WO2019200745A1 - Mri病变位置检测方法、装置、计算机设备和存储介质 - Google Patents

Mri病变位置检测方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2019200745A1
WO2019200745A1 PCT/CN2018/095493 CN2018095493W WO2019200745A1 WO 2019200745 A1 WO2019200745 A1 WO 2019200745A1 CN 2018095493 W CN2018095493 W CN 2018095493W WO 2019200745 A1 WO2019200745 A1 WO 2019200745A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
training
mri
sample
neural network
Prior art date
Application number
PCT/CN2018/095493
Other languages
English (en)
French (fr)
Inventor
王健宗
吴天博
刘新卉
刘莉红
马进
肖京
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2019200745A1 publication Critical patent/WO2019200745A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10088—Magnetic resonance imaging [MRI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30081—Prostate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30096—Tumor; Lesion

Definitions

  • the present application relates to the field of computer technology, and in particular, to a MRI lesion position detecting method, apparatus, computer device and storage medium.
  • Prostate cancer is the second most common cancer among men worldwide, with risk factors including advanced age, family history and ethnicity. About 99% of patients with prostate cancer are over 50 years old, and parents and other relatives are two to three times more likely to have a child's risk than ordinary people.
  • prostate cancer is usually not the direct cause of death, it is important to detect the site of cancer damage.
  • Existing traditional detection methods include Prostate Specific Antigen (PSA, prostate specific antigen) detection and Digital Rectal Examination (DRE, digital rectal examination) detection. These methods have a low detection accuracy and are at risk of over-detection (in hospitals receiving tests that are not related to their condition).
  • Prostate Imaging Reporting And Data System PRADS, prostate imaging report and data system
  • PRADS Prostate Imaging Reporting And Data System
  • prostate imaging report and data system refers to the use of a structural reporting system for comprehensive diagnosis of the prostate, the accuracy of which depends largely on The professional level of radiologists and medical experts is therefore very limited; at the same time, it takes a lot of manpower and material resources.
  • the main purpose of the present application is to provide a method, a device, a computer device and a storage medium for detecting a position of an MRI lesion, which overcome the defects of low detection accuracy and high detection cost in the prior art.
  • the present application provides a method for detecting an MRI lesion position, comprising the following steps:
  • the training samples are input into a convolutional neural network for parameter training, and the training parameters of the convolutional neural network are calculated through training, and the convolutional neural network training the training parameters is used as a detection model for detecting the lesion position in the MRI data.
  • the training sample is MRI data of a known lesion location;
  • the sample to be detected is MRI data of an unknown lesion position.
  • the application also provides an MRI lesion position detecting device, comprising:
  • a training unit configured to input training samples into a convolutional neural network for parameter training, calculate training parameters of the convolutional neural network through training, and use a convolutional neural network that trains the training parameters as detecting MRI data a lesion location detection model; the training sample is MRI data of a known lesion location;
  • a detecting unit configured to receive a sample to be detected, input the sample to be detected into the detection model for prediction, and predict a lesion position in the sample to be detected; the sample to be detected is an MRI data of an unknown lesion position .
  • the application further provides a computer device comprising a memory and a processor, wherein the memory stores computer readable instructions, and the processor implements the steps of the method when the computer readable instructions are executed.
  • the present application also provides a computer non-transitory readable storage medium having stored thereon computer readable instructions that, when executed by a processor, implement the steps of the above methods.
  • the MRI lesion position detecting method, device, computer equipment and storage medium provided in the present application input training samples into a convolutional neural network for parameter training, and calculate training parameters of the convolutional neural network by training, and training
  • the convolutional neural network of the training parameter is used as a detection model for detecting a lesion position in the MRI data; the sample to be detected is input into the detection model for prediction, and the position of the lesion in the sample to be detected is predicted;
  • the MRI data automatically detects the location of the lesion in the patient's MRI data.
  • the automatic detection replaces the subjective diagnosis of the expert, makes full use of the existing data, saves a lot of human and material costs, and improves the accuracy of predicting the location of the lesion from the MRI data. , has a high practical value.
  • FIG. 1 is a schematic diagram showing the steps of a method for detecting an MRI lesion position in an embodiment of the present application
  • step S1a is a schematic diagram of specific steps of step S1a in an embodiment of the present application.
  • FIG. 3 is a block diagram showing the structure of an MRI lesion position detecting device according to an embodiment of the present application.
  • FIG. 4 is a block diagram showing the structure of an MRI lesion position detecting device in another embodiment of the present application.
  • FIG. 5 is a structural block diagram of a preprocessing unit in an embodiment of the present application.
  • FIG. 6 is a schematic block diagram showing the structure of a computer device according to an embodiment of the present application.
  • the MRI lesion position detecting method provided in the embodiment of the present application is suitable for detecting a lesion position in any MRI data.
  • the following describes the lesion position in the MRI data of the prostate site as an example for detecting other parts. The specific process of lesion location in MRI data will not be described again.
  • an embodiment of the present application provides a method for detecting an MRI lesion position, including the following steps:
  • step S1 the training samples are input into the convolutional neural network for parameter training, and the training parameters of the convolutional neural network are calculated through training, and the convolutional neural network training the training parameters is used as the detection of the lesion position in the MRI data.
  • the model is tested; the training sample is MRI data of a known location of a prostate cancer lesion, which is typically an MRI picture.
  • the above convolutional neural network (Convolutional Neural Networks, CNN) is a deep learning network that includes a convolutional layer and a pooling layer that provides sound for image processing.
  • the above training sample is MRI (Magnetic Resonance Imaging) data of a position in which a prostate cancer lesion is known in advance, wherein the MRI data can be obtained by nuclear magnetic resonance.
  • the source of the MRI data of the above known prostate cancer lesion location may be selected from the data of the prostate cancer lesion that has been detected in the historical detection database, or the training sample obtained in advance, or may be the prostate for the prostate cancer patient by the expert.
  • MRI data collected at the site may be selected from the data of the prostate cancer lesion that has been detected in the historical detection database, or the training sample obtained in advance, or may be the prostate for the prostate cancer patient by the expert.
  • the original MRI data is obtained by performing nuclear magnetic resonance on the prostate site of the prostate cancer patient, and the original MRI data includes four different MRI sequence data, and the four kinds of the sequence data are diffusion-weighted imaging data respectively.
  • DWI diffusion Weighted Images
  • ADC Apparent Diffusion Coefficient
  • Ktrans dynamic enhanced quantitative parameter data
  • T2 Weighted Images, T2WI weighted image data
  • the training samples are input into a convolutional neural network for parameter training, and MRI data of known prostate cancer lesion locations are trained, and the position of the lesion in the MRI data is different from the training result of the lesion position. Therefore, the training results of different positions in the MRI data are different, and since the training result is known (ie, the known lesion position), the training parameters of the convolutional neural network can be deduced based on the training result, and then The trained training parameters are input into the convolutional neural network, and the lesion position detection model in the detected MRI data is obtained.
  • the detection model is used for fully automatic detection of MRI data of the prostate site of the patient, and detecting prostate cancer lesions in the MRI data. position. It can replace manual detection, reduce manpower and material cost; and the detection speed is obviously improved, the detection efficiency is improved, the detection accuracy is improved, and even exceeds the expert level.
  • Step S2 receiving a sample to be detected, inputting the sample to be detected into the detection model for prediction, and predicting a position of a prostate cancer lesion in the sample to be detected; the sample to be detected is an MRI data of an unknown lesion position .
  • the detection model is the detection model obtained by the training completed in the above step S1.
  • the MRI of the prostate site can be collected by the medical device. Data, the MRI data is taken as a sample to be detected and sent through the medical device, and when the terminal of the detection model receives the sample to be detected, it is input into the detection model for prediction, and the last layer of the detection model is outputted.
  • the position of the prostate cancer lesion in the sample is detected, that is, whether the patient has prostate cancer can be predicted.
  • the whole detection process instead of manually detecting the location of prostate cancer lesions in MRI data, not only the detection cost is reduced, but also the detection accuracy is high, and the detection speed and detection efficiency are also obviously improved.
  • the method before the step S1 of inputting the training sample into the convolutional neural network for parameter training and calculating the training parameter of the convolutional neural network by training, the method includes:
  • Step S1a pre-processing the original MRI data to obtain the training sample;
  • the original MRI data is prostate tissue MRI data of a prostate cancer patient.
  • the training samples in the above embodiments may be pre-made, and in this embodiment, an implementation process of making original MRI data into training samples is proposed.
  • the original MRI data refers to MRI data obtained by magnetic resonance imaging of the prostate site of a prostate cancer patient by a medical device, and the original MRI data is only marked with a lesion position, but the data usually has heterogeneous characteristics or contains some Abnormal data, abnormal data includes blurred pictures, broken pictures, etc., so the original MRI data needs to be preprocessed to obtain training samples suitable for training convolutional neural networks.
  • the foregoing step S1a of pre-processing the original MRI data to obtain the training sample includes:
  • Step S101 performing correct alignment on the four sequence data included in the original MRI data, and eliminating heterogeneity between the four types of sequence data; the four types of sequence data are diffusion-weighted imaging data and apparent diffusion, respectively. Coefficient data, dynamic enhanced quantitative parameter data, and weighted image data.
  • the four sequence data contained in the original MRI data are usually not aligned and heterogeneous.
  • the present embodiment is based on the mutuality proposed by Chappelow et al. (2011).
  • the alignment method of information corrects alignment of the four sequence data.
  • the step of removing the abnormal data is required before the step S101, and the step of removing the abnormal data is a conventional method, and details are not described herein. .
  • the region is found on the diffusion-weighted imaging data using the region growing method and the morphological operation, and the lesion is damaged.
  • the center of the area is defined as the center of damage. This step is beneficial to better distinguish between the lesion position and the non-lesion position during the training.
  • Step S102 randomly selecting three different data into the RGB three-channel image in the diffusion-weighted imaging data, the apparent diffusion coefficient data, the dynamic enhanced quantitative parameter data, and the weighted image data to obtain the training sample;
  • the RGB three-channel image is a 3D image.
  • the above four sequence data are randomly combined into an RGB three-channel image, and the combined RGB three-channel image is used as a training sample.
  • the above diffusion-weighted imaging data, apparent diffusion coefficient data, dynamic enhanced quantitative parameter data, and weighted image data are respectively represented by D, A, K, and T; and the above four kinds of sequence data are randomly combined into RGB three.
  • the channel image can be represented as DAK, DAT, AKT, DKT, and the above RGB three-channel image is used as a training sample of the input convolutional neural network.
  • the images expressed by the combination of different sequence data are different, so that the training samples input into the convolutional neural network are diverse, and the training samples using the diversity training are more comprehensive and more beneficial to the training model.
  • the MRI data to be detected may be pre-processed in the above-mentioned step S1a before the detection of the sample to be detected.
  • step S1a For the specific implementation, reference may be made to the above steps S101 and S102, and details are not described herein.
