A disease screening and diagnosis system, model training method

Abstract

The present invention relates to the field of magnetic resonance imaging technology, and particularly to a disease screening and diagnosis system using magnetic resonance data, including: a data acquisition module for acquiring training data and detection data obtained by scanning a subject's body part; a model training module connected to the data acquisition module for inputting the training data into a detection model to train the detection model; and a detection module respectively connected to the data acquisition module and the model training module for inputting the detection data into the trained detection model to output a detection result. The beneficial effects of the present invention are as follows: By using a pre-trained detection model, the k-space data is directly input into the detection model as detection data to output a detection result, without reconstructing the k-space data into an image and then having it judged manually by a doctor. On the one hand, the time for image reconstruction is saved, significantly reducing the time for disease screening and diagnosis, and also saving costs. The accuracy of the detection result is higher.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic resonance imaging, and particularly to a disease screening and diagnosis system using magnetic resonance data. Background Art

[0002] Magnetic Resonance Imaging (MRI) technology is a non-invasive imaging method that can display the structure and properties of human tissues. Clinically, magnetic resonance imaging systems are usually first used to obtain k-space data from the body parts of the scanned person (disease patients or normal subjects), and image reconstruction is performed to obtain images, which are then read by doctors to determine whether there are various abnormalities, including those caused by stroke, tumors or other diseases. However, the process of image reconstruction can be extremely time-consuming and expensive. It takes a long scanning time to form images with good enough image quality (i.e., high enough spatial resolution, high enough contrast, high enough signal-to-noise ratio, and few enough artifacts). The k-space data collected by full sampling is used to reconstruct images and then provided to doctors for diagnosis. Secondly, doctors need additional time to observe the reconstructed images and draw diagnostic conclusions from them.

[0003] Different from cameras used in daily life, in magnetic resonance imaging, the acquired raw data is not image pixels, but k-space data, that is, the two-dimensional or three-dimensional spatial frequency domain representation of the image. If the k-space data is fully sampled, it can be converted into an image through inverse Fourier transform. Some methods in the prior art attempt to reduce the MRI scanning time by only collecting partial k-space data and use appropriate image reconstruction models to eliminate the artifacts brought by only collecting partial k-space data, such as parallel imaging methods (including SENSE and GRAPPA that can be applied to in-plane accelerated acquisition reconstruction or multi-slice simultaneous acquisition) and compressive sensing to restore images. However, when the k-space samples become very sparse, the noise and artifacts introduced by these methods become more and more obvious, and the image quality reconstructed from highly sparse k-space data is poor, making it difficult for doctors to judge diseases or misjudge. In fact, in order to obtain clinically available magnetic resonance images, their maximum acceleration factors are limited between 2 and 8, so the speed of magnetic resonance imaging is still significantly limited. That is to say, the diagnostic time in the prior art is also longer.

[0004] On the other hand, there are also other methods that attempt to use machine learning to detect anomalies in magnetic resonance images, and such methods require less or no intervention from doctors. However, these methods are intended to use high-quality images as input, and these images are similar to the images presented to human doctors. However, this anomaly detection method using machine learning requires the formation of complete and high-quality images, which also requires a long magnetic resonance imaging scanning time. In cases of high noise and obvious artifacts, even if the detection does not completely fail, the detection accuracy will still be significantly reduced. Summary of the Invention

[0005] In view of the above problems existing in the prior art, there is provided a disease screening and diagnosis system using magnetic resonance data. Through a pre-trained detection model, k-space data can be directly input into the detection model as detection data, and the detection result can be directly output by the detection model, without the need to reconstruct the k-space data into an image and then have it judged manually by a doctor.

[0006] The specific technical solution is as follows:

[0007] The present invention includes a disease screening and diagnosis system using magnetic resonance data, comprising:

[0008] A data acquisition module, configured to acquire training data and detection data obtained by scanning a subject's body part;

[0009] A model training module, connected to the data acquisition module, configured to input the training data into a detection model to train the detection model;

[0010] A detection module, respectively connected to the data acquisition module and the model training module, configured to input the detection data into the trained detection model to output a detection result.

[0011] Optionally, the detection module further includes a preprocessing unit, configured to perform normalization and reordering processing on the training data and / or the detection data before inputting the training data and / or the detection data into the detection model.

[0012] Optionally, the sampling trajectory used in the data acquisition module includes a Cartesian trajectory, a radial trajectory, or a spiral trajectory.

