A Smart Diagnostic Analysis Method for Parkinson's Disease Based on MRI Images

Abstract

This invention discloses an intelligent diagnostic analysis method for Parkinson's disease based on MRI images. First, medical image data is read. A DICOM or DaT file is selected using a file selector. After acquiring the selected file, the software starts a file processing thread to obtain information such as voxels, spatial spacing, window width, and window level from the medical image data, and then visualizes this information. Next, a lesion region detection module is used. A YOLO v5 model is used to detect lesions, outputting a disease risk score and the lesion region, which is then visualized. Finally, the lesion region is segmented, refining DaT image features. A U-Net network is used to segment the striatum region, dividing the feature region into the right substantia nigra for more precise localization of the lesion area, and then visualizing this information. This invention reduces computational load and increases speed, thereby assisting doctors in making more accurate diagnoses.

Description

Technical Field

[0001] This invention belongs to the field of biomedical electro-technical technology, specifically relating to an intelligent diagnostic and analysis method for Parkinson's disease. Background Technology

[0002] Magnetic resonance imaging (MRI) is a type of tomographic imaging. Newer MRI techniques, such as susceptibility-weighted imaging (SWI), can quantify magnetic susceptibility to visually represent the iron content of tissues, while increasingly powerful mathematical methods can characterize the topological structure of brain networks at the system level.

[0003] Early detection and treatment of Parkinson's disease (PD) can improve patients' quality of life and slow disease progression, which is of great practical significance. The development of functional magnetic resonance imaging (fMRI) of the brain has made the detection of degenerative lesions, including those in Parkinson's disease, more sensitive and intuitive, providing important diagnostic evidence for clinical practice and playing a crucial role in the early diagnosis and treatment of PD. Because the early clinical manifestations of PD are not obvious, the rates of missed diagnosis and misdiagnosis are high. With the continuous development of medical imaging technology in recent years, many studies have shown that imaging applications in the diagnosis of PD have characteristic changes and have become an indispensable auxiliary examination method for PD. The new MRI method, susceptibility-weighted imaging (SWI), can quantify magnetic susceptibility to visually represent the iron content of tissues, while increasingly powerful mathematical methods can characterize the topological structure of brain networks at the system level. This new method is receiving increasing attention in the early diagnosis of Parkinson's disease.

[0004] However, there are still many problems in applying MRI images to diagnose PD. (1) The resolution and contrast of brain MRI medical images used in hospitals for diagnosis and treatment are low, making it difficult to capture subtle differences between lesion features; (2) The lesion areas in MRI images of PD patients are smaller and less obvious compared to other brain diseases; (3) MRI image data of PD patients are scarce and difficult to collect, and labeled PD patient data is more valuable. Therefore, using existing models to classify and detect MRI images results in poor classification results and inaccurate labeled lesion areas, leading to low confidence in model diagnosis of PD.

[0005] Brief introduction of existing technology:

[0006] [1] Zhang Xiaobo, Zhang Zhehao, Li Wei. A method for classifying Parkinson's disease and annotating lesion regions in MRI images [P]. Sichuan Province: CN113269774B, 2022-04-26.

[0007] This invention discloses a method for classifying and labeling Parkinson's disease (PD) lesion regions in MRI images, comprising the steps of: constructing a PD classification model for MRI images; constructing a PD lesion labeling model; classifying the MRI image to be tested using the PD classification model; and labeling the lesion regions in MRI images classified as PD or PD latency images using the PD lesion labeling model. The step of constructing the PD classification model for MRI images includes constructing a classification module and inputting the training set into the classification module, updating the network parameters through backpropagation to obtain the MRI image classification model. The beneficial effects of this invention are: by using a classification-then-labeling method, the accuracy of the classification model is improved, resulting in more accurate and effective localization of lesion regions, while avoiding unnecessary training, greatly improving training efficiency, and making the model more robust.

[0008] The lesion areas in MRI images of PD patients are smaller than those of other brain diseases, and the lesion features are not obvious. Traditional pixel annotation is difficult to capture subtle differences and it is difficult to accurately locate small lesion areas.