  • the step of pre-processing the original MRI data to obtain the training sample, after the step S1a includes:
  • Step S1b performing data augmentation processing on the training samples to increase the data amount of the training samples.
  • the data volume of the MRI data collected is usually small, and the MRI data in the sample to be detected is inherently diverse.
  • the convolutional neural network is trained using a large number of training samples to obtain a detection model. Therefore, under the premise of reducing the cost, it is necessary to perform data augmentation processing on the above training samples, and the data augmentation processing is an incremental processing of the data amount, expanding the data amount of the training samples, and using a sufficient amount of training samples for training.
  • Facilitating the diversity of MRI data is conducive to improving the accuracy of subsequent detection models for detecting prostate cancer.
  • the step S1b of performing data augmentation processing on the training sample to increase the data amount of the training sample includes:
  • Planar rotation, clipping, and normalization are performed on each of the slices, and each of the slices is used as a training sample.
  • the RGB three-channel image in step a is the image combined in the above step S102.
  • slices are performed from 7 different directions for the RGB three-channel image. Slicing from multiple different directions can dramatically increase the amount of data.
  • Each of the slices is then subjected to planar rotation, shearing, and normalization as described in step b. Since the slices are in different directions, they need to be planarly rotated such that all slices are on the same plane.
  • the normalization process is to transform the lesion position to positive and negative pixels, and the normalization process is used to speed up the training convergence speed, which is a commonly used data processing method in deep learning.
  • each of the slices is used as a training sample, and each slice corresponds to a two-dimensional Region of Interests (ROIs) picture data.
  • ROIs Region of Interests
  • the convolutional neural network performs parameter training using the Adam method
  • the loss function used by the convolutional neural network is a cross entropy loss function.
  • the cross entropy loss function is a way to measure the predicted and actual values of a convolutional neural network (CNN). Compared with the quadratic cost function, it can promote the training of CNN more effectively.
  • the Adam method dynamically adjusts the learning rate for each parameter based on the first-order moment estimate and the second-order moment estimate of the gradient of each parameter according to the loss function. The learning rate will gradually decrease after Loss (loss function) is no longer reduced. The reason that Loss does not decrease is that the learning rate is too large, so the way to reduce the learning rate is to make the loss continue to decrease.
  • the optimization method can also use SGD (random steepest descent method), Momentum (momentum optimization) and other methods, but after experiment comparison, it is found that using Adam method is the best.
  • SGD random steepest descent method
  • Momentum mimentum optimization
  • Algorithm greyedy baggingalgorithm
  • the training samples are input into a convolutional neural network for parameter training, the training parameters of the convolutional neural network are calculated through training, and the convolutional neural network training the training parameters is used as the detection MRI.
  • the method includes:
  • step S1c a test sample is input into the detection model for verification to verify the training parameter; the test sample is MRI data of a known prostate cancer lesion location.
  • the training sample and the test sample are set, and the ratio of the training sample to the test sample can be set to 3:1; in the deep learning, the ratio of the training sample to the test sample needs to be set reasonably, so that the training sample is trained.
  • the model is reasonably tested on the test sample to select the optimal test model.
  • the test samples are consistent with the data in the above training samples and are MRI data of known prostate cancer lesion locations.
  • the test sample is input into the above-mentioned trained detection model for training, the detection model outputs the prediction result, and the prediction result is compared with the known result of the test sample to determine the training parameter. is it right or not.
  • the AUC of the test model trained in this embodiment (the index for judging the classification effect) is higher than the traditional PIRADS method.
  • the detection model in this embodiment it is possible to automatically detect whether the prostate cancer site is contained in the MRI data of the prostate site of the patient.
  • the corresponding patient has prostate cancer.
  • body information age, weight, living habits, medical history, etc.
  • similar patient cases are matched in the historical detection database, and the risk factors of the patients are easily analyzed according to the commonality.
  • a treatment plan designed for patients with similar body information is called from the database and pushed to the doctor to assist the doctor in treating the patient.
  • the body information of the patient collected in advance is stored in the historical detection database, and the body information of a large number of patients in the historical detection data is analyzed by big data, and the weight of the patient's causative factors is analyzed. For example, to count the proportion of patients in a certain age group in a large number of patients, it is possible to roughly analyze the risk factor of the patient's disease as a weight of the age, and similarly calculate the weight of each risk factor.
  • the training sample is input into the convolutional neural network for parameter training, and the training parameters of the convolutional neural network are calculated through training, and the training parameters are trained.
  • the convolutional neural network of the training parameter is used as a detection model for detecting a lesion position in the MRI data; the sample to be detected is input into the detection model for prediction, and the position of the prostate cancer lesion in the sample to be detected is predicted;
  • the MRI data to be detected automatically detects the position of the prostate cancer lesion in the patient's MRI data, and the automatic detection replaces the expert subjective diagnosis, makes full use of the existing data, saves a lot of human and material costs, and improves the prediction from the MRI data.
  • the accuracy of prostate cancer lesion location has a high practical value.
  • the MRI lesion position detecting device provided in the embodiment of the present application is suitable for detecting the position of a lesion in any MRI data.
  • the following describes the position of the lesion in the MRI data of the prostate site as an example.
  • an embodiment of the present application further provides an MRI lesion position detecting apparatus, including:
  • the training unit 10 is configured to input the training samples into the convolutional neural network for parameter training, calculate the training parameters of the convolutional neural network by training, and use the convolutional neural network that trains the training parameters as the detection MRI data.
  • a lesion location detection model the training sample is MRI data of a known prostate cancer lesion location, and the MRI data is usually an MRI image.
  • the Convolutional Neural Networks is a deep learning network, including a convolutional layer and a pooling layer, which have an acoustic performance for image processing.
  • the above training sample is MRI (Magnetic Resonance Imaging) data of a position in which a prostate cancer lesion is known in advance, wherein the MRI data can be obtained by nuclear magnetic resonance.
  • the source of the MRI data of the above known prostate cancer lesion location may be selected from the data of the prostate cancer lesion that has been detected in the historical detection database, or the training sample obtained in advance, or may be the prostate for the prostate cancer patient by the expert.
  • MRI data collected at the site is MRI (Magnetic Resonance Imaging) data of a position in which a prostate cancer lesion is known in advance, wherein the MRI data can be obtained by nuclear magnetic resonance.
  • the source of the MRI data of the above known prostate cancer lesion location may be selected from the data of the prostate cancer lesion that has been detected in the historical detection database, or the training sample obtained in advance
  • the original MRI data is obtained by performing nuclear magnetic resonance on the prostate site of the prostate cancer patient, and the original MRI data includes four different MRI sequence data, and the four kinds of the sequence data are diffusion-weighted imaging data respectively.
  • DWI diffusion Weighted Images
  • ADC Apparent Diffusion Coefficient
  • Ktrans dynamic enhanced quantitative parameter data
  • T2 Weighted Images, T2WI weighted image data
  • the training unit 10 inputs the above training samples into the convolutional neural network for parameter training, and trains the MRI data of the known prostate cancer lesion position, and the position of the lesion in the MRI data and the position where the lesion does not occur.
  • the training results are different. Therefore, the training results of different positions in the MRI data are different, and since the training result is known (ie, the known lesion position), the training of the convolutional neural network can be reversed according to the training result.
  • the location of prostate cancer lesions It can replace manual detection, reduce manpower and material cost; and the detection speed is obviously improved, the detection efficiency is improved, the detection accuracy is improved, and even exceeds the expert level.
  • the detecting unit 20 is configured to receive a sample to be detected, input the sample to be detected into the detection model for prediction, and predict a position of a prostate cancer lesion in the sample to be detected; the sample to be detected is an unknown lesion position MRI data.
  • the detection model is a detection model obtained by training the training unit 10, and at this time, if a new patient needs to detect whether there is prostate cancer, the MRI of the prostate site can be collected by the medical device. Data, the MRI data is taken as a sample to be detected and sent through the medical device.
  • the detecting unit 20 receives the sample to be detected, it is input into the detection model for prediction, and the last layer of the detection model is output to be detected.
  • the location of the prostate cancer lesion in the sample can predict whether the patient has prostate cancer. In the whole detection process, instead of manually detecting the location of prostate cancer lesions in MRI data, not only the detection cost is reduced, but also the detection accuracy is high, and the detection speed and detection efficiency are also obviously improved.
  • the MRI lesion position detecting apparatus further includes:
  • the pre-processing unit 30 is configured to pre-process the original MRI data to obtain the training sample; the original MRI data is the prostate site MRI data of the prostate cancer patient.
  • the training samples in the above embodiments may be pre-made, and in this embodiment, an implementation process of making original MRI data into training samples is proposed.
  • the original MRI data refers to MRI data obtained by magnetic resonance imaging of the prostate site of a prostate cancer patient by a medical device, and the original MRI data is only marked with a lesion position, but the data usually has heterogeneous characteristics or contains some Abnormal data, the abnormal data includes a blurred picture, a broken picture, etc., so the original MRI data needs to be preprocessed by the pre-processing unit 30 to obtain a training sample suitable for training the convolutional neural network.
  • the pre-processing unit 30 includes:
  • the alignment module 301 is configured to perform orthogonal alignment on the four sequence data included in the original MRI data, and eliminate heterogeneity between the four types of sequence data; the four types of sequence data are diffusion-weighted imaging data, respectively. Apparent diffusion coefficient data, dynamic enhanced quantitative parameter data, and weighted image data.
  • the four sequence data included in the original MRI data are usually not aligned and have heterogeneity.
  • the alignment module 301 is proposed by Chappelow et al. (2011).
  • the four sequence data is corrected and aligned based on the mutual information alignment method.
  • the step of removing the abnormal data is required before the alignment module 301 is aligned, and the step of removing the abnormal data is a conventional means, and Repeat them.
  • the region is found on the diffusion-weighted imaging data using the region growing method and the morphological operation, and the lesion is damaged.
  • the center of the area is defined as the center of damage. This step is beneficial to better distinguish between the lesion position and the non-lesion position during the training.
  • the combining module 302 is configured to randomly select three different data into the RGB three-channel image in the diffusion-weighted imaging data, the apparent diffusion coefficient data, the dynamic enhanced quantitative parameter data, and the weighted image data, to obtain the training sample.
  • the RGB three-channel image is a 3D image.
  • the combining module 302 randomly combines the above four sequence data into an RGB three-channel image, and uses the combined RGB three-channel image as a training. sample.
  • the above diffusion-weighted imaging data, apparent diffusion coefficient data, dynamic enhanced quantitative parameter data, and weighted image data are respectively represented by D, A, K, and T; and the above four kinds of sequence data are randomly combined into RGB three.
  • the channel image can be represented as DAK, DAT, AKT, DKT, and the above RGB three-channel image is used as a training sample of the input convolutional neural network.