[0013] Optionally, the detection data includes k-space data and / or image domain data obtained by reconstructing the k-space data;

[0014] The training data includes the k-space data and / or the image domain data obtained by reconstructing the k-space data.

[0015] Optionally, the training data includes normal k-space data obtained by scanning a normal subject with a magnetic resonance imaging system and abnormal k-space data obtained by scanning a diseased patient.

[0016] Optionally, the training data includes normal k-space data and abnormal k-space data obtained by simulating magnetic resonance images of normal subjects and magnetic resonance images of diseased patients.

[0017] Optionally, the training data includes normal k-space data obtained by simulating magnetic resonance images of normal subjects, and also includes simulated magnetic resonance images of lesions obtained by simulating magnetic resonance images of the normal subjects, and then abnormal k-space data obtained by simulating the simulated magnetic resonance images of the lesions.

[0018] Optionally, the detection model is a classification model, and the detection result output after the classification model inputs the detection data is a classification result; or the detection model is a regression model, and the detection result output after the classification model inputs the detection data is a confidence level.

[0019] Optionally, the k-space data is obtained by two-dimensional acquisition or three-dimensional acquisition; at the same time, the k-space data is acquired by single-channel acquisition or multi-channel acquisition; the k-space data includes data of a single contrast or data of multiple contrasts.

[0020] Optionally, the detection model is divided into multiple sub-detection models according to the anatomical parts of the target detection object, and the detection module divides the detection data according to the anatomical parts, and inputs the detection data of different anatomical parts into the corresponding sub-detection models to output the detection results corresponding to the anatomical parts.

[0021] The present invention also includes a method for training a detection model, including the following steps:

[0022] Collect training data obtained by scanning a subject's body part;

[0023] Perform preprocessing of normalizing and reordering the training data;

[0024] Input the training data into the detection model to train the detection model.

[0025] The technical solution of the present invention has the following advantages or beneficial effects: A disease screening and diagnosis system using magnetic resonance data is provided. Through a pre-trained detection model, the k-space data is directly input into the detection model as detection data, and the detection result is directly output by the detection model, without reconstructing the k-space data into an image and then manually judged by a doctor. On the one hand, the time for image reconstruction is saved, the time for disease screening and diagnosis is greatly reduced, and the cost is also saved; on the other hand, compared with manual diagnosis, the detection result output by the detection model has better stability, reduces the subjectivity of manual diagnosis, and makes the detection result more accurate. Brief Description of the Drawings

[0026] Referring to the accompanying drawings, the embodiments of the present invention will be described more fully. However, the accompanying drawings are only for illustration and explanation, and do not constitute a limitation on the scope of the present invention.

[0027] Figure 1 It is a detection flowchart of the disease screening and diagnosis system in the embodiment of the present invention;

[0028] Figure 2 It is a schematic diagram of model training using the first training data in the embodiment of the present invention;

[0029] Figure 3 It is a schematic diagram of model training using the second training data in the embodiment of the present invention;

[0030] Figure 4 It is a schematic diagram of model training using the third training data in the embodiment of the present invention;

[0031] Figure 5 It is a schematic diagram of the principle of detection using k-space data in the embodiment of the present invention;

[0032] Figure 6 It is a schematic diagram of the principle of detection using k-space data and image domain data in the embodiment of the present invention;

[0033] Figure 7 It is a full-sampled k-space data graph;

[0034] Figure 8 It is a magnetic resonance image reconstructed using full-sampled k-space data;

[0035] Figure 9 It is a highly sparse-sampled k-space data graph;

[0036] Figure 10 It is a magnetic resonance image reconstructed using highly sparse-sampled k-space data. Detailed Embodiments

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other.

[0039] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but it is not a limitation of the present invention.

[0040] The present invention includes a disease screening and diagnosis system using magnetic resonance data, as Figure 1 shown, including:

[0041] A data acquisition module 1 for acquiring training data and detection data obtained by scanning a subject's body part;

[0042] A model training module 2 connected to the data acquisition module 1 for inputting the training data into a detection model to train the detection model;

[0043] A detection module 3 respectively connected to the data acquisition module 1 and the model training module 2 for inputting the detection data into the trained detection model to output a detection result.

