WO2016192612A1 - 基于深度学习对医疗数据进行分析的方法及其智能分析仪 - Google Patents

基于深度学习对医疗数据进行分析的方法及其智能分析仪 Download PDF

Info

Publication number
WO2016192612A1
WO2016192612A1 PCT/CN2016/084000 CN2016084000W WO2016192612A1 WO 2016192612 A1 WO2016192612 A1 WO 2016192612A1 CN 2016084000 W CN2016084000 W CN 2016084000W WO 2016192612 A1 WO2016192612 A1 WO 2016192612A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
medical
model
layer
output
Prior art date
Application number
PCT/CN2016/084000
Other languages
English (en)
French (fr)
Inventor
陈宽
Original Assignee
陈宽
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 陈宽 filed Critical 陈宽
Priority to EP16802535.1A priority Critical patent/EP3306500A4/en
Priority to JP2017559611A priority patent/JP6522161B2/ja
Priority to US15/579,212 priority patent/US11200982B2/en
Publication of WO2016192612A1 publication Critical patent/WO2016192612A1/zh
Priority to IL255856A priority patent/IL255856B/en
Priority to US17/519,873 priority patent/US20220059229A1/en

Links

Images

Classifications

    • 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
    • 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
    • G06N3/08—Learning methods
    • 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
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16Z—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
    • G16Z99/00—Subject matter not provided for in other main groups of this subclass

Definitions

  • the present invention relates to a smart device for analyzing medical or medical data, and more particularly to an intelligent analyzer that automatically aggregates a large number of medical or medical data obtained by a large hospital or medical research institution and gives a matching analysis conclusion.
  • doctors or researchers in large hospitals or medical research institutions need to do a lot of work every day.
  • doctors in the clinical department of the hospital need to conduct research, analysis and decision-making on the collected medical data every day.
  • the following is the medical data generated by randomly taking one day in a large top three hospital in Beijing:
  • the technical problem to be solved by the present invention is to provide a medical study based on deep learning based on deep learning, which can effectively alleviate the work pressure of hospital doctors or medical researchers and can scientifically analyze and obtain a large number of medical or medical data. Analytical methods and their intelligent analyzers.
  • the technical solution adopted by the present invention is:
  • the method for analyzing medical data based on deep learning of the present invention comprises the following steps:
  • a setting a basic framework for deep learning, and establishing a data model including an input layer, at least one hidden layer, and an output layer according to the data feature, the input layer includes a plurality of nodes having data features, and the output layer includes a plurality of a node having medical diagnostic data features, each hidden layer comprising a plurality of nodes having a mapping correspondence with an output value of the upper layer;
  • Each node uses a mathematical equation to establish a data model of the node, and manually or randomly determines the relevant parameter values in the mathematical equation, and the input values of the nodes in the input layer are the data features, and each hidden layer And the input value of each node in the output layer is an output value of the upper layer, and the output value of each node in each layer is a value obtained by the operation of the mathematical equation of the node;
  • the method of optimizing the parameter value A i is an unsupervised learning method.
  • the unsupervised learning method employs a noise reduction automatic coding generator or a Berman machine to perform self-learning.
  • the method of optimizing the parameter value A i is a supervised learning method.
  • the mathematical equation is a parametric mathematical equation or a nonparametric mathematical equation, wherein the parametric mathematical equation can be a linear model, a neuron model or a convolution operation, and the nonparametric mathematical equation can be an extreme value operation equation, and the mathematical model is set as follows:
  • y is the medical diagnostic data feature in the output layer
  • the dimension is M n
  • X is the training material data
  • the dimension is M 0
  • f 1 to f n are the set operation equations of each layer
  • each layer equation f i is the dimension M i-1 ⁇ M i, f 1 as the first layer is to dimension M 0 is converted into the X dimension output of M 1 Z 1, Z 1 and the equation becomes f 2 of the second layer Input, and so on, where each layer model f i has a parameter set A i that matches it.