  • the images expressed by the combination of different sequence data are different, so that the training samples input into the convolutional neural network are diverse, and the training samples using the diversity training are more comprehensive and more beneficial to the training model.
  • the MRI data to be detected may be pre-processed as described above.
  • the MRI lesion position detecting device further includes:
  • the augmenting unit is configured to perform data augmentation processing on the training sample to increase the data amount of the training sample.
  • the data volume of the MRI data collected is usually small, and the MRI data in the sample to be detected is inherently diverse.
  • the convolutional neural network is trained using a large number of training samples to obtain a detection model. Therefore, under the premise of reducing the cost, it is necessary to perform data augmentation processing on the above training samples, and the data augmentation processing is an incremental processing of the data amount, expanding the data amount of the training samples, and using a sufficient amount of training samples for training.
  • Facilitating the diversity of MRI data is conducive to improving the accuracy of subsequent detection models for detecting prostate cancer.
  • the augmenting unit comprises:
  • a slicing module for slicing from a plurality of different directions for the RGB three-channel image
  • a normalization module is used for planar rotation, clipping, and normalization of each of the slices, and each of the slices is used as a training sample.
  • the RGB three-channel image is an image obtained by combining the combination modules 302.
  • the slicing module slices from the 7 different directions for the RGB three-channel image. Slicing from multiple different directions can dramatically increase the amount of data.
  • Each of the slices is then subjected to planar rotation, shearing, and normalization as described in step b. Since the slices are in different directions, they need to be planarly rotated such that all slices are on the same plane.
  • the normalization process is to transform the lesion position to positive and negative pixels, and the normalization process is used to speed up the training convergence speed, which is a commonly used data processing method in deep learning.
  • each of the slices is used as a training sample, and each slice corresponds to a two-dimensional Region of Interests (ROIs) picture data.
  • ROIs Region of Interests
  • the convolutional neural network performs parameter training using the Adam method
  • the loss function used by the convolutional neural network is a cross entropy loss function.
  • the cross entropy loss function is a way to measure the predicted and actual values of a convolutional neural network (CNN). Compared with the quadratic cost function, it can promote the training of CNN more effectively.
  • the Adam method dynamically adjusts the learning rate for each parameter based on the first-order moment estimate and the second-order moment estimate of the gradient of each parameter according to the loss function. The learning rate will gradually decrease after Loss (loss function) is no longer reduced. The reason that Loss does not decrease is that the learning rate is too large, so the way to reduce the learning rate is to make the loss continue to decrease.
  • the optimization method can also use SGD (random steepest descent method), Momentum (momentum optimization) and other methods, but after experiment comparison, it is found that using Adam method is the best.
  • SGD random steepest descent method
  • Momentum mimentum optimization
  • Algorithm greyedy baggingalgorithm
  • the MRI lesion position detecting device further includes:
  • a verification unit for inputting a test sample into the detection model for verification to verify the training parameter;
  • the test sample is MRI data of a known prostate cancer lesion location.
  • the training sample and the test sample are set, and the ratio of the training sample to the test sample can be set to 3:1; in the deep learning, the ratio of the training sample to the test sample needs to be set reasonably, so that the training sample is trained.
  • the model is reasonably tested on the test sample to select the optimal test model.
  • the test samples are consistent with the data in the above training samples and are MRI data of known prostate cancer lesion locations.
  • the verification unit inputs the test sample into the above-mentioned trained detection model for training, detects the model output prediction result, compares the prediction result with the known result of the test sample, and determines Whether the above training parameters are correct.
  • the AUC of the test model trained in this embodiment (the index for judging the classification effect) is higher than the traditional PIRADS method.
  • the detection model in this embodiment it is possible to automatically detect whether the prostate cancer site is contained in the MRI data of the prostate site of the patient.
  • the detecting unit 20 after the detecting unit 20 detects the position of the prostate cancer lesion in the sample to be detected, it can be determined that the corresponding patient has prostate cancer. According to the body information (age, weight, living habits, medical history, etc.) of the pre-collected patients, similar patient cases are matched in the historical detection database, and the risk factors of the patients are easily analyzed according to the commonality. Or a treatment plan designed for patients with similar body information is called from the database and pushed to the doctor to assist the doctor in treating the patient.
  • the detecting unit 20 detects the position of the prostate cancer lesion in the sample to be detected, it is determined that the corresponding patient has prostate cancer.
  • the body information of the patient collected in advance is stored in the historical detection database, and the body information of a large number of patients in the historical detection data is analyzed by big data, and the weight of the patient's causative factors is analyzed. For example, to count the proportion of patients in a certain age group in a large number of patients, it is possible to roughly analyze the risk factor of the patient's disease as a weight of the age, and similarly calculate the weight of each risk factor.
  • the MRI lesion position detecting apparatus inputs the training sample into the convolutional neural network for parameter training, and calculates the training parameters of the convolutional neural network by training, and training is performed.
  • the convolutional neural network of the training parameter is used as a MRI lesion position detection model in detecting MRI data; the sample to be detected is input into a detection model for prediction, and the position of the prostate cancer lesion in the sample to be detected is predicted;
  • the MRI data of the part to be detected automatically detects the position of the prostate cancer lesion in the patient's MRI data, and the automatic detection replaces the expert subjective diagnosis, makes full use of the existing data, saves a lot of human and material costs, and improves the MRI data. Predicting the accuracy of prostate cancer lesion location has a high practical value.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 6.
  • the computer device includes a processor, memory, network interface, and database connected by a system bus. Among them, the computer designed processor is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium, an internal memory.
  • the non-volatile storage medium stores an operating system, computer readable instructions, and a database.
  • the internal memory provides an environment for operation of an operating system and computer readable instructions in a non-volatile storage medium.
  • the database of the computer device is used to store data such as a convolutional neural network.
  • the network interface of the computer device is used to communicate with an external terminal via a network connection.
  • the computer readable instructions are executed by the processor to implement an MRI lesion location detection method.
  • the processor performs the steps of the MRI lesion position detecting method: inputting a training sample into a convolutional neural network for parameter training, and calculating a training parameter of the convolutional neural network by training, and training the volume of the training parameter
  • the neural network is used as a lesion position detection model for detecting MRI data;
  • the training sample is MRI data of a known prostate cancer lesion location;
  • the sample to be detected is MRI data of an unknown lesion position.
  • the processor inputs the training samples into the convolutional neural network for parameter training, and before the step of calculating the training parameters of the convolutional neural network by training, the method includes:
  • the raw MRI data is pre-processed to obtain the training sample; the original MRI data is the prostate site MRI data of the prostate cancer patient.
  • the processor performs pre-processing on the original MRI data to obtain the training sample, and includes:
  • the four kinds of sequence data are diffusion weighted imaging data, apparent diffusion coefficient data, Dynamically enhancing quantitative parameter data and weighted image data;
  • the RGB three-channel image is a 3D image.
  • the method includes:
  • Data augmentation processing is performed on the training samples to increase the amount of data of the training samples.
  • the step of the processor performing data augmentation processing on the training sample to increase the data amount of the training sample includes:
  • Planar rotation, clipping, and normalization are performed on each of the slices, and each of the slices is used as a training sample.
  • the convolutional neural network performs parametric training using the Adam method.
  • the processor inputs training samples into a convolutional neural network for parameter training, calculates training parameters of the convolutional neural network by training, and constructs a convolutional neural network that trains the training parameters.
  • the method includes:
  • test sample is input into the test model for verification to verify the training parameter; the test sample is MRI data of a known prostate cancer lesion location.
  • FIG. 6 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation of the computer device to which the solution of the present application is applied.
  • An embodiment of the present application further provides a computer non-volatile readable storage medium having computer readable instructions stored thereon.
  • an MRI lesion position detecting method is implemented, specifically:
  • the training samples are input into a convolutional neural network for parameter training, and the training parameters of the convolutional neural network are calculated through training, and the convolutional neural network training the training parameters is used as a detection model for detecting lesions in the MRI data;
  • the training sample is an MRI data of a known location of a prostate cancer lesion;
  • the sample to be detected is MRI data of an unknown lesion position.
  • the processor inputs the training samples into the convolutional neural network for parameter training, and before the step of calculating the training parameters of the convolutional neural network by training, the method includes:
  • the raw MRI data is pre-processed to obtain the training sample; the original MRI data is the prostate site MRI data of the prostate cancer patient.
  • the processor performs pre-processing on the original MRI data to obtain the training sample, and includes:
  • the four kinds of sequence data are diffusion weighted imaging data, apparent diffusion coefficient data, Dynamically enhancing quantitative parameter data and weighted image data;
  • the RGB three-channel image is a 3D image.
  • the processor after the step of pre-processing the original MRI data to obtain the training sample, the processor includes:
  • Data augmentation processing is performed on the training samples to increase the amount of data of the training samples.
  • the step of the processor performing data augmentation processing on the training sample to increase the data amount of the training sample includes:
  • Planar rotation, clipping, and normalization are performed on each of the slices, and each of the slices is used as a training sample.
  • the convolutional neural network performs parametric training using the Adam method.
  • the processor inputs training samples into a convolutional neural network for parameter training, calculates training parameters of the convolutional neural network by training, and constructs a convolutional neural network that trains the training parameters.
  • the method includes:
  • test sample is input into the test model for verification to verify the training parameter; the test sample is MRI data of a known prostate cancer lesion location.
  • the MRI lesion position detecting method, apparatus, computer device and storage medium input training samples into a convolutional neural network for parameter training, and calculate the convolutional nerve through training.
  • a training parameter of the network the convolutional neural network training the training parameter is used as a lesion position detection model in the MRI data;
  • the sample to be detected is input into the detection model for prediction, and the prostate cancer in the sample to be detected is predicted
  • the MRI data of the patient's prostate is automatically detected by the MRI data of the patient's prostate, and the automatic detection of the patient's MRI data replaces the subjective diagnosis of the expert, making full use of the existing data, saving a lot of human and material costs. It also improves the accuracy of predicting the location of prostate cancer lesions from MRI data, and has a high practical value.
  • Non-volatile memory can include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM is available in a variety of formats, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronization.
  • SRAM static RAM
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • SSRSDRAM dual-speed SDRAM
  • ESDRAM enhanced SDRAM
  • SLDRAM Link (Synchlink) DRAM
  • SLDRAM Memory Bus
  • RDRAM Direct RAM
  • DRAM Direct Memory Bus Dynamic RAM
  • RDRAM Memory Bus Dynamic RAM