[0044] Specifically, the present invention relates to detecting abnormalities of a subject in magnetic resonance imaging. Specifically, it relates to using a machine learning classification model to directly detect abnormalities from magnetic resonance imaging k-space data. This solution uses a magnetic resonance imaging system to scan a subject's body part to obtain highly sparse k-space data. Different from traditional methods, this method does not require the step of reconstructing images, but directly uses the k-space data as the input of the detection model to train the detection model to predict abnormalities in the subject's k-space data. In short, traditional diagnostic methods need to use k-space data for image reconstruction, and then doctors diagnose through the images. The method of the present invention is to train a detection model to let the detection model judge whether there are abnormalities in the k-space data scanned from the subject, and the detection model can output a detection result. For example, if the currently scanned k-space data is from the subject's brain, the detection model judges whether there is a disease in the subject's brain based on the input k-space data, or the detection result of what kind of disease. Through the above technical solutions, a large amount of image reconstruction time and the cost of reconstructing images can be saved, and in addition, the time for doctors to make a diagnosis can also be saved, and the diagnostic accuracy is higher.

[0045] Specifically, Figure 1It represents the design of the entire system, which may include three modules: a data acquisition module 1, a model training module 2, and a detection module 3. In the data acquisition module 1, a magnetic resonance imaging system can be used to scan a subject's body part to obtain sparse k-space data as training data or detection data, and then the k-space data is input into the detection model for training the detection model or for disease detection.

[0046] As an alternative, the detection data includes k-space data and / or image domain data obtained by image reconstruction using the k-space data; the training data includes k-space data and / or image domain data obtained by image reconstruction using the k-space data. The k-space data collected by the data acquisition module 1 is also reconstructed into an image, which may contain a large number of artifacts but can also be used as an input to the detection model. In the model training module 2, the detection model is pre-trained, training data with corresponding labels is prepared and fed back to the detection model, and the model parameters are iteratively updated to minimize the difference between the model output and the labels. Here, the labels are labels indicating whether there is a disease in the training data or what kind of disease exists, for example, the k-space data of a brain tumor patient. The present invention can input a large amount of training data into the detection model, including the k-space data of normal subjects and the k-space data of patients with various diseases, for training, so that the detection model can directly diagnose and output the detection results during actual use.

[0047] As an alternative embodiment, the detection module 3 further includes a preprocessing unit for normalizing and reordering the training data and / or detection data before inputting the training data and / or detection data into the detection model.

[0048] As an alternative embodiment, the sampling trajectory for the data acquisition module 1 to collect k-space data may include a Cartesian trajectory, a radial trajectory, a spiral trajectory, or other non-Cartesian undersampling trajectories.

[0049] As an alternative embodiment, as Figure 2 shown, the training data includes normal k-space data obtained by scanning a normal subject with a magnetic resonance imaging system and abnormal k-space data obtained by scanning a disease patient. Figure 2 Illustrates the process of training the model in some embodiments, where the k-space data is obtained by scanning normal subjects and disease patients with an actual magnetic resonance imaging system, and then undergoes preprocessing including normalization and reordering, and is used as an input to the detection model.

[0050] As an alternative embodiment, as Figure 3As shown, the training data includes the magnetic resonance images of normal subjects and the normal k-space data and abnormal k-space data obtained by simulating the magnetic resonance images of disease patients. The k-space data for training can also be obtained by simulation. First, the magnetic resonance images from normal subjects and disease patients are acquired, and then the corresponding k-space data is obtained through simulation. The other steps are the same as Figure 2 the description of.

[0051] As an alternative implementation, as Figure 4 shown, magnetic resonance images with lesions can also be obtained by simulating the magnetic resonance images of normal subjects. The other steps are similar to Figure 3 . The training data includes the normal k-space data obtained by simulating the magnetic resonance images of normal subjects, and also includes the magnetic resonance images with lesions obtained by simulating the magnetic resonance images of normal subjects. Then, the abnormal k-space data obtained by simulating the magnetic resonance images with lesions is included.

[0052] As an alternative implementation, Figure 5 illustrates the operation of detecting abnormalities from k-space data using a detection model. The highly sparse k-space data is preprocessed (normalized, reordered) and input into the detection model. The detection model then outputs detection results, such as confidence levels or classification results, etc. In this invention, a fully connected multi-layer perceptron is taken as an example. Assuming that the scale of single-shot spiral trajectory data is 256, and the real and imaginary parts of the data are used as two independent input channels (i.e., the input size is 2×256), then a multi-layer perceptron with the parameters shown in Table 1 can be used for abnormality detection:

[0053] Table 1. Example of a fully connected multi-layer perceptron

[0054]