  • the medical material data includes relevant information records of the doctor's diagnosis, examination and treatment process by the doctor in the clinical and medical technology stages; the diagnosis data includes the doctor's diagnosis of the patient's initial diagnosis, the discharge result, and the disease treatment effect in the clinical and medical technology stage. Relevant information records and textual diagnosis data written by doctors and Follow-up data was tracked.
  • the data characteristics include changes in the time and space of the medical training data, and various mathematical statistics of the data itself. For example, as time goes by, the trend of data rising or falling.
  • the structured data involved in the medical to-be-analyzed data and the matching analysis results are fed back into the deep learning model to form new training data.
  • the intelligent analyzer for analyzing medical data based on deep learning includes an input device that can import medical training data and medical to-be-analyzed data into a computer, and separately or collectively save the medical training data and medical to-be-analyzed data.
  • a storage module a deep learning model module that invokes medical training data in the storage module for self-learning, an output device that derives medical pathological analysis results that match the medical data to be analyzed, and a processor including a CPU and/or a GPU, wherein ,
  • the medical training data includes medical material data and medical diagnostic data matched thereto;
  • the medical training data and the medical to-be-analyzed data are structured data matrices that can be understood by a computer;
  • the self-learning employs a parametric mathematical equation including a linear model, a neuron model, a convolution operation, and/or seeking a maximum operation;
  • the input device includes a computer device disposed in a hospital, a medical institution, various medical examination devices and a pathological analysis device networked with the computer;
  • the output device includes a stationary computer output terminal and a mobile smart terminal that are disposed in a hospital, a medical institution, and networked with the input device.
  • the intelligent analyzer of the present invention is further provided with a network connection module which can be connected to the Internet or Ethernet, including a fiber connection, a WIFI connection or a GPRS module connection.
  • a network connection module which can be connected to the Internet or Ethernet, including a fiber connection, a WIFI connection or a GPRS module connection.
  • the core content of the method of the present invention and its intelligent analyzer is the application of deep convolution in deep learning
  • the meta-algorithm (English full name: Deep Convolution Neural Network, DCNN for short) establishes a model in the computer.
  • the model uses massive medical data to select and optimize model parameters, and automatically learns the pathological analysis process of doctors or medical researchers through the “training” model, which in turn helps them process large amounts of medical or medical data, and ultimately assists doctors in making large amounts of medical data. Correct judgment and effective decision making.
  • the invention can greatly reduce the work pressure of the doctor or the medical researcher and improve the work efficiency thereof, and the invention can free the doctor or the medical researcher from the heavy analysis work of the medical or medical data, thereby using more energy. In other more important work.
  • Figure 1 is a block diagram showing the operation of the intelligent analyzer of the present invention.
  • Figure 2 shows image data generated by brain nuclear magnetic resonance.
  • FIG. 3 is image data after deleting the target body in the image data.
  • Figure 4 is a schematic diagram of the basic mathematical structure of a DNN based on graphical data.
  • Figure 5 is a schematic diagram of a convolutional block operation.
  • Figure 6 is a schematic diagram of the logical operation of the interconnected multilayer perceptron.
  • Figure 7 is a schematic diagram of the workflow of the automatic code generator.
  • the method for analyzing medical data based on deep learning is to select and optimize model parameters by using massive medical data, and automatically learn the pathological analysis process of doctors or medical researchers through the “training” model, and then help them. Handling large amounts of medical or medical data ultimately helps doctors make the right judgments and effective decisions for large amounts of medical data.
  • medical data intelligence analysis systems are a very important area of medical technology.
  • the field of more people's research is the analysis of CT nodules in the lungs, which are mainly divided into two major technical modules: Image segmentation and intelligent analysis.
  • image segmentation is to intelligently segment key parts of the lungs such as the trachea, lungs, blood vessels, etc., and model the 3D images to help clinicians and imaging doctors better analyze the lungs. Structure and preparation before surgery.