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)
  • Image Processing (AREA)

Abstract

一种MRI病变位置检测方法，包括：将训练样本输入至卷积神经网络中进行参数训练，得到检测MRI数据中病变位置的检测模型(S1)；将待检测样本输入至检测模型中进行预测，预测出所述待检测样本中的病变位置(S2)。该方法提高了从MRI数据中预测病变位置的准确率。还公开了MRI病变位置检测装置以及计算机设备和存储介质。

Description

MRI病变位置检测方法、装置、计算机设备和存储介质
本申请要求于2018年4月20日提交中国专利局、申请号为2018103614272，发明名称为“MRI病变位置检测方法、装置、计算机设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，特别涉及一种MRI病变位置检测方法、装置、计算机设备和存储介质。
 背景技术
前列腺癌是世界范围内男士群体患病率第二高的癌症，其风险因子包括高龄、家族病史和种族等。约99%的前列腺癌病例中的患者年龄超过50岁，而父母等亲属患病时，其子发病风险较常人超出两至三倍。
尽管前列腺癌通常不是导致死亡的直接原因，但是检测癌症损伤部位非常重要。现有传统检测手段包括Prostate Specific Antigen（PSA，前列腺特异抗原）检测和Digital RectalExamination（DRE，直肠指检）检测。这些方法的检测准确率较低，而且会有过度检测（在医院接受与自身病情无关的检查）的风险。
另外，传统方法还有Prostate Imaging Reporting And Data System（PIRADS，前列腺影像报告和数据系统），指的是用一个结构性的报告体系对前列腺进行全方位的诊断，其准确性很大程度上取决于放射专家和医疗专家的专业水平，因此有很大的局限性；同时，耗费大量的人力物力。
 技术问题
本申请的主要目的为提供一种MRI病变位置检测方法、装置、计算机设备和存储介质，克服现有技术中检测准确率低，检测成本高的缺陷。
 技术解决方案
为实现上述目的，本申请提供了一种MRI病变位置检测方法，包括以下步骤：
将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置的检测模型；所述训练样本为已知病变位置的MRI数据；
接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的病变位置；所述待检测样本为未知病变位置的MRI数据。
本申请还提供了一种MRI病变位置检测装置，包括：
训练单元，用于将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型；所述训练样本为已知病变位置的MRI数据；
检测单元，用于接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的病变位置；所述待检测样本为未知病变位置的MRI数据。
本申请还提供一种计算机设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现上述方法的步骤。
本申请还提供一种计算机非易失性可读存储介质，其上存储有计算机可读指令，所述计算机可读指令被处理器执行时实现上述方法的步骤。
 有益效果
本申请中提供的MRI病变位置检测方法、装置、计算机设备和存储介质，将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置的检测模型；将待检测样本输入至检测模型中进行预测，预测出所述待检测样本中的病变位置；通过对病人的待检测MRI数据进行自动检测该病人MRI数据中的病变位置，全自动检测替代了专家主观诊断，充分利用已有数据，节约了大量的人力物力成本，并提高了从MRI数据中预测病变位置的准确率，具有很高的实际应用价值。
 附图说明
图1 是本申请一实施例中MRI病变位置检测方法步骤示意图；
图2 是本申请一实施例中的步骤S1a的具体步骤示意图；
图3 是本申请一实施例中MRI病变位置检测装置结构框图；
图4 是本申请另一实施例中MRI病变位置检测装置结构框图；
图5 是本申请一实施例中的预处理单元结构框图；
图6 为本申请一实施例的计算机设备的结构示意框图。
 本发明的最佳实施方式
本申请实施例中提供MRI病变位置检测方法适用于检测任意MRI数据中的病变位置，为了便于阐述，下文中均以检测前列腺部位的MRI数据中的病变位置为例进行阐述，对于检测其它部位的MRI数据中的病变位置的具体过程不再进行赘述。
参照图1，本申请实施例中提供了一种MRI病变位置检测方法，包括以下步骤：
步骤S1，将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型；所述训练样本为已知前列腺癌病变位置的MRI数据，该MRI数据通常为MRI图片。
在本步骤S1中，上述卷积神经网络（Convolutional Neural Networks，CNN）是一种深度学习网络，包括卷积层以及池化层，对图像处理具有出声表现。上述训练样本为预先已知前列腺癌病变位置的MRI（Magnetic Resonance Imaging，核磁共振图像）数据，其中，MRI数据通过核磁共振可以获得。上述已知前列腺癌病变位置的MRI数据的来源可以是从历史检测数据库中已经检测出前列腺癌病变的数据中选取的，或者是预先得到的训练样本，也可以是通过专家针对前列腺癌患者的前列腺部位采集的MRI数据。例如，在具体一个实施例中，对前列腺癌患者的前列腺部位进行核磁共振获取原始MRI数据，原始MRI数据中包括有四种不同的MRI序列数据，四种所述序列数据分别为扩散加权成像数据（Diffusion Weighted Images，DWI）、表观扩散系数数据（Apparent Diffusion Coefficient，ADC）、动态增强定量参数数据（Ktrans）以及加权图像数据（T2 Weighted Images，T2WI）。通过对患者的前列腺部分进行核磁共振可以获取到上述四种MRI序列数据，由专家/专业医生对上述数据中的前列腺癌病变位置进行标注，如此，则获取到已知前列腺癌病变位置的MRI数据，将其作为训练样本。
在本步骤中，将上述训练样本输入至卷积神经网络中进行参数训练，对已知前列腺癌病变位置的MRI数据进行训练，MRI数据中发生病变的位置与未发生病变位置的训练结果是不同的，因此，MRI数据中不同位置的训练结果不同，而由于训练结果为已知的（即已知的病变位置），则可以根据训练结果反推出所述卷积神经网络的训练参数，然后将训练出的训练参数输入至卷积神经网络中，则得到检测MRI数据中病变位置检测模型，该检测模型用于对患者前列腺部位的MRI数据进行全自动检测，检测出MRI数据中的前列腺癌病变位置。其可以替代人工检测，降低人力、物力成本；且检测速度明显提升，检测效率得到提高，检测的准确率提高，甚至超过专家水平。
步骤S2，接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的前列腺癌病变位置；所述待检测样本为未知病变位置的MRI数据。
在本步骤S2中，所述检测模型即为上述步骤S1中训练完成得到的检测模型，此时，若有新的患者需要检测是否患有前列腺癌，则可以通过医疗设备采集其前列腺部位的MRI数据，将该MRI数据作为待检测样本并通过医疗设备发送过来，检测模型的终端接收到该待检测样本时，则将其输入至检测模型中进行预测，该检测模型的最后一层则输出待检测样本中的前列腺癌病变位置，即可以预测出该患者是否患有前列腺癌。整个检测过程中，替代人工进行检测MRI数据中的前列腺癌病变位置，不仅降低检测成本，而且其检测准确率高，检测速度、检测效率也明显得到提高。
在一实施例中，上述将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数的步骤S1之前，包括：
步骤S1a，对原始MRI数据进行预处理，得到所述训练样本；所述原始MRI数据为前列腺癌患者的前列腺部位MRI数据。
上述实施例中的训练样本可以是预先制作好的，而本实施例中提出一种将原始的MRI数据制作成训练样本的实现过程。原始MRI数据指的是通过医疗设备对前列腺癌患者的前列腺部位进行核磁共振得到的MRI数据，该原始MRI数据中仅标注有病变位置，然而该数据中通常会具有异质性的特点或者含有一些异常数据，异常数据包括模糊图片、残缺图片等，因此需要对原始MRI数据进行预处理，以得到适合训练卷积神经网络的训练样本。
具体地，参照图2，上述对原始MRI数据进行预处理，得到所述训练样本的步骤S1a，包括：
步骤S101，将所述原始MRI数据中包含的四种序列数据进行矫正对齐，消除四种所述序列数据之间的异质性；四种所述序列数据分别为扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据。
在本步骤中，原始MRI数据中包含的四种序列数据通常没有对齐，具有异质性，为了消除上述异质性，本实施例中按照Chappelow et al.(2011) 提出的基于mutual information（互信息）的对齐方法对四种序列数据进行矫正对齐。在其它实施例中，由于原始MRI数据中具有一些模糊图片、残缺图片等，因此在该步骤S101之前，需要进行去除异常数据的步骤，该去除异常数据的步骤为常规手段，在此不进行赘述。
在一实施例中，为了精细化上述原始MRI数据的损伤中心（病变位置），本实施例中使用region growing法（区域增长法）和形态学操作在扩散加权成像数据上找到损伤区域，将损伤区域的圆心定为损伤中心，此步骤有益于训练过程中更好的区分出病变位置与非病变位置。
步骤S102，在所述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据中随机选出三种不同的数据组合成RGB三通道图像，得到所述训练样本；所述RGB三通道图像为3D图像。
在本步骤中，为了使得输入至卷积神经网络中的训练样本具有多样性，将上述四种序列数据随机组合成RGB三通道图像，将该组合后的RGB三通道图像作为训练样本。具体地，为了便于阐述，上述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据分别以D、A、K以及T表示；上述四种序列数据随机组合成的RGB三通道图像则可以表示为DAK、DAT、AKT、DKT，将上述RGB三通道图像作为输入卷积神经网络的训练样本。不同序列数据组合表达的图像不同，使得输入至卷积神经网络中的训练样本具有多样性，使用多样性的训练样本训练时更加全面，更加有利于训练模型。
应当理解的是，在对待检测样本进行检测之前，也可以如上述步骤S1a中对待检测MRI数据进行预处理，其具体实现可参照上述步骤S101、步骤S102，在此不再进行赘述。
在一实施例中，上述对原始MRI数据进行预处理，得到所述训练样本的步骤S1a之后，包括：