[0055] As an alternative implementation, as Figure 6 shown, in addition to the highly sparse k-space data, the reconstructed image can also be used as one of the inputs to the detection model. As an Figure 5 effective supplement, it should be noted that the step of reconstructing the image is optional, and in this invention, images of relatively low quality can be input for detection. For example, Figure 10Images reconstructed from such highly sparse k-space data can save certain time and cost compared to images reconstructed from fully sampled k-space data. Although doctors cannot visually judge diseases directly from such images, the detection model can directly diagnose the results. Further, the image reconstruction in this embodiment can be implemented by non-uniform fast Fourier transform (NUFFT), parallel imaging image reconstruction, or compressive sensing reconstruction. Here, a convolutional neural network is taken as an example for illustration. Assuming that the size of the complex image after image reconstruction is 128×128, and the real part and the imaginary part of the complex image are used as two independent input channels (i.e., the input size is 2×128×128), then the convolutional neural network shown in Table 2 can be used for anomaly detection:

[0056] Table 2. Example of Convolutional Neural Network

[0057]

[0058] Figures 9 - 10 is a schematic result of anomaly detection from highly sparse sampled k-space data. This figure shows that the present invention can detect brain tumors directly from highly sparse sampled k-space data. Among them Figure 7 is the original fully sampled k-space data, Figure 8 is the image reconstructed from the fully sampled k-space data, which contains a simulated tumor with a diameter of 5-10 mm and is reflected as a high-brightness signal in the image. Figure 9 is the highly sparse sampled k-space data obtained from two k-space lines, Figure 10 is the image directly reconstructed from the above-mentioned highly sparse sampled k-space data. This image is too blurred to be provided to doctors as a basis for directly diagnosing diseases, but it can be diagnosed by the detection model trained by the present invention and the diagnostic results can be output.

[0059] As an optional implementation manner, the detection model is a classification model, and the detection result output after the classification model inputs the detection data is a classification result; or the detection model is a regression model, and the detection result output after the classification model inputs the detection data is a confidence level.

[0060] The detection model can be a Support Vector Machine (SVM); it can be a Decision Tree (DT), such as a Gradient Boosting Decision Tree (GBDT) and a Random Forest; it can also be various types of Artificial Neural Networks (ANN), such as a Multilayer Perceptron (MLP), a Convolutional Neural Network (CNN), and a Recurrent Neural Network (RNN), or a combination of the above models.

[0061] As an alternative implementation, the k-space data is obtained by two-dimensional acquisition or three-dimensional acquisition; at the same time, the k-space data is acquired through single-channel or multi-channel acquisition; the k-space data includes data of a single contrast or data of multiple contrasts. It should be noted that, on the one hand, during magnetic resonance imaging, there may be multiple receiving channels (such as when using a phased array coil), so the above k-space data can be single-channel or multi-channel. On the other hand, conventional clinical magnetic resonance imaging usually uses multiple sequences to obtain images with different contrasts, so as to obtain complementary diagnostic information; therefore, the k-space data can also contain data of multiple contrasts. If it is necessary to support data with multiple contrasts, it is only necessary to input the data of different contrasts as data of different channels when inputting into the detection model.

[0062] As an alternative implementation, both the k-space data and the image domain data are complex numbers. When inputting into the detection model, either the real part and the imaginary part of the data can be input as different channels respectively, or a detection model that supports complex numbers (such as a Complex-valued CNN) can be used to utilize the complex characteristics of the data.

[0063] As an alternative implementation, the detection model is divided into multiple sub-detection models according to the anatomical parts of the target detection object, and the detection module divides the detection data according to the anatomical parts and inputs the detection data of different anatomical parts into the corresponding sub-detection models to output the detection results corresponding to the anatomical parts.

[0064] Specifically, in this embodiment, after obtaining the training data, it is divided according to anatomical parts, and the training data of a specific anatomical part is used to train a sub-detection model dedicated to this anatomical part; while scanning the subject, the anatomical part is identified according to the collected detection data, and a suitable sub-detection model is selected; in each anatomical part, the corresponding sub-detection model is used, the magnetic resonance imaging data is input to perform the detection, and the results of each anatomical part are combined. When using the above method, at least one positioning scan or navigation signal needs to be obtained during the actual scan to identify the anatomical part currently being scanned. The magnetic resonance data obtained by scanning can be fully sampled k-space data or highly sparse k-space data obtained by downsampling. According to the different anatomical parts, the corresponding detection data is input for detection. On the one hand, the detection angle can be improved, and on the other hand, the detection time can be further shortened.

