Stimulation Target Location System and Method Based on Point Cloud Deep Learning

Abstract

The present invention provides a stimulation target positioning system and method based on point cloud deep learning, including a data set production module configured to be connected to a collection device for constructing a data set by using individual MRI images and individual point cloud data; a deep learning module provided on a processor for training and evaluating a pre-constructed deep learning model by using the data set to obtain an optimally evaluated deep learning model; and a target prediction module for predicting current point cloud data collected from the collection device by using the optimally evaluated deep learning model to locate a stimulation target and displaying the located stimulation target on the corresponding individual point cloud data. The present invention saves the process of repeatedly collecting individual MRI nuclear magnetic images, alleviates the process of complex image processing using MRI nuclear magnetic images in the past, and reduces the complexity of determining the stimulation target by using deep learning of point clouds for prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical imaging, and particularly relates to a stimulation target positioning system and method based on point cloud deep learning. Background Art

[0002] Transcranial Magnetic Stimulation (TMS) is a non-invasive neuromodulation technique. In the application of TMS, accurately determining the position of the stimulation target is very important to ensure the accuracy and effectiveness of the stimulated target area.

[0003] Deep learning is a machine learning method that processes and analyzes large-scale data by simulating the working mode of the human brain neural network. It uses a neural network structure to learn and extract features from data and is commonly used in image recognition, natural language processing, prediction, etc. The advantage of using deep learning is that it can learn complex feature representations from raw data. With the rapid development of deep learning in various fields, many scholars have studied the application of deep learning in the field of medical images. The prediction of medical images is an important research direction of deep learning in the medical field. By using a deep learning model to analyze and predict medical images. The prediction of medical images based on deep learning requires a large-scale and high-quality medical image dataset, and professional doctors need to accurately annotate the images.

[0004] By using anatomical landmark points for positioning, initially in the clinical research of TMS, scalp measurements were used to locate the TMS stimulation target. This method relies on specific landmark points on the head anatomy. This method is relatively simple, but individual differences may lead to lower accuracy.

[0005] There is also a method of directly using the target area already marked on the 10-20 electroencephalogram positioning cap by wearing it. The advantage of this method is that the positioning cap can be reused, but due to the different head sizes and shapes of each person, there are some errors when using the standard 10-20 electroencephalogram positioning cap.

[0006] With the development of brain imaging technology, Magnetic Resonance Imaging (MRI) has played an important role in determining the stimulation target. By obtaining the head MRI nuclear magnetic resonance images of the user, the position and structure of the target brain region can be accurately determined. The user's MRI nuclear magnetic resonance images can be registered with standard nuclear magnetic resonance images (such as ICBM152, Colin27, etc.) and standard brain atlas images (such as AAL template, Brodmann template, etc.). By registering the standard brain region images in the standard brain atlas to the user's MRI nuclear magnetic resonance images, the determination of the stimulation target can be achieved according to the standard brain atlas template. However, due to the differences between the standard brain atlas and the individual brain, and the process of image registration requires a more accurate registration algorithm, a more accurate method is also needed to determine the TMS stimulation target.

[0007] Since functional Magnetic Resonance Imaging (fMRI) can provide information on brain functional connectivity, processing the task-state fMRI images can obtain the corresponding task-state activation maps; combining with the MRI nuclear magnetic resonance images to construct the user's brain activation map and mark the point coordinates, thereby determining the transcranial magnetic stimulation target. Using fMRI to determine the stimulation target usually requires combining other anatomical or functional information to reduce errors and ensure the accuracy of positioning. The cost of collecting MRI images and fMRI images is high, and a new positioning method is needed to reduce the cost of collecting MRI images and fMRI images and obtain more accurate positioning.