步骤S1b，对所述训练样本进行数据增广处理，以增加所述训练样本的数据量。
考虑到采集MRI数据的人力成本、物力成本等，通常采集的MRI数据的数据量较小，而待检测样本中的MRI数据本身就具有多样性，为了提高检测模型检测时的准确性，则需要使用大量的训练样本对卷积神经网络进行训练以得到检测模型。因此，在降低成本的前提下，需要对上述训练样本进行数据增广处理，数据增广处理即是对数据量的增量处理，扩大训练样本的数据量，使用足够量的训练样本进行训练，便于应对MRI数据的多样性，有利于提高后续检测模型检测前列腺癌的准确性。
具体的一个实施例中，上述对所述训练样本进行数据增广处理，以增加训练样本的数据量的步骤S1b，包括：
a、针对所述RGB三通道图像，从多个不同方向进行切片；
b、对每个所述切片进行平面旋转、剪切以及归一化处理，并将每个所述切片作为一个训练样本。
在本实施例中，步骤a中的RGB三通道图像，即为上述步骤S102中组合而成的图像。具体实施例中，为了对训练样本进行数据增广处理，针对该RGB三通道图像，从7个不同方向进行切片。从多个不同方向进行切片，则可以使得数据量急剧增加。然后如步骤b所述对每个所述切片进行平面旋转、剪切以及归一化处理，由于上述切片位于不同方向上，因此，需要对其进行平面旋转，使得所有切片位于同一平面上。归一化处理的过程为转化病变位置到像素正负1，归一化处理用于加快训练收敛速度，为深度学习中常用的数据处理方法。经过上述归一化处理之后，将每个所述切片作为一个训练样本，每一个切片对应一个二维的感兴趣部位（Region of Interests，ROIs）图片数据。
在上述实施例中，上述卷积神经网络使用Adam方法进行参数训练，所述卷积神经网络使用的损失函数为交叉熵损失函数。交叉熵损失函数是用来衡量卷积神经网络（CNN）的预测值与实际值的一种方式。与二次代价函数相比，它能更有效地促进CNN的训练。Adam方法是根据损失函数对每个参数的梯度的一阶矩估计和二阶矩估计动态调整针对于每个参数的学习速率。学习速率在Loss（损失函数）不再减小之后会逐渐减小，Loss不减小的原因是learning rate（学习速率）过大，所以常用减小学习速率的方式是使loss继续减小，因此使用Adam方法进行优化。优化方法也可以用SGD（随机最速下降法）、Momentum（动量优化）等方法，但是经过实验对比发现使用Adam方法效果最好。在一实施例中，训练模型时，使用多种不同参数的卷积神经网络进行训练，在验证时对这些卷积神经网络用加权平均来确定最终的训练参数，加权平均的权重由贪婪装袋算法（greedy baggingalgorithm）来确定。
在一实施例中，上述将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型的步骤S1之后，包括：
步骤S1c，将测试样本输入至所述检测模型中进行验证，验证所述训练参数；所述测试样本为已知前列腺癌病变位置的MRI数据。
在本实施例中，设置有训练样本以及测试样本，训练样本与测试样本比例可设置为3:1；在深度学习中，需要合理的设置训练样本与测试样本的比例，以使训练样本训练的模型在测试样本上得到合理的测试，进而选出最优的检测模型。测试样本与上述训练样本中的数据一致，均为已知前列腺癌病变位置的MRI数据。为了验证上述检测模型的有效性，以及准确性，将测试样本输入至上述训练出的检测模型中进行训练，检测模型输出预测结果，对比该预测结果与测试样本的已知结果，判断上述训练参数是否正确。经过测试样本的验证，本实施例中训练的检测模型的AUC（评判分类效果优劣的指标）高于传统的PIRADS法。使用本实施例中的检测模型可以实现全自动检测病人前列腺部位处MRI数据中是否包含有前列腺癌变位置。
在一实施例中，经上述检测模型检测出待检测样本中的前列腺癌病变位置之后，则判断其对应的患者患有前列腺癌。根据预先收集好的患者的身体信息（年龄、体重、生活习性、病史等），在历史检测数据库中匹配相似的患者案例，根据共通性，便于分析出患者患病的风险因子。或者从数据库中调用针对类似身体信息的患者而设计的调理方案，推送给医生，以便辅助医生对患者治疗。
在另一实施例中，经上述检测模型检测出待检测样本中的前列腺癌病变位置之后，则判断其对应的患者患有前列腺癌。将预先收集好的该患者的身体信息存入历史检测数据库中，对历史检测数据中的大量患者的身体信息进行大数据分析，分析，分析出患者的致病因子所占权重。例如，统计大量患者中在某一个年龄段的患者占比，则可以大致分析出患者患病的风险因子为年龄的一个权重，同理，计算出各个患病风险因子的权重。
综上所述，为本申请实施例中提供的MRI病变位置检测方法，将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型；将待检测样本输入至检测模型中进行预测，预测出所述待检测样本中的前列腺癌病变位置；通过对病人前列腺部位的待检测MRI数据进行自动检测该病人MRI数据中的前列腺癌病变位置，全自动检测替代了专家主观诊断，充分利用已有数据，节约了大量的人力物力成本，并提高了从MRI数据中预测前列腺癌病变位置的准确率，具有很高的实际应用价值。
本申请实施例中提供MRI病变位置检测装置适用于检测任意MRI数据中的病变位置，为了便于阐述，下文中均以检测前列腺部位的MRI数据中的病变位置为例进行阐述。
参照图3，本申请实施例中还提供了一种MRI病变位置检测装置，包括：
训练单元10，用于将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型；所述训练样本为已知前列腺癌病变位置的MRI数据，该MRI数据通常为MRI图片。
在本实施例中，上述卷积神经网络（Convolutional Neural Networks，CNN）是一种深度学习网络，包括卷积层以及池化层，对图像处理具有出声表现。上述训练样本为预先已知前列腺癌病变位置的MRI（Magnetic Resonance Imaging，核磁共振图像）数据，其中，MRI数据通过核磁共振可以获得。上述已知前列腺癌病变位置的MRI数据的来源可以是从历史检测数据库中已经检测出前列腺癌病变的数据中选取的，或者是预先得到的训练样本，也可以是通过专家针对前列腺癌患者的前列腺部位采集的MRI数据。例如，在具体一个实施例中，对前列腺癌患者的前列腺部位进行核磁共振获取原始MRI数据，原始MRI数据中包括有四种不同的MRI序列数据，四种所述序列数据分别为扩散加权成像数据（Diffusion Weighted Images，DWI）、表观扩散系数数据（Apparent Diffusion Coefficient，ADC）、动态增强定量参数数据（Ktrans）以及加权图像数据（T2 Weighted Images，T2WI）。通过对患者的前列腺部分进行核磁共振可以获取到上述四种MRI序列数据，由专家/专业医生对上述数据中的前列腺癌病变位置进行标注，如此，则获取到已知前列腺癌病变位置的MRI数据，将其作为训练样本。
在本实施例中，训练单元10将上述训练样本输入至卷积神经网络中进行参数训练，对已知前列腺癌病变位置的MRI数据进行训练，MRI数据中发生病变的位置与未发生病变位置的训练结果是不同的，因此，MRI数据中不同位置的训练结果不同，而由于训练结果为已知的（即已知的病变位置），则可以根据训练结果反推出所述卷积神经网络的训练参数，然后将训练出的训练参数输入至卷积神经网络中，则得到检测MRI数据中病变位置检测模型，该检测模型用于对患者前列腺部位的MRI数据进行全自动检测，检测出MRI数据中的前列腺癌病变位置。其可以替代人工检测，降低人力、物力成本；且检测速度明显提升，检测效率得到提高，检测的准确率提高，甚至超过专家水平。
检测单元20，用于接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的前列腺癌病变位置；所述待检测样本为未知病变位置的MRI数据。
在本实施例中，所述检测模型即为上述训练单元10训练完成得到的检测模型，此时，若有新的患者需要检测是否患有前列腺癌，则可以通过医疗设备采集其前列腺部位的MRI数据，将该MRI数据作为待检测样本并通过医疗设备发送过来，检测单元20接收到该待检测样本时，则将其输入至检测模型中进行预测，该检测模型的最后一层则输出待检测样本中的前列腺癌病变位置，即可以预测出该患者是否患有前列腺癌。整个检测过程中，替代人工进行检测MRI数据中的前列腺癌病变位置，不仅降低检测成本，而且其检测准确率高，检测速度、检测效率也明显得到提高。
参照图4，在一实施例中，上述MRI病变位置检测装置还包括：
预处理单元30，用于对原始MRI数据进行预处理，得到所述训练样本；所述原始MRI数据为前列腺癌患者的前列腺部位MRI数据。
上述实施例中的训练样本可以是预先制作好的，而本实施例中提出一种将原始的MRI数据制作成训练样本的实现过程。原始MRI数据指的是通过医疗设备对前列腺癌患者的前列腺部位进行核磁共振得到的MRI数据，该原始MRI数据中仅标注有病变位置，然而该数据中通常会具有异质性的特点或者含有一些异常数据，异常数据包括模糊图片、残缺图片等，因此需要通过预处理单元30对原始MRI数据进行预处理，以得到适合训练卷积神经网络的训练样本。
具体地，参照图5，在一实施例中，所述预处理单元30包括：
对齐模块301，用于将所述原始MRI数据中包含的四种序列数据进行矫正对齐，消除四种所述序列数据之间的异质性；四种所述序列数据分别为扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据。
在本实施例中，原始MRI数据中包含的四种序列数据通常没有对齐，具有异质性，为了消除上述异质性，本实施例中，对齐模块301按照Chappelow et al.(2011) 提出的基于mutual information（互信息）的对齐方法对四种序列数据进行矫正对齐。在其它实施例中，由于原始MRI数据中具有一些模糊图片、残缺图片等，因此在经对齐模块301对齐之前，需要进行去除异常数据的步骤，该去除异常数据的步骤为常规手段，在此不进行赘述。
在一实施例中，为了精细化上述原始MRI数据的损伤中心（病变位置），本实施例中使用region growing法（区域增长法）和形态学操作在扩散加权成像数据上找到损伤区域，将损伤区域的圆心定为损伤中心，此步骤有益于训练过程中更好的区分出病变位置与非病变位置。
组合模块302，用于在所述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据中随机选出三种不同的数据组合成RGB三通道图像，得到所述训练样本；所述RGB三通道图像为3D图像。
在本实施例中，为了使得输入至卷积神经网络中的训练样本具有多样性，组合模块302将上述四种序列数据随机组合成RGB三通道图像，将该组合后的RGB三通道图像作为训练样本。具体地，为了便于阐述，上述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据分别以D、A、K以及T表示；上述四种序列数据随机组合成的RGB三通道图像则可以表示为DAK、DAT、AKT、DKT，将上述RGB三通道图像作为输入卷积神经网络的训练样本。不同序列数据组合表达的图像不同，使得输入至卷积神经网络中的训练样本具有多样性，使用多样性的训练样本训练时更加全面，更加有利于训练模型。