[0009] In recent years, cutting-edge technologies such as big data and artificial intelligence have been gradually applied to the medical field. With the rise of deep neural networks and 3D visualization technology, new medical imaging technologies are being combined with artificial intelligence for more medical-assisted diagnosis and evaluation. In particular, deep learning technology, which has emerged in the field of artificial intelligence in recent years, automatically learns the inherent patterns and representational levels of sample data by constructing neural networks with deep structures to achieve abstract feature modeling. It has a deeper structure and more powerful feature fitting capabilities, and can design features without prior knowledge, achieving fully automated target detection and recognition. It has wide applications in fields such as computer vision. Applying artificial intelligence to the evaluation and diagnosis of Parkinson's disease has significant theoretical and practical value for early diagnosis and treatment, improving patients' quality of life, and reducing their economic burden. Summary of the Invention

[0010] To overcome the shortcomings of existing technologies, this invention provides an intelligent diagnostic analysis method for Parkinson's disease based on MRI images. First, medical image data is read. A DICOM or DaT file is selected using a file selector. After acquiring the selected file, the software starts a file processing thread to obtain information such as voxels, spatial spacing, window width, and window level from the medical image data, and then visualizes this information. Next, a lesion region detection module is used, employing a YOLO v5 model to detect lesions, outputting a disease risk score and the lesion region, and visualizing this information. Finally, the lesion region is segmented, refining DaT image features. A U-Net network is used to segment the striatum region, dividing the feature region into the right substantia nigra for more precise localization of the lesion area, and then visualizing this information. This invention reduces computational load and increases speed, thereby assisting doctors in making more accurate diagnoses.

[0011] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0012] Step 1: Obtain voxel, spatial spacing, window width and window level information of medical image data, and visualize it;

[0013] Step 1-1: Read MRI image data and DaT data from DICOM format medical images, and extract features from the two types of data using different feature extraction methods;

[0014] Steps 1-2: Use the FSL-bet tool to perform craniotomy preprocessing on the MRI image data;

[0015] Steps 1-3: Obtain information on voxels, spatial spacing, and 3D image size of medical images;

[0016] Steps 1-4: Display the retrieved medical image data;

[0017] Step 2: Lesion area detection module, using the YOLO v5 model to detect lesion areas, outputting disease risk scores and lesion areas, and displaying them visually;

[0018] Step 2-1: Convert MRI data and DaT data into PNG format images, annotate the red nucleus and substantia nigra regions of the brain MRI images, and divide the annotated brain MRI image data into training set, test set and validation set;

[0019] Step 2-2: Use a localization model based on the improved YOLOv5 network to localize the red nucleus and substantia nigra regions in brain MRI images;

[0020] The localization model based on the improved YOLOv5 network includes an input module, an improved Backbone module, a Neck module, a YOLO head module, and a localization module. The input module uses the Mosaic data augmentation algorithm. The improved Backbone module includes a Focus layer, two sets of CBL and CSP_1, and one set of CBL and SPP. The Neck module includes an FPN+PAN structure and two sets of CBL+CSP2_X. The YOLO head module includes two-scale detection heads. The localization module includes non-maximum suppression, IOU thresholding, evaluation of location loss using the DIOU loss function, and evaluation of object category confidence loss for ground truth and predicted bounding boxes using the BCE loss function. The ground truth bounding box refers to the minimum bounding rectangle corresponding to the red nucleus and substantia nigra region in a brain MRI image.

[0021] Steps 2-3: Input the test set into the localization model based on the improved YOLOv5 network to complete the localization of the red nucleus and substantia nigra region in the MRI image and extract the relative position information of the predicted box in the localization result; the predicted box refers to the rectangular border generated by the YOLOv5 network model, which is used to select the localized red nucleus and substantia nigra region in the MRI image.

[0022] A classification model based on a one-dimensional convolutional neural network obtains the relative position information of the prediction box. After coordinate transformation and region extraction, the red nucleus and substantia nigra region within the prediction box of the brain MRI image is obtained, the substantia nigra region is located, and its health status is displayed. The one-dimensional convolutional neural network model reduces the dimensionality of the obtained average spectral value through one-dimensional convolution, and finally detects the substantia nigra region. The mAP value is used as an evaluation index for the network to predict the disease score and determine whether it is healthy.

[0023] Steps 2-4: Use a YOLOv5-based detection network model to locate the red nucleus and substantia nigra regions in the DaT data; detect the red nucleus and substantia nigra regions and display the disease score to identify whether they are healthy;

[0024] The YOLOv5-based detection network model comprises three parts: a backbone, a neck, and an output module. The backbone includes a BottleneckCSP module and a Focus module. The BottleneckCSP module enhances the learning performance of the entire convolutional neural network. The Focus module slices the image, expanding the input channels to four times their original size, and then performs a convolution to obtain a downsampled feature map. The neck employs a structure combining FPN and PAN, integrating the FPN layer with a bottom-up feature pyramid to fuse extracted semantic and positional features, while also fusing features between the backbone layer and the detection layer. The output module predicts image features and outputs a vector containing the target object's class probability, object score, and the location of its bounding box.