  • Image segmentation has very mature technologies and algorithms. However, the main use is in very traditional algorithms such as cascade models, and can not fully exploit the use of intelligent analyzers.
  • the analysis system for graphic segmentation is only for a small part of medical data processing, and the value for doctors is limited.
  • Deep Learning is a revolutionary technology recognized in the field of artificial intelligence. It has subverted traditional application methods in the fields of image recognition and speech recognition, and has successfully brought many breakthrough technology applications: Google Image Content Analysis, Google No People driving cars, Google Book, Google Brain, etc.
  • DNN Deep Neural Network
  • the invention applies the most advanced deep learning algorithm to medical data analysis, and models with massive data to construct a medical data analysis system. It can greatly reduce the pressure on doctors and increase the number of doctors Efficiency.
  • model training modules pre-training
  • model improvement modules fine-tuning
  • the model training module mainly uses medical training data to find the mathematical expression that best represents the medical analysis process.
  • the model application module is a main application module in the intelligent analyzer system, which inputs the medical to-be-analyzed data into the model training module and automatically outputs the medical pathological analysis result that matches the medical to-be-analyzed data.
  • the method of the invention comprises the following steps:
  • the purpose of medical training is to enable the computer to automatically calculate the corresponding medical diagnostic analysis data from the medical material data.
  • the medical material data includes relevant information records of the doctor's diagnosis, examination and treatment process by the doctor in the clinical and medical technology stages; the diagnosis data includes the doctor's diagnosis of the patient's initial diagnosis, the discharge result, and the disease treatment effect in the clinical and medical technology stage. Relevant information records and textual outpatient data and follow-up data written by doctors.
  • the medical material data includes: the doctor writes the input patient information, such as the current medical history, past medical history, physical examination, laboratory and device examination, and the treatment process after admission.
  • the medical diagnosis data (also referred to as target data) includes: a record of the doctor's initial diagnosis and discharge of the patient, and the effect of the disease treatment.
  • the patient is referred to the patient, and the patient's relevant information, such as age, gender, weight, current medical history, past medical history, physical examination information, etc., are integrated, and the analysis data is integrated to provide the patient's disease type analysis, admission advice, and treatment plan. For example, enter information about a patient, 65-year-old male patient, cough, chest tightness, recent weight loss, long-term smoking history, and no previous examinations.
  • the medical material data includes: original image data, pathological types, disease-related test data, specific location of the lesion, presence or absence of metastasis or multiple occurrences.
  • the medical diagnosis data textual diagnosis data written by a doctor, and follow-up data.
  • the intelligent analyzer Through the analysis and training of the original image data of different body parts and different image inspection methods, the intelligent analyzer has the function of recognition and analysis for the lesion, and gives the next step of diagnosis and treatment. For example, CT intelligent diagnosis of single nodules in the lungs, the intelligent analyzer can retrieve all the original images in a very short time, and determine the location, size, internal density, edge morphology, and other parts of the image.
  • the medical training data corresponding to each individual and the change value are summarized into one unit data.
  • the medical training data and the change values to be associated with a person or a series of cases are summarized into one unit of data.
  • the data characteristics include changes in the time and space of the medical training data, and various mathematical statistics of the data itself.
  • Data characteristics include changes in medical training data over time, such as the trend of rising or falling data; spatial changes, such as the relationship between one image and one pixel from one image.
  • the data characteristics also include various mathematical statistics of the data itself, such as individual data and other individual data comparison values.
  • These data features will be formatted as a computer-understood structure in the form of vectors, matrices, or series.
  • the collection of data features also includes image processing or initial data statistics processing. In image processing, segmenting the image content related to the medical diagnosis data is the first step in finding the characteristics of the image data. In document file processing, TF-IDF (term frequency–inverse document frequency), a method of quantitative data retrieval and text mining, can also be applied. The above initial image text processing will greatly facilitate the computer to collect data features.