应当理解的是，在对待检测样本进行检测之前，也可以如上述预处理单元中对待检测MRI数据进行预处理，其具体实现可参照上述对齐模块301、组合模块302，在此不再进行赘述。
在一实施例中，上述MRI病变位置检测装置还包括：
增广单元，用于对所述训练样本进行数据增广处理，以增加所述训练样本的数据量。
考虑到采集MRI数据的人力成本、物力成本等，通常采集的MRI数据的数据量较小，而待检测样本中的MRI数据本身就具有多样性，为了提高检测模型检测时的准确性，则需要使用大量的训练样本对卷积神经网络进行训练以得到检测模型。因此，在降低成本的前提下，需要对上述训练样本进行数据增广处理，数据增广处理即是对数据量的增量处理，扩大训练样本的数据量，使用足够量的训练样本进行训练，便于应对MRI数据的多样性，有利于提高后续检测模型检测前列腺癌的准确性。
在一具体实施例中，所述增广单元包括：
切片模块，用于针对所述RGB三通道图像，从多个不同方向进行切片；
归一化模块，用于对每个所述切片进行平面旋转、剪切以及归一化处理，并将每个所述切片作为一个训练样本。
在本实施例中，上述RGB三通道图像，即为上述组合模块302组合而成的图像。具体实施例中，为了对训练样本进行数据增广处理，切片模块针对该RGB三通道图像，从7个不同方向进行切片。从多个不同方向进行切片，则可以使得数据量急剧增加。然后如步骤b所述对每个所述切片进行平面旋转、剪切以及归一化处理，由于上述切片位于不同方向上，因此，需要对其进行平面旋转，使得所有切片位于同一平面上。归一化处理的过程为转化病变位置到像素正负1，归一化处理用于加快训练收敛速度，为深度学习中常用的数据处理方法。经过上述归一化处理之后，将每个所述切片作为一个训练样本，每一个切片对应一个二维的感兴趣部位（Region of Interests，ROIs）图片数据。
在上述实施例中，所述卷积神经网络使用Adam方法进行参数训练，所述卷积神经网络使用的损失函数为交叉熵损失函数。交叉熵损失函数是用来衡量卷积神经网络（CNN）的预测值与实际值的一种方式。与二次代价函数相比，它能更有效地促进CNN的训练。Adam方法是根据损失函数对每个参数的梯度的一阶矩估计和二阶矩估计动态调整针对于每个参数的学习速率。学习速率在Loss（损失函数）不再减小之后会逐渐减小，Loss不减小的原因是learning rate（学习速率）过大，所以常用减小学习速率的方式是使loss继续减小，因此使用Adam方法进行优化。优化方法也可以用SGD（随机最速下降法）、Momentum（动量优化）等方法，但是经过实验对比发现使用Adam方法效果最好。在一实施例中，训练模型时，使用多种不同参数的卷积神经网络进行训练，在验证时对这些卷积神经网络用加权平均来确定最终的训练参数，加权平均的权重由贪婪装袋算法（greedy baggingalgorithm）来确定。
在上述实施例中，上述MRI病变位置检测装置还包括：
验证单元，用于将测试样本输入至所述检测模型中进行验证，验证所述训练参数；所述测试样本为已知前列腺癌病变位置的MRI数据。
在本实施例中，设置有训练样本以及测试样本，训练样本与测试样本比例可设置为3:1；在深度学习中，需要合理的设置训练样本与测试样本的比例，以使训练样本训练的模型在测试样本上得到合理的测试，进而选出最优的检测模型。测试样本与上述训练样本中的数据一致，均为已知前列腺癌病变位置的MRI数据。为了验证上述检测模型的有效性，以及准确性，上述验证单元将测试样本输入至上述训练出的检测模型中进行训练，检测模型输出预测结果，对比该预测结果与测试样本的已知结果，判断上述训练参数是否正确。经过测试样本的验证，本实施例中训练的检测模型的AUC（评判分类效果优劣的指标）高于传统的PIRADS法。使用本实施例中的检测模型可以实现全自动检测病人前列腺部位处MRI数据中是否包含有前列腺癌变位置。
在一实施例中，经上述检测单元20检测出待检测样本中的前列腺癌病变位置之后，则可以判断其对应的患者患有前列腺癌。根据预先收集好的患者的身体信息（年龄、体重、生活习性、病史等），在历史检测数据库中匹配相似的患者案例，根据共通性，便于分析出患者患病的风险因子。或者从数据库中调用针对类似身体信息的患者而设计的调理方案，推送给医生，以便辅助医生对患者治疗。
在另一实施例中，经上述检测单元20检测出待检测样本中的前列腺癌病变位置之后，则判断其对应的患者患有前列腺癌。将预先收集好的该患者的身体信息存入历史检测数据库中，对历史检测数据中的大量患者的身体信息进行大数据分析，分析，分析出患者的致病因子所占权重。例如，统计大量患者中在某一个年龄段的患者占比，则可以大致分析出患者患病的风险因子为年龄的一个权重，同理，计算出各个患病风险因子的权重。
综上所述，为本申请实施例中提供的MRI病变位置检测装置，将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中MRI病变位置检测模型；将待检测样本输入至检测模型中进行预测，预测出所述待检测样本中的前列腺癌病变位置；通过对病人前列腺部位的待检测MRI数据进行自动检测该病人MRI数据中的前列腺癌病变位置，全自动检测替代了专家主观诊断，充分利用已有数据，节约了大量的人力物力成本，并提高了从MRI数据中预测前列腺癌病变位置的准确率，具有很高的实际应用价值。
参照图6，本申请实施例中还提供一种计算机设备，该计算机设备可以是服务器，其内部结构可以如图6所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设计的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储卷积神经网络等数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种MRI病变位置检测方法。
上述处理器执行上述MRI病变位置检测方法的步骤：将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型；所述训练样本为已知前列腺癌病变位置的MRI数据；
接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的前列腺癌病变位置；所述待检测样本为未知病变位置的MRI数据。
在一实施例中，所述处理器将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数的步骤之前，包括：
对原始MRI数据进行预处理，得到所述训练样本；所述原始MRI数据为前列腺癌患者的前列腺部位MRI数据。
在一实施例中，所述处理器对原始MRI数据进行预处理，得到所述训练样本的步骤，包括：
将所述原始MRI数据中包含的四种序列数据进行矫正对齐，消除四种所述序列数据之间的异质性；四种所述序列数据分别为扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据；
在所述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据中随机选出三种不同的数据组合成RGB三通道图像，得到所述训练样本；所述RGB三通道图像为3D图像。
在一实施例中，所述处理器对原始MRI数据进行预处理，得到所述训练样本的步骤之后，包括：
对所述训练样本进行数据增广处理，以增加所述训练样本的数据量。
在一实施例中，所述处理器对所述训练样本进行数据增广处理，以增加所述训练样本的数据量的步骤，包括：
针对所述RGB三通道图像，从多个不同方向进行切片；
对每个所述切片进行平面旋转、剪切以及归一化处理，并将每个所述切片作为一个训练样本。
在一实施例中，所述卷积神经网络使用Adam方法进行参数训练。
在一实施例中，所述处理器将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型的步骤之后，包括：
将测试样本输入至所述检测模型中进行验证，验证所述训练参数；所述测试样本为已知前列腺癌病变位置的MRI数据。
本领域技术人员可以理解，图6中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定。
本申请一实施例还提供一种计算机非易失性可读存储介质，其上存储有计算机可读指令，计算机可读指令被处理器执行时实现一种MRI病变位置检测方法，具体为：将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型；所述训练样本为已知前列腺癌病变位置的MRI数据；
接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的前列腺癌病变位置；所述待检测样本为未知病变位置的MRI数据。
在一实施例中，所述处理器将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数的步骤之前，包括：
对原始MRI数据进行预处理，得到所述训练样本；所述原始MRI数据为前列腺癌患者的前列腺部位MRI数据。
在一实施例中，所述处理器对原始MRI数据进行预处理，得到所述训练样本的步骤，包括：
将所述原始MRI数据中包含的四种序列数据进行矫正对齐，消除四种所述序列数据之间的异质性；四种所述序列数据分别为扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据；
在所述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据中随机选出三种不同的数据组合成RGB三通道图像，得到所述训练样本；所述RGB三通道图像为3D图像。
在一实施例中，所述处理器在对原始MRI数据进行预处理，得到所述训练样本的步骤之后，包括：
对所述训练样本进行数据增广处理，以增加所述训练样本的数据量。
在一实施例中，所述处理器对所述训练样本进行数据增广处理，以增加所述训练样本的数据量的步骤，包括：
针对所述RGB三通道图像，从多个不同方向进行切片；
对每个所述切片进行平面旋转、剪切以及归一化处理，并将每个所述切片作为一个训练样本。
在一实施例中，所述卷积神经网络使用Adam方法进行参数训练。
在一实施例中，所述处理器将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型的步骤之后，包括：
将测试样本输入至所述检测模型中进行验证，验证所述训练参数；所述测试样本为已知前列腺癌病变位置的MRI数据。