[0025] Step 3: Lesion area segmentation module: refine DaT image features, use U-Net network to segment the striatum region, divide the feature region into the right substantia nigra to locate the lesion area, and visualize the display.

[0026] Step 3-1: Based on DaT data, the model outputs the segmentation and recognition results of the left and right striatum respectively; the prediction results of DaT data YOLOv5 and U-Net are combined for judgment;

[0027] Step 3-2: Prediction results of UNET using EfficientNetB4 as the backbone; the segmentation map has 5 different categories: background, healthy left striatum, healthy right striatum, unhealthy left striatum, and unhealthy right striatum.

[0028] The beneficial effects of this invention are as follows:

[0029] (1) The method of the present invention can accurately identify changes in the lesion area of ​​Parkinson's disease in the brain, which can avoid the situation where doctors may overlook some minor lesions due to fatigue or lack of experience during direct observation, leading to inaccurate final diagnosis.

[0030] (2) In the method of the present invention, the BottleneckCSP module significantly reduces the amount of computation while enhancing the learning performance of the entire convolutional neural network; the Focus module performs slicing operation on the image, expands the input channel to 4 times the original, and obtains the downsampled feature map through one convolution, which reduces the amount of computation and improves the speed while realizing downsampling, thereby assisting doctors to complete the diagnosis of the disease more accurately. Attached Figure Description

[0031] Figure 1 The medical imaging data is for an embodiment of the present invention.

[0032] Figure 2 This is a test result diagram of YOLOv5 for T2-weighted MRI images of patients and normal individuals, according to an embodiment of the present invention.

[0033] Figure 3 This is a test result diagram of YOLOv5 for DaT images of patients and normal individuals, according to an embodiment of the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] This invention utilizes brain MRI imaging combined with deep learning technology for the diagnosis of Parkinson's disease. It detects and identifies lesions in the substantia nigra and erythrocyte-nigrostriatal regions and outputs a disease risk score. This can help doctors make a more intuitive and efficient assessment of a patient's condition from an imaging perspective, and provides a method for annotating Parkinson's disease lesion regions on MRI images.

[0036] Step A: The medical imaging data reading module uses the file selector to choose a DICOM or DaT file. After acquiring the selected file, the software starts a file processing thread. It then obtains information such as voxels, spatial spacing, window width, and window level from the medical imaging data and visualizes it.

[0037] Step B: Lesion area detection module, using YOLO v5 model to detect lesion areas, outputting disease risk scores and lesion areas, and displaying them visually.

[0038] Step C: Lesion area segmentation module, refines DaT image features, uses U-Net network to segment the striatum region, divides the feature region into the right substantia nigra to more accurately locate the lesion area, and displays it visually.

[0039] The original image input in step A is DICOM format image data, and step A includes the following steps:

[0040] Step A-1: ​​Read in MRI images and DaT data, and preprocess the two modalities.

[0041] Step A-2: Use the FSL-bet tool to perform craniotomy preprocessing on the raw MRI data.

[0042] Step A-3: Start the file processing thread to obtain information such as voxels, spatial spacing, and 3D image size of the medical image data.

[0043] Step A-4: Display the read medical image data.

[0044] The results are as follows Figure 1 .

[0045] Step B includes the following steps:

[0046] Step B-1: Convert MRI images and DaT data into PNG format images. Under the guidance of a professional physician, annotate the red nucleus and substantia nigra regions of the brain MRI images. Divide the annotated data into training, testing, and validation sets.

[0047] Step B-2: Localize the substantia nigra region using a localization model based on the improved YOLOv5 network. The improved YOLOv5 localization model includes an input layer, an improved Backbone module, a Neck module, a YOLOhead module, and a localization module. The input layer refers to the Mosaic data augmentation algorithm. The improved Backbone module includes a Focus layer, two sets of CBL and CSP_1, and one set of CBL and SPP. The Neck module includes an FPN+PAN structure and two sets of CBL+CSP2_X. The YOLOhead module includes two-scale detection heads. The localization module includes non-maximum suppression, IOU thresholding, evaluation of location loss using the DIOU loss function, and evaluation of object category confidence loss for ground truth and predicted bounding boxes using the BCE loss function. The ground truth bounding box refers to the smallest bounding rectangle corresponding to the location of the substantia nigra region in the brain MRI image. The prediction results are as follows: Figure 2 .