  • Each node uses a mathematical equation to establish a data model of the node, and uses artificial or random methods to preset relevant parameter values in the mathematical equation.
  • the input values of the nodes in the input layer are the data features, and each hidden layer
  • the input value of each node in the output layer is the output value of the upper layer, and the output value of each node in each layer is The value obtained by the operation of the mathematical equation by the node;
  • the method of optimizing the parameter value A i is an unsupervised learning method and a supervised learning method.
  • the unsupervised learning method employs a noise reduction automatic coding generator or a Berman machine to perform self-learning.
  • the mathematical equation is a parametric mathematical equation or a nonparametric mathematical equation, wherein the parametric mathematical equation can be a linear model, a neuron model or a convolution operation, and the nonparametric mathematical equation can be an extreme value operation equation, and the mathematical model is set as follows:
  • y is the medical diagnostic data feature in the output layer
  • the dimension is M n
  • X is the training material data
  • the dimension is M 0
  • f 1 to f n are the set operation equations of each layer
  • each layer equation f i is the dimension M i-1 ⁇ M i, f 1 as the first layer is to dimension M 0 is converted into the X dimension output of M 1 Z 1, Z 1 and the equation becomes f 2 of the second layer Input, and so on, where each layer model f i has a parameter set A i that matches it.
  • x m is the input value of the equation
  • y is the output value of the equation
  • a m is the basic parameter of the equation.
  • the depth learning model parameters A 1 to A n are initialized, and the model parameters, model depth, etc. can be set arbitrarily, and the initialization parameter model can also be selected in some way.
  • the invention belongs to the artificial intelligence technology, and the ultimate purpose of the data operation is to "train" the model to automatically identify the lesion in the medical image, give the probability and mark, and assist the doctor's diagnosis and treatment work. Therefore, in the process of model construction, massive data is equivalent to teaching material, and the model framework is the specific process of abstracting and summarizing specific information by algorithms. Therefore, in the process of intelligent computing, massive data and intelligent algorithms are indispensable.
  • MRI Magnetic Resonance Imaging, Chinese name MRI
  • graphics can be three-dimensional matrix (grayscale), that is, two-dimensional gray-scale index and one-dimensional cross-section; or four-dimensional matrix (rgb), that is, two-dimensional color The index is followed by three color indices, and finally a one-dimensional cross section.
  • Any medical data can be abstracted abstractly into such a matrix.
  • Such a matrix constitutes the original data source that the model reads.
  • Fig. 2 and Fig. 3 the image data generated by the brain MRI is displayed. It is assumed that the system generates one MRI slice into a pixel of 512 ⁇ 512, and one brain scan is 200 slices, then one gray scale.
  • such medical raw matrices are the basic data for modeling.
  • an analysis target that matches the graph is also needed.
  • the simplest binary analysis information is: measuring the lesion (a slightly more complicated information can be the probability of a lesion). Later, more complex medical information such as the type of lesion, treatment effect, and the specific location of the lesion can be incorporated.
  • the longitudinal time series data of the patient's past physical examination can be matched, and the algorithm can learn to predict the development of medical phenomena.
  • Simulated simulated data This type of data is processed or simulated by a computer, using simulated data as training data for modeling.
  • the most typical example of such a model is Microsoft's Xbox Kinect system.
  • the basic data of the hand gesture recognition model in the development stage is all 3D modeling completed.
  • simulation data can be understood as new data constructed based on the original medical data through deformation, distortion, and noise superposition.
  • analog data There are two reasons for using analog data: First, adding deformed data is beneficial to the church algorithm to more stably identify the core changes in medical data; second, one The general DNN model needs to guide more than a few million parameters. In the case of limited data, it is easy to cause over-fitting, that is, the model over-learns the existing historical data, and can not summarize the core change law well. And the abstract summary, adding the simulated deformation data is equivalent to adding noise during the training process, and the forcing algorithm can better distinguish the noise and effective information, which helps to solve the over-fitting problem.