综上所述，为本申请实施例中提供的MRI病变位置检测方法、装置、计算机设备和存储介质，将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型；将待检测样本输入至检测模型中进行预测，预测出所述待检测样本中的前列腺癌病变位置；通过对病人前列腺部位的待检测MRI数据进行自动检测该病人MRI数据中的前列腺癌病变位置，全自动检测替代了专家主观诊断，充分利用已有数据，节约了大量的人力物力成本，并提高了从MRI数据中预测前列腺癌病变位置的准确率，具有很高的实际应用价值。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储与一非易失性计算机可读取存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的和实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可以包括只读存储器（ROM）、可编程ROM（PROM）、电可编程ROM（EPROM）、电可擦除可编程ROM（EEPROM）或闪存。易失性存储器可包括随机存取存储器（RAM）或者外部高速缓冲存储器。作为说明而非局限，RAM通过多种形式可得，诸如静态RAM（SRAM）、动态RAM（DRAM）、同步DRAM（SDRAM）、双速据率SDRAM（SSRSDRAM）、增强型SDRAM（ESDRAM）、同步链路（Synchlink）DRAM（SLDRAM）、存储器总线（Rambus）直接RAM（RDRAM）、直接存储器总线动态RAM（DRDRAM）、以及存储器总线动态RAM（RDRAM）等。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其它变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、装置、物品或者方法不仅包括那些要素，而且还包括没有明确列出的其它要素，或者是还包括为这种过程、装置、物品或者方法所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、装置、物品或者方法中还存在另外的相同要素。
以上所述仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其它相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种MRI病变位置检测方法，其特征在于，包括以下步骤：
    将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置的检测模型；所述训练样本为已知病变位置的MRI数据；
    接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的病变位置；所述待检测样本为未知病变位置的MRI数据。
  2. 根据权利要求1所述的MRI病变位置检测方法，其特征在于，所述将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数的步骤之前，包括：
    对原始MRI数据进行预处理，得到所述训练样本；所述原始MRI数据为患者的患病部位MRI数据。
  3. 根据权利要求2所述的MRI病变位置检测方法，其特征在于，所述对原始MRI数据进行预处理，得到所述训练样本的步骤，包括：
    将所述原始MRI数据中包含的四种序列数据进行矫正对齐，消除四种所述序列数据之间的异质性；四种所述序列数据分别为扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据；
    在所述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据中随机选出三种不同的数据组合成RGB三通道图像，得到所述训练样本；所述RGB三通道图像为3D图像。
  4. 根据权利要求3所述的MRI病变位置检测方法，其特征在于，所述对原始MRI数据进行预处理，得到所述训练样本的步骤之后，包括：
    对所述训练样本进行数据增广处理，以增加所述训练样本的数据量。
  5. 根据权利要求4所述的MRI病变位置检测方法，其特征在于，所述对所述训练样本进行数据增广处理，以增加所述训练样本的数据量的步骤，包括：
    针对所述RGB三通道图像，从多个不同方向进行切片；
    对每个所述切片进行平面旋转、剪切以及归一化处理，并将每个所述切片作为一个训练样本。
  6. 根据权利要求1所述的MRI病变位置检测方法，其特征在于，所述卷积神经网络使用Adam方法进行参数训练。
  7. 根据权利要求1所述的MRI病变位置检测方法，其特征在于，所述将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型的步骤之后，包括：
    将测试样本输入至所述检测模型中进行验证，验证所述训练参数；所述测试样本为已知病变位置的MRI数据。
  8. 一种MRI病变位置检测装置，其特征在于，包括：
    训练单元，用于将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置检测模型；所述训练样本为已知病变位置的MRI数据；
    检测单元，用于接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的病变位置；所述待检测样本为未知病变位置的MRI数据。
  9. 根据权权要求8所述的MRI病变位置检测装置，其特征在于，还包括：
    预处理单元，用于对原始MRI数据进行预处理，得到所述训练样本；所述原始MRI数据为患者的患病部位MRI数据。
  10. 根据权权要求9所述的MRI病变位置检测装置，其特征在于，所述预处理单元包括：
    对齐模块，用于将所述原始MRI数据中包含的四种序列数据进行矫正对齐，消除四种所述序列数据之间的异质性；四种所述序列数据分别为扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据；
    组合模块，用于在所述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据中随机选出三种不同的数据组合成RGB三通道图像，得到所述训练样本；所述RGB三通道图像为3D图像。
  11. 根据权权要求10所述的MRI病变位置检测装置，其特征在于，还包括：
    增广单元，用于对所述训练样本进行数据增广处理，以增加所述训练样本的数据量。
  12. 根据权权要求11所述的MRI病变位置检测装置，其特征在于，所述增广单元包括：
    切片模块，用于针对所述RGB三通道图像，从多个不同方向进行切片；
    归一化模块，用于对每个所述切片进行平面旋转、剪切以及归一化处理，并将每个所述切片作为一个训练样本。
  13. 根据权权要求8所述的MRI病变位置检测装置，其特征在于，所述卷积神经网络使用Adam方法进行参数训练。
  14. 根据权权要求8所述的MRI病变位置检测装置，其特征在于，还包括：
    验证单元，用于将测试样本输入至所述检测模型中进行验证，验证所述训练参数；所述测试样本为已知病变位置的MRI数据。
  15. 一种计算机设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现MRI病变位置检测方法，所述方法包括：
    将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置的检测模型；所述训练样本为已知病变位置的MRI数据；
    接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的病变位置；所述待检测样本为未知病变位置的MRI数据。
  16. 根据权利要求15所述的计算机设备，其特征在于，所述处理器将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数的步骤之前，包括：
    对原始MRI数据进行预处理，得到所述训练样本；所述原始MRI数据为前列腺癌患者的前列腺部位MRI数据。
  17. 根据权利要求16所述的计算机设备，其特征在于，所述处理器对原始MRI数据进行预处理，得到所述训练样本的步骤，包括：
    将所述原始MRI数据中包含的四种序列数据进行矫正对齐，消除四种所述序列数据之间的异质性；四种所述序列数据分别为扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据；
    在所述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据中随机选出三种不同的数据组合成RGB三通道图像，得到所述训练样本；所述RGB三通道图像为3D图像。
  18. 一种计算机非易失性可读存储介质，其上存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现MRI病变位置检测方法，所述方法包括：
    将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数，将训练出所述训练参数的卷积神经网络作为检测MRI数据中病变位置的检测模型；所述训练样本为已知病变位置的MRI数据；
    接收待检测样本，将所述待检测样本输入至所述检测模型中进行预测，预测出所述待检测样本中的病变位置；所述待检测样本为未知病变位置的MRI数据。
  19. 根据权利要求18所述的计算机非易失性可读存储介质，其特征在于，所述处理器将训练样本输入至卷积神经网络中进行参数训练，通过训练计算出所述卷积神经网络的训练参数的步骤之前，包括：
    对原始MRI数据进行预处理，得到所述训练样本；所述原始MRI数据为前列腺癌患者的前列腺部位MRI数据。
  20. 根据权利要求19所述的计算机非易失性可读存储介质，其特征在于，所述处理器对原始MRI数据进行预处理，得到所述训练样本的步骤，包括：
    将所述原始MRI数据中包含的四种序列数据进行矫正对齐，消除四种所述序列数据之间的异质性；四种所述序列数据分别为扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据；
    在所述扩散加权成像数据、表观扩散系数数据、动态增强定量参数数据以及加权图像数据中随机选出三种不同的数据组合成RGB三通道图像，得到所述训练样本；所述RGB三通道图像为3D图像。
PCT/CN2018/095493 2018-04-20 2018-07-12 Mri病变位置检测方法、装置、计算机设备和存储介质 WO2019200745A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810361427.2A CN108765368A (zh) 2018-04-20 2018-04-20 Mri病变位置检测方法、装置、计算机设备和存储介质
CN201810361427.2 2018-04-20