[0065] The embodiment of the present invention also provides a training method for a detection model. As Figures 2 - 4 shown, the disease screening and diagnosis system in the above example executes this training method to obtain the detection model, including the following steps:

[0066] Step S1, collect the training data obtained by scanning the subject's body part;

[0067] Step S2, perform preprocessing of normalization and reordering on the training data;

[0068] Step S3, input the training data into the detection model to train the detection model.

[0069] Specifically, the training data in this embodiment is diversified. As Figure 2 shown, the training data can be normal k-space data obtained by scanning normal subjects and abnormal k-space data obtained by scanning disease patients; it can also be Figure 3 shown, the normal k-space data and abnormal k-space data obtained by simulating the magnetic resonance images of normal subjects and disease patients; it can also be Figure 4 shown, the normal k-space data obtained by simulating the magnetic resonance images of normal subjects, and also includes the simulated lesion magnetic resonance images obtained by simulating the magnetic resonance images of normal subjects, and then the abnormal k-space data obtained by simulating the lesion magnetic resonance images.

[0070] The embodiment of the present invention has the following beneficial effects:

[0071] (1) Since only a very small amount of k-space data is required and a large amount of k-space data is not needed for image reconstruction, the MRI scanning time and related costs are reduced, and the reconstruction time and related costs for converting k-space data into images are also saved;

[0072] (2) The detection result is directly output by the detection model, saving the doctor's time for observing the image and related costs;

[0073] (3) The detection model of the present invention can achieve high detection accuracy at a high acquisition acceleration factor or a very short scanning time.

[0074] The above are only the preferred embodiments of the present invention, and do not limit the implementation manners and protection scope of the present invention. For those skilled in the art, it should be realized that all equivalent replacements and obvious changes made by using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A disease screening and diagnosis system using magnetic resonance data, characterized in that, Including: A data acquisition module, configured to acquire training data and detection data obtained by scanning a body part of a subject. A model training module, connected to the data acquisition module, configured to input the training data into a detection model to train the detection model. A detection module, respectively connected to the data acquisition module and the model training module, configured to input the detection data into the trained detection model to output a detection result. The detection data includes k-space data. The training data includes the k-space data.

2. The disease screening and diagnosis system according to claim 1, wherein The detection module further includes a preprocessing unit, configured to perform normalization and reordering processing on the training data and / or the detection data before inputting the training data and / or the detection data into the detection model.

3. The disease screening and diagnosis system according to claim 1, wherein The sampling trajectory used in the data acquisition module includes a Cartesian trajectory, a radial trajectory, or a spiral trajectory.

4. The disease screening and diagnosis system according to claim 1, characterized in that, The training data includes normal k-space data obtained by scanning a normal subject with a magnetic resonance imaging system and abnormal k-space data obtained by scanning a disease patient.

5. The disease screening and diagnosis system according to claim 1, characterized in that, The training data includes normal k-space data and abnormal k-space data obtained by simulating magnetic resonance images of a normal subject and a disease patient.

6. The disease screening and diagnosis system according to claim 1, wherein The training data includes normal k-space data obtained by simulating a magnetic resonance image of a normal subject, and further includes a lesion magnetic resonance image obtained by simulating the magnetic resonance image of the normal subject, and then abnormal k-space data obtained by simulating the lesion magnetic resonance image.

7. The disease screening and diagnosis system according to claim 1, characterized in that, The detection model is a classification model, and the detection result output after inputting the detection data into the classification model is a classification result; or the detection model is a regression model, and the detection result output after inputting the detection data into the classification model is a confidence level.

8. The disease screening and diagnosis system according to any one of claims 4-6, characterized in that, The k-space data is obtained by two-dimensional acquisition or three-dimensional acquisition; at the same time, the k-space data is acquired by single-channel acquisition or multi-channel acquisition; the k-space data includes data of a single contrast or data of multiple contrasts.

9. The disease screening and diagnosis system according to claim 1, characterized in that, The detection model is divided into multiple sub-detection models according to the anatomical parts of the target detection object, and the detection module divides the detection data according to the anatomical parts, and inputs the detection data of different anatomical parts into the corresponding sub-detection models to output the detection result corresponding to the anatomical part.

10. A training method for a detection model, characterized in that, Applied to the disease screening and diagnosis system according to any one of claims 1-9, the disease screening and diagnosis system executes the training method to obtain the detection model, including the following steps: Acquire training data obtained by scanning a body part of a subject. Perform preprocessing of normalization and reordering on the training data. Input the training data into a detection model to train the detection model.

Images