[0048] Step B-3: Input the test set into the YOLOv5 network model to locate the basal nucleus and substantia nigra region in MRI images and extract the relative position information of the predicted bounding boxes from the localization results. The predicted bounding box refers to the rectangular border generated by the YOLOv5 network model, used to select the located basal nucleus and substantia nigra region within the MRI image. A classification model based on a one-dimensional convolutional neural network obtains the relative position information of the lymphocyte predicted bounding boxes. After coordinate transformation and region extraction, the basal nucleus and substantia nigra region within the predicted bounding box of the brain MRI image is obtained, locating the substantia nigra region and indicating its health status. In normal individuals, the substantia nigra region is clear and symmetrical, while in patients, the substantia nigra region is darker and less uniform in color. The one-dimensional convolutional neural network model reduces the dimensionality of the obtained average spectral values ​​through one-dimensional convolution, ultimately detecting the substantia nigra region and using the mAP value as an evaluation index for the network's predicted disease score to determine health status.

[0049] Step B-4: Localize the red nucleus and substantia nigra regions in the DaT data using a YOLOv5-based detection network model. In a healthy individual, both regions will light up and display a comma-like shape when observed using SPECT. In early-stage PD individuals, the light intensity from these regions is lower, resulting in very small areas that no longer exhibit comma-like features. This region was localized and trained using a YOLOv5 model. The YOLOv5-based detection network model consists of three parts: a backbone, a neck, and an output module. The backbone includes a BottleneckCSP module and a Focus module. The BottleneckCSP module enhances the overall learning performance of the convolutional neural network. The Focus module slices the image, expanding the input channels to four times their original size, and then performs a convolution to obtain a downsampled feature map. The neck employs a structure combining FPN and PAN, integrating a conventional FPN layer with a bottom-up feature pyramid to fuse extracted semantic and positional features. It also fuses features from the backbone layer and the detection layer, enabling the model to acquire richer feature information. The output module predicts image features and outputs a vector containing the target object's category probability, object score, and the location of its bounding box. It detects feature regions and displays the network's perceived disease score, thus identifying whether the object is healthy.

[0050] The results are as follows Figure 3 .

[0051] Step C includes the following steps:

[0052] Step C-1: Based on DaT data, the model outputs the segmentation and recognition results for the left and right striatum, respectively. Combining the prediction results from DaT data using YOLOv5 and U-Net further improves the reliability and persuasiveness of the model.

[0053] Step C-2: Prediction results of UNET using EfficientNetB4 as the backbone. The segmentation map has 5 different categories: background, healthy left striatum, healthy right striatum, unhealthy left striatum, and unhealthy right striatum. Specific implementation examples:

[0055] Step A: Read the medical imaging data module.

[0056] Select either a DaT file or an MRI file using the file selector, and extract features using different feature extraction methods for the two types of data. After acquiring the selected file, the software starts a file processing thread to obtain information such as voxels, spatial spacing, and 3D image size from the medical image data. The dataset is then divided into training, testing, and validation sets in a 4:1:1 ratio using a randomized method.

[0057] Displays the retrieved medical image data.

[0058] Step B: Red nucleus-substance lesion detection module. In this embodiment, the detection network model based on YOLOv5 consists of three parts: backbone, Neck, and output module. The backbone includes the BottleneckCSP module and the Focus module. The BottleneckCSP module significantly reduces the computational load while enhancing the learning performance of the entire convolutional neural network. The Focus module performs image slicing, expanding the input channels to four times their original size, and then performs a convolution to obtain a downsampled feature map, reducing computational load and improving speed while achieving downsampling. The Neck adopts a structure combining FPN and PAN, combining the conventional FPN layer with a bottom-up feature pyramid to fuse the extracted semantic and positional features. It also fuses the backbone layer and the detection layer to obtain richer feature information. The output module predicts image features and outputs a vector containing the target object's category probability, object score, and the position of the object's bounding box.

[0059] The network is trained using a pre-defined training set based on the YOLOv5 detection network model. The network performance is tested using a test set to determine if the model meets engineering metrics. If it does, training is terminated; otherwise, the network parameters are modified and retrained. These engineering metrics include: accurately identifying the red nucleus and substantia nigra regions in MRI and DaT images; and accurately identifying the diseased lesion area and disease risk score.