  • the machine learning algorithm model is the basic mathematical framework used by the invention to summarize and summarize the information.
  • the main purpose is to express the pattern recognition process in a mathematical structure that the computer can understand.
  • the training process is to estimate the parameters in the model. After the parameters are estimated, the model will become the core part of the method of the present invention.
  • Machine learning algorithms can be classified into two categories, supervised learning and unsupervised learning, according to different purposes.
  • the present invention covers two types of algorithms.
  • Supervised learning algorithms emphasize the target law that people seek to set up. As described in the previous section, in addition to the original graphics matrix data, the supervised learning algorithm also requires matching analysis conclusion data (such as the medical diagnostic data described).
  • This patent mainly includes the following supervised learning algorithms.
  • the basic principle of this algorithm mimics the process of human brain discrimination.
  • the input of the DNN algorithm is the original medical data and the doctor's historical analysis results, and finally the analysis process can be completed automatically.
  • the abstract summary of DNN is
  • x is the original medical matrix data
  • y is the intelligent system analysis result
  • DNN is the mathematical mapping expression of the equation f, x to y.
  • the DNN algorithm simulates the neuron structure of the human brain.
  • the basic mathematical structure of DNN based on graphical data is shown in Fig. 4.
  • the first layer is the convolution layer and the second layer is the largest pool layer. This loops the structure of the DNN from the left graph raw data to the rightmost analysis result. Can be divided into multiple layers, each layer to complete different mathematical operations.
  • the model has a total of multiple layers of neuronal structures.
  • the first layer performs multiple parallel inner product operations for medical data.
  • the most commonly used algorithm in the first layer is convolution.
  • the convolution algorithm outputs the inner product of the new equation and the original series data by sliding a new equation over the input series of values. As in 3D medical graphics, the algorithm constructs multiple convolutional squares, each of which is a matrix of measurements.
  • the convolutional block x and y axes cover the equation of an image's spatial variation, while the z-axis of the convolutional block covers the equation of the image's spatial variation.
  • Each convolutional square matrix slides along the data dimension itself to calculate the inner product of the values of the respective dimensions of the 3D graphics and the convolutional square.
  • the value of the inner product operation can be understood as the similarity between the data dimension and the convolutional square, and
  • the inner product value output by each part of the data will be the input value of the next layer of neurons. From the perspective of intuitive image, the parallel convolution matrix is equivalent to a specific shape, and the outer calculation of the convolution matrix is equivalent to judging whether different regions are similar to a specific shape in the data.
  • the second common layer of the DNN model is the level of the pooling operation (English name is Max Pooling, referred to as MP operation).
  • the MP operation process synthesizes the dimensional information into a larger range of squares, and performs a Max (seeking maximum) operation in each square.
  • MP operations mainly mimic the active economic characteristics of neurons in visual neural networks. Within a certain range of information frameworks, only the most active information units are retained and proceed to the next level. From a graphical point of view, the MP operation causes the result of the operation to no longer change due to the rotation of the data itself. From the perspective of operation, the MP operation is equivalent to the dimensionality reduction processing, combined with the first layer of neuron operations, the MP operation removes the region information with lower similarity to the first layer convolutional block, and reduces the invalid information in each region. content.
  • the structure of the DNN is often combined and continuously repeated by the convolutional layer and the pooled layer to extract medical diagnosis.
  • Related data characteristics Intuitively, for example, in Figure 4, the middle layer algorithm uses the nonlinear elements of the first and second layers, and so on. This is to build a more abstract framework element. Numerous layers of neurons can be constructed by the method of this patent.
  • the combination of the convolutional layer and the pooled layer should be repeated as much as possible.