[0008] The purpose of determining the transcranial magnetic stimulation target is to ensure that the stimulation is accurately applied to the target brain region to achieve the desired neuromodulation effect. During TMS, the stimulation of non-target regions may cause unexpected effects or adverse reactions. Accurately determining the stimulation target helps to avoid the stimulation acting on non-target regions, thereby reducing unnecessary interference and side effects. Currently, the 10-20 electroencephalogram positioning system is usually used for the positioning of the stimulation target. Due to the differences in the brain structure and size of each individual, it may not be able to fully match the individual's anatomical structure. Summary of the Invention

[0009] To solve the above problems existing in the prior art, the present invention provides a stimulation target positioning system and method based on point cloud deep learning. The technical problems to be solved by the present invention are achieved through the following technical solutions:

[0010] The present invention provides a stimulation target positioning system based on point cloud deep learning, including:

[0011] A dataset creation module configured to be connected to a collection device, which is used to obtain multiple individual MRI images and corresponding individual point cloud data from the collection device, and construct a dataset by using the individual MRI images and the individual point cloud data;

[0012] A deep learning module set on a processor, which is used to train and evaluate a pre-constructed deep learning model by using the dataset to obtain an optimally evaluated deep learning model;

[0013] A target prediction module configured to be connected to the collection device and a display device, which is used to predict the current point cloud data collected from the collection device by using the optimally evaluated deep learning model, locate the stimulation target, and display the located stimulation target on the corresponding individual point cloud data through the display device.

[0014] The present invention provides a method for locating a stimulation target based on point cloud deep learning, which is realized by using a system for locating a stimulation target based on point cloud deep learning.

[0015] Beneficial effects:

[0016] The present invention provides a system and method for locating a stimulation target based on point cloud deep learning, including a dataset creation module configured to be connected to a collection device, which is used to form a dataset by using individual MRI images and individual point cloud data; a deep learning module set on a processor, which is used to train and evaluate a pre-constructed deep learning model by using the dataset to obtain an optimally evaluated deep learning model; a target prediction module, which is used to predict the current point cloud data collected from the collection device by using the optimally evaluated deep learning model, locate the stimulation target, and display the located stimulation target on the corresponding individual point cloud data. The present invention saves the process of repeatedly collecting individual MRI nuclear magnetic resonance images, reduces the process of complex image processing using MRI nuclear magnetic resonance images in the past, and reduces the complexity of determining the stimulation target by using deep learning of point clouds for prediction.

[0017] The following will further describe the present invention in detail with reference to the drawings and embodiments. Description of the drawings

[0018] Figure 1 is an overall process schematic diagram of the operation of a system for locating a stimulation target based on point cloud deep learning provided by the present invention;

[0019] Figure 2 is a process schematic diagram of a cortical mapping algorithm provided by the present invention. Specific embodiments

[0020] The following further describes the present invention in detail with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0021] Combined with Figure 1 and Figure 2 , the present invention provides a stimulation target localization system based on point cloud deep learning, comprising:

[0022] A dataset production module configured to be connected to an acquisition device, for obtaining multiple individual MRI images and corresponding individual point cloud data from the acquisition device, and constructing a dataset by using the individual MRI images and the individual point cloud data;

[0023] A deep learning module provided on a processor, for training and evaluating a pre-constructed deep learning model by using the dataset, and obtaining an optimally evaluated deep learning model;

[0024] A target prediction module configured to be connected to the acquisition device and a display device, for predicting the currently acquired point cloud data from the acquisition device by using the optimally evaluated deep learning model, locating the stimulation target, and displaying the located stimulation target on the corresponding individual point cloud data through the display device.

[0025] The present invention provides a stimulation target localization system based on point cloud deep learning, comprising

[0026] A dataset production module configured to be connected to an acquisition device, for constructing a dataset by using individual MRI images and individual point cloud data; a deep learning module provided on a processor, for training and evaluating a pre-constructed deep learning model by using the dataset, and obtaining an optimally evaluated deep learning model; a target prediction module, for predicting the currently acquired point cloud data from the acquisition device by using the optimally evaluated deep learning model, locating the stimulation target, and displaying the located stimulation target on the corresponding individual point cloud data. The present invention saves the process of repeatedly acquiring individual MRI nuclear magnetic images, reduces the process of complex image processing using MRI nuclear magnetic images in the past, and reduces the complexity of determining the stimulation target by using deep learning of point clouds.