Publications (1)

Publication Number Publication Date
WO2019200745A1 true WO2019200745A1 (zh) 2019-10-24

Family

ID=64011104

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/095493 WO2019200745A1 (zh) 2018-04-20 2018-07-12 Mri病变位置检测方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN108765368A (zh)
WO (1) WO2019200745A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111275677A (zh) * 2020-01-17 2020-06-12 哈尔滨工业大学 一种基于卷积神经网络的天花板震害的识别方法
CN111415333A (zh) * 2020-03-05 2020-07-14 北京深睿博联科技有限责任公司 乳腺x射线影像反对称生成分析模型训练方法和装置
US20210391078A1 (en) * 2018-10-11 2021-12-16 Jlk Inc. Deep learning model learning device and method for cancer region

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109712128B (zh) * 2018-12-24 2020-12-01 上海联影医疗科技有限公司 特征点检测方法、装置、计算机设备和存储介质
CN110070101B (zh) * 2019-03-12 2024-05-14 平安科技（深圳）有限公司 植物种类的识别方法及装置、存储介质、计算机设备
CN109994201B (zh) * 2019-03-18 2021-06-11 浙江大学 一种基于深度学习的糖尿病与高血压概率计算系统

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102947840A (zh) * 2010-01-22 2013-02-27 纽约州立大学研究基金会 前列腺可视化和癌症检测的系统和方法
CN104424386A (zh) * 2013-08-23 2015-03-18 北京大学 基于多参数磁共振影像的前列腺癌计算机辅助识别系统
CN106096616A (zh) * 2016-06-08 2016-11-09 四川大学华西医院 一种基于深度学习的磁共振影像特征提取及分类方法
CN107133638A (zh) * 2017-04-01 2017-09-05 中南民族大学 基于两分类器的多参数mri前列腺癌cad方法及系统
WO2017151757A1 (en) * 2016-03-01 2017-09-08 The United States Of America, As Represented By The Secretary, Department Of Health And Human Services Recurrent neural feedback model for automated image annotation
CN107563123A (zh) * 2017-09-27 2018-01-09 百度在线网络技术（北京）有限公司 用于标注医学图像的方法和装置
US20180033144A1 (en) * 2016-09-21 2018-02-01 Realize, Inc. Anomaly detection in volumetric images
CN107767378A (zh) * 2017-11-13 2018-03-06 浙江中医药大学 基于深度神经网络的gbm多模态磁共振图像分割方法

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102947840A (zh) * 2010-01-22 2013-02-27 纽约州立大学研究基金会 前列腺可视化和癌症检测的系统和方法
CN104424386A (zh) * 2013-08-23 2015-03-18 北京大学 基于多参数磁共振影像的前列腺癌计算机辅助识别系统
WO2017151757A1 (en) * 2016-03-01 2017-09-08 The United States Of America, As Represented By The Secretary, Department Of Health And Human Services Recurrent neural feedback model for automated image annotation
CN106096616A (zh) * 2016-06-08 2016-11-09 四川大学华西医院 一种基于深度学习的磁共振影像特征提取及分类方法
US20180033144A1 (en) * 2016-09-21 2018-02-01 Realize, Inc. Anomaly detection in volumetric images
CN107133638A (zh) * 2017-04-01 2017-09-05 中南民族大学 基于两分类器的多参数mri前列腺癌cad方法及系统
CN107563123A (zh) * 2017-09-27 2018-01-09 百度在线网络技术（北京）有限公司 用于标注医学图像的方法和装置
CN107767378A (zh) * 2017-11-13 2018-03-06 浙江中医药大学 基于深度神经网络的gbm多模态磁共振图像分割方法

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210391078A1 (en) * 2018-10-11 2021-12-16 Jlk Inc. Deep learning model learning device and method for cancer region
US11769594B2 (en) * 2018-10-11 2023-09-26 Jlk Inc. Deep learning model learning device and method for cancer region
CN111275677A (zh) * 2020-01-17 2020-06-12 哈尔滨工业大学 一种基于卷积神经网络的天花板震害的识别方法
CN111415333A (zh) * 2020-03-05 2020-07-14 北京深睿博联科技有限责任公司 乳腺x射线影像反对称生成分析模型训练方法和装置
CN111415333B (zh) * 2020-03-05 2023-12-01 北京深睿博联科技有限责任公司 乳腺x射线影像反对称生成分析模型训练方法和装置

Also Published As

Publication number Publication date
CN108765368A (zh) 2018-11-06

Similar Documents

Publication Publication Date Title
WO2019200745A1 (zh) Mri病变位置检测方法、装置、计算机设备和存储介质
KR101857624B1 (ko) 임상 정보를 반영한 의료 진단 방법 및 이를 이용하는 장치
US10755410B2 (en) Method and apparatus for acquiring information
US8798345B2 (en) Diagnosis processing device, diagnosis processing system, diagnosis processing method, diagnosis processing program and computer-readable recording medium, and classification processing device
US20200402204A1 (en) Medical imaging using neural networks
Polat et al. COVID-19 diagnosis from chest X-ray images using transfer learning: Enhanced performance by debiasing dataloader
CN109447962A (zh) 一种基于卷积神经网络的眼底图像硬性渗出物病变检测方法
CN110827250A (zh) 基于轻量级卷积神经网络的智能医学图像质量评估方法
CN112348785B (zh) 一种癫痫病灶定位方法及系统
CN107862665B (zh) Ct图像序列的增强方法及装置
CN110634144B (zh) 一种卵圆孔定位方法、装置和存储介质
CN116703901B (zh) 肺部医学ct影像分割及分类装置及设备
WO2021138087A1 (en) Systems and methods for processing electronic images for generalized disease detection
CN112950644B (zh) 基于深度学习的新生儿大脑图像分割方法及模型构建方法
CN111652775A (zh) 居家服务过程管理体系模型建构方法
CN116721761A (zh) 一种放疗数据处理方法、系统、设备及介质
CN108510489A (zh) 一种基于深度学习的尘肺检测方法与系统
CN111370098A (zh) 一种基于边缘侧计算和服务装置的病理诊断系统及方法
CN114121225A (zh) Mri影像预测前列腺癌根治术后生化复发风险方法
CN116797521A (zh) 肺炎重症化预测方法、装置、电子设备及介质
WO2023226217A1 (zh) 微卫星不稳定预测系统及其构建方法、终端设备及介质
CN116543154A (zh) 一种基于多层次语义特征的医学影像分割方法
US11416983B2 (en) Server-client architecture in digital pathology
US11501451B2 (en) Detecting subject motion in medical imaging
CN107256544A (zh) 一种基于vcg16的前列腺癌图像诊断方法及系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18915250

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18915250

Country of ref document: EP

Kind code of ref document: A1