  • the human brain belongs to a very deep neuron structure. Therefore, the deeper the neuron model, the stronger the power. However, the deeper the neuronal structural parameters, the more difficult it is to train and estimate the parameters, and it is easy to have derivative demise and over-fitting problems.
  • the remaining information enters the final Multi-Layer-Perceptron MLP.
  • the basic structure of the perceptron is a two-layer logistic regression operation, which is equivalent to appending the contribution of different abstract graphic elements to the final evaluation result, and the final output value of the MLP operation is the medical analysis result of the model.
  • the perceptron is generally a layer of implicit perceptual layer. The variables of each unit are completely interconnected with all variables in the upper layer, and each layer performs logical operations to obtain the next layer of values.
  • Unsupervised learning The concept of neural networks has existed for many years, but it is limited by the amount of data that can be used and the computing power of the processor, making the derivative demise problem very serious and cannot be used to solve practical problems.
  • the error between the output predicted value and the actual value of the model constitutes the basis of model parameter optimization, and the excessively deep neural network structure cannot reverse the parameter optimization information to the underlying network, ie, the surface layer. Information cannot be transmitted to the deep network structure layer by layer, which brings great difficulties to model training.
  • the volume of medical data is often very large. It is unrealistic to perform a complete optimization search, and the problem of high computational difficulty is more serious than other fields.
  • the invention performs preliminary optimization on parameters in the model through unsupervised learning, so that the model The initial conditions of the parameters during the optimization process become very advantageous, allowing the model optimization process to find local minima faster.
  • the two most unsupervised learning methods are Denoising Autoencoders (dAE) and Restricted Boltzmann Machine (RBM).
  • dAE Noise reduction automatic coding generator
  • the principle of the automatic code generator is to find valid implicit variables of a certain data variable.
  • the working principle of the automatic code generator is completely presented.
  • the automatic code generator looks for the recessive element representative y and the parameter W to map the new data z, and automatically generates the code.
  • the ultimate goal is to find the parameter W to minimize the difference between z and x, in other words to find the parameter that best represents the data variable information within the limited information.
  • These parameters can be considered to cover the largest amount of raw data information within the model range.
  • the noise reduction automatic coding generator artificially introduces a large amount of noise in the working principle of the simple coding generator.
  • the noise is forced to find a more valuable potential law through a large amount of noise, without being affected by the invalid law in the noise.
  • the final trained parameters will become the initial parameter starting values for supervised learning, which is equivalent to finding a good starting point for the first step of the model and greatly speeding up the optimization of the parameters of the trained model.
  • a further improvement of the optimization data is to establish a loss equation or a target equation, and perform supervised learning data model parameter optimization according to the loss equation.
  • the loss equation can be set as the difference between the analysis result output by the deep learning model and the actual target variable in the training data.
  • the parameter value in the model is adjusted according to the change of the loss equation and the optimization method.
  • the loss equation can be set as the difference between the measured value and the actual value generated by the model (such as the variance difference), and the parameter optimization by gradient descent
  • the method moves the parameter values in each cycle, and stops the parameter optimization process after the parameter optimization cycle meets certain conditions (for example, the value of the loss equation before the cycle and the cycle is less than a certain threshold, or the number of cycles exceeds a certain number to stop the optimization operation) ), retain the best value.
  • g(x, A) is the analytical output of the basic deep learning
  • Y is the actual value of the analysis target
  • L[Y;g(X;A)] is mainly used to calculate the difference between the analytical output and the actual value of the deep learning. The cost incurred.
  • R(A) is mainly a regularization expression, the main function is to avoid over-fitting of the model.
  • the method of parameter optimization can be arbitrarily chosen.
  • the most common method is the gradient reduction method.
  • the mathematical expression of the steps is as follows:
  • the parameter is moved in the opposite direction to the differential of the target equation. After repeated rounds of movement, the movement is stopped if a specific stop condition is satisfied.