[0027] In a specific embodiment of the present invention, the dataset production module is specifically configured to:

[0028] S110, acquiring multiple individual MRI images and point cloud data from the acquisition device;

[0029] The present invention can acquire individual nuclear magnetic images of a certain number of subjects, and use an infrared binocular camera to acquire the individual point cloud data of the subjects. In order to facilitate subsequent processing, it is necessary to convert the DICOM file of the original MRI image into a NIfTI file to facilitate the preprocessing of the MRI nuclear magnetic image.

[0030] S120. Preprocess each individual MRI image to obtain a preprocessed individual MRI image, and save the individual MRI image in a file format;

[0031] During the process of acquiring MRI images, due to factors such as random interference in the imaging device, the obtained MRI images will be interfered by noise and distortion. Therefore, preprocessing can remove the interference.

[0032] To facilitate subsequent determination of the stimulation target in the individual space, individual head model reconstruction is required.

[0033] S130. Mark the stimulation target in the preprocessed individual MRI image, and form a point set of the coordinates of the stimulation target and save it as a label file;

[0034] S140. Convert the marked stimulation target to the individual point cloud data space to convert the label file into point cloud data to obtain the labeled point cloud data of the stimulation target;

[0035] The present invention converts the marked stimulation target obtained by using the individual nuclear magnetic image to the acquired individual point cloud data space, and converts the stimulation target label file into point cloud data.

[0036] S150. Combine the individual point cloud data and the labeled point cloud data of the marked stimulation target to form a data set.

[0037] The data set of the present invention is composed of individual point cloud data and labeled point cloud data of the marked stimulation target, and the format of the data set is a point cloud PLY file.

[0038] As a specific implementation manner of the present invention, S120 includes:

[0039] S121. Denoise each individual MRI image using median filtering to remove the salt-and-pepper noise in the individual MRI image;

[0040] The present invention first needs to use median filtering to remove the salt-and-pepper noise in the MRI image to enhance the image effect, and can well retain the edges and details of the image. Secondly, due to the uneven gray distribution caused by using magnetic resonance scanning, bias field correction is required.

[0041] S122. Select the starting seed points in the individual MRI image after removing noise, and use the isosurface extraction algorithm to extract the scalp file from the individual MRI image according to the set extraction value;

[0042] S123. Set the neighborhood range, and search for target points with similar pixels within the neighborhood range of the starting seed points;

[0043] S124. Connect the searched target points to obtain a target region, and segment the target region from the individual MRI image to obtain a target region image;

[0044] The target region image is the image of three parts: gray matter, white matter, and cerebrospinal fluid.

[0045] S125. Use a three-dimensional reconstruction algorithm to perform three-dimensional reconstruction on the target region image to obtain a three-dimensional reconstruction image, and determine the three-dimensional reconstruction image as the preprocessed individual MRI image.

[0046] To facilitate the subsequent determination of the stimulation target in the individual space, an individual head model reconstruction is required. First, the process of segmenting gray matter, white matter, and cerebrospinal fluid is implemented using a region growing algorithm, where the region growing algorithm includes isolated connection and confidence connection. First, select a starting seed point in the individual nuclear magnetic resonance image, that is, a certain point on the target region, and the seed point has obvious features; then, by setting the neighborhood range, search for points with similar pixels within the neighborhood range of the starting seed point; finally, connect the searched points, and finally segment the target region. Second, the process of segmenting the scalp and skull is implemented using an isosurface extraction algorithm. By setting a specific extraction value, the input individual nuclear magnetic resonance image is extracted, and the part of the image equal to the extraction value is extracted and retained to obtain the corresponding scalp and skull images. Finally, perform a three-dimensional reconstruction algorithm on the images obtained after the above segmentation respectively to obtain a three-dimensional reconstruction image. Since the surface of the three-dimensional reconstruction image is not smooth, the Laplacian smoothing algorithm is used for mesh smoothing in this process, and a smoothed image can be obtained. The smoothing iteration times and relaxation factor can be set, where the larger the iteration times, the better the smoothing effect, but too many iterations will cause some details to be lost. At this time, the preprocessing of the individual nuclear magnetic resonance image is completed.