[0060] Step C: Refine the DaT image features and use the U-Net network to segment the striatum region. The feature region is divided into the right substantia nigra for more precise localization of the lesion area. A detection and segmentation accuracy of 68.8% was achieved on the validation set, used to determine the presence of Parkinson's disease. The prediction results based on YOLOv5 and U-Net are combined to further improve the model's reliability and persuasiveness. The model automatically interprets images, automatically locates, segments, and detects lesion areas. The accuracy output on the test set can be used to evaluate the model's reliability. The classification probability of [0~1] output during prediction can be used as an indicator to assess the risk score of Parkinson's disease. The prediction results are visualized.

Claims

1. A smart diagnostic analysis method for Parkinson's disease based on MRI images, characterized in that, Includes the following steps: Step 1: Obtain voxel, spatial spacing, window width and window level information of medical image data, and visualize it; Step 1-1: Read MRI image data and DaT data from DICOM format medical images, and extract features from the two types of data using different feature extraction methods; Steps 1-2: Use the FSL-bet tool to perform craniotomy preprocessing on the MRI image data; Steps 1-3: Obtain information on voxels, spatial spacing, and 3D image size of medical images; Steps 1-4: Display the retrieved medical image data; Step 2: Lesion area detection module, using the YOLO v5 model to detect lesion areas, outputting disease risk scores and lesion areas, and displaying them visually; Step 2-1: Convert MRI data and DaT data into PNG format images, annotate the red nucleus and substantia nigra regions of the brain MRI images, and divide the annotated brain MRI image data into training set, test set and validation set; Step 2-2: Use a localization model based on the improved YOLOv5 network to localize the red nucleus and substantia nigra regions in brain MRI images; The localization model based on the improved YOLOv5 network includes an input module, an improved Backbone module, a Neck module, a YOLO head module, and a localization module. The input module uses the Mosaic data augmentation algorithm. The improved Backbone module includes a Focus layer, two sets of CBL and CSP_1, and one set of CBL and SPP. The Neck module includes an FPN+PAN structure and two sets of CBL+CSP2_X. The YOLO head module includes two-scale detection heads. The localization module includes non-maximum suppression, IOU thresholding, evaluation of location loss using the DIOU loss function, and evaluation of object category confidence loss for ground truth and predicted bounding boxes using the BCE loss function. The ground truth bounding box refers to the minimum bounding rectangle corresponding to the red nucleus and substantia nigra region in a brain MRI image. Steps 2-3: Input the test set into the localization model based on the improved YOLOv5 network to complete the localization of the red nucleus and substantia nigra region in the MRI image and extract the relative position information of the predicted box in the localization result; the predicted box refers to the rectangular border generated by the YOLOv5 network model, which is used to select the localized red nucleus and substantia nigra region in the MRI image. A classification model based on a one-dimensional convolutional neural network obtains the relative position information of the prediction box. After coordinate transformation and region extraction, the red nucleus and substantia nigra region within the prediction box of the brain MRI image is obtained, the substantia nigra region is located, and its health status is displayed. The one-dimensional convolutional neural network model reduces the dimensionality of the obtained average spectral value through one-dimensional convolution, and finally detects the substantia nigra region. The mAP value is used as an evaluation index for the network to predict the disease score and determine whether it is healthy. Steps 2-4: Use a YOLOv5-based detection network model to locate the red nucleus and substantia nigra regions in the DaT data; detect the red nucleus and substantia nigra regions and display the disease score to identify whether they are healthy; The YOLOv5-based detection network model comprises three parts: a backbone, a neck, and an output module. The backbone includes a BottleneckCSP module and a Focus module. The BottleneckCSP module enhances the learning performance of the entire convolutional neural network. The Focus module slices the image, expanding the input channels to four times their original size, and then performs a convolution to obtain a downsampled feature map. The neck employs a structure combining FPN and PAN, integrating the FPN layer with a bottom-up feature pyramid to fuse extracted semantic and positional features, while also fusing features between the backbone layer and the detection layer. The output module predicts image features and outputs a vector containing the target object's class probability, object score, and the location of its bounding box. Step 3: Lesion area segmentation module, refines DaT image features, uses U-Net network to segment the striatum region, divides the feature region into the right substantia nigra to locate the lesion area, and displays it visually; Step 3-1: Based on DaT data, the model outputs the segmentation and recognition results of the left and right striatum respectively; the prediction results of DaT data YOLOv5 and U-Net are combined for judgment; Step 3-2: Prediction results of UNET using EfficientNetB4 as the backbone; the segmentation map has 5 different categories: background, healthy left striatum, healthy right striatum, unhealthy left striatum, and unhealthy right striatum.

Images