  • Another step-by-step improvement in optimizing data is the data noise-increasing method, which artificially adds noise to the model and data to stabilize the model and over-fitting the data model.
  • the data noise-increasing method which artificially adds noise to the model and data to stabilize the model and over-fitting the data model.
  • the original data can be deformed and distorted, and the model is forced to recognize valid information other than noise.
  • test sample segmentation Another improvement of the optimized data is test sample segmentation, which can further segment the training data out of the test sample, build the model by using the remaining training data, test the validity of the model by testing the sample, and adjust the depth learning model automatically or manually according to the result.
  • Core framework Another improvement of the optimized data is test sample segmentation, which can further segment the training data out of the test sample, build the model by using the remaining training data, test the validity of the model by testing the sample, and adjust the depth learning model automatically or manually according to the result.
  • the medical learning result obtained by the depth learning model is matched with the medical to-be-analyzed data by the output device.
  • the structured data related to the medical to-be-analyzed data and the matching analysis result are fed back into the deep learning model to form new training data to further optimize the deep learning model.
  • the intelligent analyzer for analyzing medical data based on deep learning includes an input device that can import medical training data and medical to-be-analyzed data into a computer, and separately or collectively save the medical training data and medical to-be-analyzed data.
  • a storage module a deep learning model module in the method of the present invention that invokes medical training data in the storage module for self-learning, an output device that derives medical pathological analysis results matching the medical data to be analyzed, and a CPU and/or GPU Processor, where
  • the medical training data includes medical material data and medical diagnostic data matched thereto;
  • the medical training data and the medical to-be-analyzed data are structured data matrices that can be understood by a computer;
  • the self-learning employs a parametric mathematical equation including a linear model, a neuron model, a convolution operation, and/or seeking a maximum operation;
  • the input device comprises a computer device installed in a hospital, a medical institution, various medical examination devices and pathological analysis devices networked with the computer; such as a computer, a color ultrasound instrument, an X-ray, a synchronous electrocardiograph, a biochemical analyzer, and an immunoassay Instruments, fiberscopes, nuclear magnetic resonance, CT Doppler diagnostics, sphygmomanometers, weight scales, etc.
  • a computer installed in a hospital, a medical institution, various medical examination devices and pathological analysis devices networked with the computer; such as a computer, a color ultrasound instrument, an X-ray, a synchronous electrocardiograph, a biochemical analyzer, and an immunoassay Instruments, fiberscopes, nuclear magnetic resonance, CT Doppler diagnostics, sphygmomanometers, weight scales, etc.
  • the output device includes a stationary computer output terminal and a mobile smart terminal that are disposed in a hospital, a medical institution, and networked with the input device.
  • a stationary computer output terminal and a mobile smart terminal that are disposed in a hospital, a medical institution, and networked with the input device.
  • a network connection module including a fiber connection, a WIFI connection or a GPRS module connection that can be connected to the Internet or the Ethernet can be installed on the intelligent analyzer of the present invention.
  • the intelligent analyzer of the present invention applies the training-derived deep learning model to the actual, and is a complete integrated system.
  • the new medical data i.e., the medical data to be analyzed
  • the intelligent analyzer becomes an additional plug-in in the analysis process.
  • it can be an additional plug-in in medical devices, or it can be an insertion interface in a system commonly used for PACS (Chinese name for image archiving and communication system) or HIS (Chinese name for hospital information), or via the Internet.
