in that, The parallel fusion attention module first performs batch normalization on the input features, and then inputs the features in parallel to three independent attention branches: channel attention module, pixel attention module and spatial attention module; The channel attention module first performs global average pooling on each channel in the spatial dimension. Then, it inputs these channel-level features into a sub-network consisting of two point convolutional layers. Finally, it generates the final channel attention weight vector through the Sigmoid activation function. The channel attention weight vector is then multiplied channel by channel with the original input features of the channel attention module. The pixel attention module learns the importance weights of each spatial location through a sub-network consisting of two point convolutional layers. The ReLU activation function is used between layers, and the output is activated by the Sigmoid activation function. The resulting pixel-level attention weight map is multiplied element-wise with the original input features of the pixel attention module. The spatial attention module first performs global average pooling and global max pooling on the feature maps in the channel dimension to obtain two single-channel spatial feature maps. Then, these two spatial feature maps are concatenated in the channel dimension and input into a 7×7 convolutional layer. Finally, after passing through the Sigmoid activation function, they are multiplied point by point with the original feature map input to the spatial attention module. The output feature maps of the three branches are concatenated along the channel dimension. The concatenated features are then fused and dimensionality reduced by two consecutive 3D point convolutional layers. The LeakyReLU activation function is used between the two 3D point convolutional layers, and finally the Sigmoid activation function is used. The feature maps are then connected with the input features of the parallel fusion attention module through an element-wise addition residual connection, and finally output a feature map that has been calibrated by multiple attention.

3. The method for detecting activation of mouse brain nuclei based on manganese-enhanced magnetic resonance imaging according to claim 1, characterized in that, The grouped large kernel convolutional attention gating module first applies grouped 3D convolution to the skip connection features and upsampled features respectively. Then, the outputs of the two grouped 3D convolutions are batch normalized. The results are added element-wise and passed through a multilayer perceptron. Finally, the attention coefficient map is output by the Sigmoid function. This coefficient map is multiplied element-wise with the upsampled features to obtain the gated features.

4. The method for detecting activation of mouse brain nuclei based on manganese-enhanced magnetic resonance imaging according to claim 1, characterized in that, The image registration stage performs rigid, affine, and nonlinear registration sequentially: In the rigid registration stage, the individual mouse brain images are subjected to six degrees of freedom spatial transformation by the gradient descent optimizer. Mutual information is used as the similarity metric to initially eliminate the overall position and orientation differences between the images to be registered and the template images. In the affine registration stage, scaling and shearing transformations are further introduced on the basis of rigid registration. A regularized gradient descent optimizer is adopted, and mutual information is still used as the similarity metric to correct the global geometric deformation between images. In the nonlinear registration stage, the mutual information similarity measure in the original SyNRA registration method is replaced with normalized cross-correlation.

5. The method for detecting activation of mouse brain nuclei based on manganese-enhanced magnetic resonance imaging according to claim 1, characterized in that, Intensity normalization uses mean normalization: For each registered individual mouse brain image, the average signal intensity of all voxels is calculated, and then the original intensity value of each voxel in the image is divided by this average signal intensity to obtain the normalized intensity value.

6. The method for detecting activation of mouse brain nuclei based on manganese-enhanced magnetic resonance imaging according to claim 1, characterized in that, Intensity normalization employs reference region normalization: an anatomical region with stable signal intensity under known experimental conditions is selected as the reference region, the average intensity value of all voxels within the reference region is calculated, and then the intensity value of each voxel in the entire image is divided by the average intensity value to obtain the normalized intensity value, thus unifying the brain signal intensity of different individuals to the same level.

Images