[0047] As a specific implementation manner of the present invention, S130 includes:

[0048] S131. Obtain a registration matrix through the registration of the individual MRI image and the standard MRI image, and use the registration matrix to convert the coordinates of the standard MRI image from the coordinates in the MNI space to the individual space, so as to obtain the coordinates of the standard MRI image converted to the individual space; wherein, the standard MRI image carries multiple given points;

[0049] Currently, most of the commonly used public data, literature on cortical activation regions, and standard brain atlas templates are in the standard space (MNI). It is necessary to convert the coordinates in the MNI space to the individual space: through the registration of the individual nuclear magnetic resonance image and the standard nuclear magnetic resonance image, obtain a registration matrix, and use the registration matrix to convert the coordinates in the MNI space to the individual space.

[0050] S132. Convert the scalp file into scalp point cloud data by reading the scalp file, and use the cortical mapping algorithm to map a given point onto the scalp to determine the surface points on the scalp point cloud data;

[0051] By performing a mapping algorithm on the activation region and the cortical stimulation region in the individual space, obtain the point coordinates on the scalp, thereby determining the stimulation target of the individual MRI image.

[0052] Use the activation regions of standard brain atlases (such as AAL template, Brodmann template, brain network group atlas of the Institute of Automation, Chinese Academy of Sciences, etc.), the heart-brain coupling and stomach-brain coupling activation regions, and the commonly used target coordinates (10-20 system), etc. as transcranial magnetic stimulation targets. Mark the stimulation targets for all individual MRI images, and form a point set with these target coordinates and save it as a spherical NIfTI file.

[0053] S133. Determine the surface points on the scalp point cloud data as the stimulation targets of the individual MRI image.

[0054] As a specific embodiment of the present invention, S132 includes:

[0055] S1321. Convert the scalp file into scalp point cloud data by reading the scalp file;

[0056] S1322. Use the point cloud convex hull algorithm to find the surface points of the scalp point cloud data;

[0057] S1323. Use the nearest neighbor search algorithm to search for the surface point closest to the given point;

[0058] S1324. Determine the closest surface point as the stimulation target of the individual MRI image.

[0059] As a specific embodiment of the present invention, S1323 includes:

[0060] S13231. Set the search range for the given point, and search for the K nearest neighbor points of the given point according to the search range;

[0061] S13232. Determine the center point of the K nearest neighbor points by calculating the average distance, and use the center point as the mapping point;

[0062] S13233. Calculate the distance from the given point to the mapping point, and determine the mapping point with the shortest distance as the surface point closest to the given point.

[0063] Reference Figure 2 , use the nearest neighbor search algorithm to search for the given point (i.e., the known cortical point), and the specific steps to determine the stimulation target are as follows:

[0064] (a) Determine the K nearest points for a given point by setting the search range, calculate the average value of the K nearest points as the final result obtained from the search (i.e., the mapped point), and calculate the distance from the given point to the mapped point;

[0065] (b) Define a structure to store the search results, including the K value for each search range, the coordinates of the mapped point, and the distance;

[0066] (c) Traverse the search results for all search ranges and find the search result with the optimal distance;

[0067] (d) Output the optimal search result.

[0068] As a specific implementation manner of the present invention, the deep learning module is specifically used for:

[0069] Use a network improved based on Unet as the deep learning model architecture to construct a deep learning model; wherein, the improved network structure uses an attention mechanism;

[0070] Iteratively train the deep learning model using a dataset, and obtain the deep learning model after each iterative training. During each iterative training process, optimize the parameters of the deep learning model through RAdam, and use the cross-entropy loss function as the loss function to measure the difference between the predicted stimulation target result output by the model during the training process and the true stimulation target label;

[0071] Use a part of the data samples in the dataset as the test set;

[0072] Evaluate the deep learning model after each iterative training using the test set to obtain the evaluation result;

[0073] Select the deep learning model with the optimal evaluation result.

[0074] The implementation process of the deep learning module of the present invention is mainly divided into three parts:

[0075] (1) Construct a deep learning model: Use a network improved based on Unet as the deep learning model architecture to perform image fusion to improve the prediction accuracy. The improved network structure uses an attention mechanism (Attention Mechanism), and the attention mechanism can make the deep learning model pay more attention to the most relevant regions or features in the point cloud data, improving the prediction performance. At the same time, add a residual convolutional layer (ResidualBlock) with a residual connection to each convolutional layer of the original Unet network to improve the training and convergence effect of the network.

[0076] (2) Model training: Use the individual nuclear magnetic resonance image dataset with labeled stimulation targets to train the deep learning model. Optimize the model parameters through RAdam to provide better convergence performance and generalization ability. Use the cross-entropy loss function as the loss function, which can measure the difference between the predicted stimulation target results output by the model during training and the true stimulation target labels. Set the learning rate to enable the model to converge quickly and obtain good performance.

[0077] (3) Model test and evaluation: Use a part of the dataset as an independent test dataset to evaluate the performance of the trained deep learning model. Use evaluation metrics to evaluate the accuracy of the model in predicting stimulation targets. For example, use IOU (Intersection over Union), which is an evaluation metric for calculating the overlap degree between the model prediction and the true annotation in the prediction of stimulation targets. In the calculation formula (IOU = Intersection / Union), Intersection refers to the area of the intersection part of the prediction result and the true result, and Union refers to the area of the union part of the prediction result and the true result, to obtain the model with the optimal training result.

[0078] As a specific implementation manner of the present invention, the target prediction module is specifically used for:

[0079] Use an infrared binocular camera and a positioning tool to collect the current point cloud data;

[0080] Input the current point cloud data into the optimal deep learning model, and use the optimal deep learning model to output the predicted stimulation target of the current point cloud data;

[0081] Determine the predicted stimulation target as the truly located stimulation target, and display the located stimulation target on the corresponding individual point cloud data through a display device.

[0082] The present invention uses an infrared binocular camera and a positioning tool to collect the point cloud data of the user, and uses the optimally trained deep learning model to predict the point cloud data of the user: Input the point cloud data of the user into the deep learning model, and the result is that the predicted stimulation target image of the user will be output. Finally, obtain the stimulation target of the user and import it into the navigation system to achieve precise positioning.

[0083] The present invention provides a method for locating stimulation targets based on deep learning, which is implemented by using a stimulation target positioning system based on point cloud deep learning.

[0084] The present invention provides a method for localizing stimulation targets based on deep learning, which constructs a data set using individual MRI images; trains and evaluates a pre-constructed deep learning model using the data set to obtain the deep learning model with the optimal evaluation; uses the deep learning model with the optimal evaluation to predict the current point cloud data collected from the acquisition device, localizes the stimulation targets, and displays the localized stimulation targets on the corresponding individual point cloud data. The present invention saves the process of repeatedly collecting individual MRI nuclear magnetic resonance images, reduces the process of complex image processing using MRI nuclear magnetic resonance images in the past, and reduces the complexity of determining stimulation targets by using deep learning for prediction.

[0085] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0086] Although the present application has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed present application, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the accompanying drawings, the disclosure content, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality of cases.

[0087] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A stimulation target positioning system based on point cloud deep learning, characterized in that, Including: A dataset production module configured to be connected to a collection device, for obtaining multiple individual MRI images and corresponding individual point cloud data from the collection device, and constructing a dataset by using the individual MRI images and the individual point cloud data; A deep learning module set on a processor, for training and evaluating a pre-constructed deep learning model by using the dataset to obtain an optimally evaluated deep learning model; A target prediction module configured to be connected to the collection device and a display device, for predicting the current point cloud data collected from the collection device by using the optimally evaluated deep learning model, locating a stimulation target, and displaying the located stimulation target on the corresponding individual point cloud data through the display device; The dataset production module is specifically configured to: S110, collect multiple individual MRI images and point cloud data from the collection device; S120, preprocess each individual MRI image to obtain a preprocessed individual MRI image, and save the individual MRI image in a file format; S130, mark the stimulation target in the preprocessed individual MRI image, and form a point set of the coordinates of the stimulation target and save it as a label file; S140, convert the marked stimulation target to the individual point cloud data space to convert the label file into point cloud data to obtain label point cloud data of the marked stimulation target; S150, form a dataset by using the individual point cloud data and the label point cloud data of the marked stimulation target; S130 includes: S131, obtain a registration matrix through the registration of the individual MRI image and a standard MRI image, and use the registration matrix to convert the coordinates of the standard MRI image from the coordinates in the MNI space to the individual space to obtain the coordinates of the standard MRI image converted to the individual space; wherein, the standard MRI image carries a plurality of given points; S132, convert the scalp file into scalp point cloud data by reading the scalp file, and use a cortical mapping algorithm to map the given points onto the scalp to determine surface points on the scalp point cloud data; S133, determine the surface points on the scalp point cloud data as the stimulation target of the individual MRI image.

2. The stimulation target positioning system based on point cloud deep learning according to claim 1, wherein S120 includes: S121, perform denoising on each individual MRI image by using median filtering to remove salt-and-pepper noise in the individual MRI image; S122, select starting seed points in the individual MRI image after removing noise, and use an isosurface extraction algorithm to extract a scalp file from the individual MRI image according to a set extraction value; S123, set a neighborhood range, and search for target points with similar pixels within the neighborhood range of the starting seed points; S124, connect the searched target points to obtain a target region, and segment the target region from the individual MRI image to obtain a target region image; S125, perform three-dimensional reconstruction on the target region image by using a three-dimensional reconstruction algorithm to obtain a three-dimensional reconstruction image, and determine the three-dimensional reconstruction image as the preprocessed individual MRI image.

3. The stimulation target positioning system based on point cloud deep learning according to claim 1, wherein S132 includes: S1321, convert the scalp file into scalp point cloud data by reading the scalp file; S1322. Use the point cloud convex hull algorithm to find the surface points of the scalp point cloud data; S1323. Use the nearest neighbor search algorithm to search for the surface points closest to a given point; S1324. Determine the closest surface point as the stimulation target of the individual MRI image.

4. The stimulation target positioning system based on point cloud deep learning according to claim 3, wherein S1323 includes: S13231. Set the search range for the given point and search for the K nearest neighbor points of the given point according to the search range; S13232. Determine the center point of the K nearest neighbor points by calculating the average distance, and use the center point as the mapping point; S13233. Calculate the distance from the given point to the mapping point, and determine the mapping point with the shortest distance as the surface point closest to the given point.

5. The stimulation target positioning system based on point cloud deep learning according to claim 1, wherein The deep learning module is specifically used for: Use a network improved based on Unet as the deep learning model architecture to build a deep learning model; among them, the improved network structure uses an attention mechanism; Use the data set to iteratively train the deep learning model to obtain the deep learning model after each iterative training. During each iterative training process, optimize the parameters of the deep learning model through RAdam, and use the cross-entropy loss function as the loss function to measure the difference between the predicted stimulation target result output by the model during the training process and the true stimulation target label; Use a part of the data samples in the data set as the test set; Use the test set to evaluate the deep learning model after each iterative training to obtain the evaluation result; Select the deep learning model with the best evaluation result.

6. The stimulation target positioning system based on point cloud deep learning according to claim 5, characterized in that, The evaluation result is the degree of overlap between the predicted stimulation target and the truly labeled stimulation target.

7. The stimulation target positioning system based on point cloud deep learning according to claim 1, characterized in that, The target prediction module is specifically used for: Use an infrared binocular camera and a positioning tool to collect the current point cloud data; Input the current point cloud data into the optimal deep learning model, and use the optimal deep learning model to output and predict the stimulation target of the current point cloud data; Determine the predicted stimulation target as the truly located stimulation target, and display the located stimulation target on the corresponding individual point cloud data through the display device.

8. A method for localizing stimulation targets based on point cloud deep learning, characterized in that, Implemented by using the stimulation target positioning system based on point cloud deep learning according to any one of claims 1 to 7.

Images