  • PACS Choinese name for image archiving and communication system
  • HIS Choinese name for hospital information

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

一种能有效减轻医院医生或医学研究人员工作压力且可对大量的医疗或医学数据进行科学分析并获得与之匹配的分析结果的基于深度学习对医疗数据进行分析的方法及其智能分析仪。其核心内容是应用深度学习中的深度卷积神经元算法在计算机内建立模型。该模型利用海量医疗数据选择及优化模型参数，通过"训练"模型自动学习医生或医学研究人员的病理分析过程，继而帮助其处理大量的医疗或医学数据，最终辅助医生做出针对大量医疗数据的正确判断和有效决策。本发明可大大降低医生或医学研究人员的工作压力，提高其工作效率，本发明可使医生或医学研究人员从繁重的对医疗或医学数据的分析工作中解脱出来，从而将更多的精力用于其它更重要的工作中。

Description

基于深度学习对医疗数据进行分析的方法及其智能分析仪 技术领域
本发明涉及对医疗或医学数据进行分析的智能设备，特别涉及将大型医院或医学研究机构获取的大量医疗或医学数据自动汇总并给出与之匹配的分析结论的智能分析仪。
背景技术
通常，大医院或医学研究机构的医生或研究人员每天需要完成大量的工作。例如医院临床科室的医生，每天需要对采集而来的医疗数据进行研究、分析和决策，以下为在北京一家大型三甲医院随意抽取一天所产生的医疗数据：
CT 1162份、X射线1461份、核磁共振325份，其中，X射线平均一份含2张二维图片，CT平均一份含50张二维图片，核磁共振平均一份含100张二维图片。这些检查数据全部都要放射科医生撰写报告，而该大型三甲医院也只有20多个放射科医生，其中还包括打字缓慢的老医生和经验不足的年轻医生。因此，每天持续处理大量的工作极易导致相关联的医生或研究人员工作压力大、体能下降快、工作效率低甚至分析结论错误率高等问题。
发明内容
本发明要解决的技术问题是提供一种能有效减轻医院医生或医学研究人员工作压力且可对大量的医疗或医学数据进行科学分析并获得与之匹配的分析结果的基于深度学习对医疗数据进行分析的方法及其智能分析仪。
为了解决上述技术问题，本发明采用的技术方案为：
本发明的基于深度学习对医疗数据进行分析的方法，其包括如下步骤：
1)采集海量已备案的同类型的医疗素材数据及与该医疗素材数据匹配的医疗诊断数据作为医疗训练数据通过输入装置存储于计算机中；
2)将所述医疗训练数据中不小于二维的影像数据与文本数据中随时间和空间的变化值与对应的数据相关联；
3)在采集的海量医疗训练数据中，将与每一个个体对应的医疗训练数据和所述变化值汇总为一条单元数据；
4)将所述医疗训练数据采用分割、关联或文本数据挖掘方法整合或格式化为计算机可以理解的结构化数据矩阵并从每个单元数据中提取数据特征；
5)将已形成结构化数据矩阵的医疗训练数据导入设置于计算机内对应深度学习模型的存储模块中；
6)通过计算机对所述深度学习模型进行优化运算，优化方法如下：
a.设定深度学习基本框架，将所述医疗训练数据按照数据特征建立包括输入层、至少一层隐层和输出层的数据模型，输入层包含若干个具有数据特征的节点，输出层包含若干个具有医疗诊断数据特征的节点，每个隐层包含若干个与上一层输出值具有映射对应关系的节点；
b.每个节点采用数学方程建立该节点的数据模型，采用人工或随机方法预设所述数学方程中的相关参数值，输入层中各节点的输入值为所述的数据特征，各隐层及输出层中各节点的输入值为上层的输出值，每层中各节点的输出值为本节点经所述数学方程运算后所得的值；
c.初始化所述参数值Ai，将所述输出层中各节点的输出值与对应节点的医疗诊断数据特征比对，反复修正各节点的所述参数值Ai，依次循环，最终获得 使所述输出层中各节点的输出值生成与所述医疗诊断数据特征相似度为局部最大时的输出值对应的各节点中的参数值Ai；
7)将获取的已形成结构化矩阵数据的医学待分析数据导入该深度学习模型中进行与之匹配的医学病理分析；
8)由该深度学习模型通过输出装置输出与所述医学待分析数据相匹配的医学病理分析结果。
对所述参数值Ai进行优化的方法为无监督学习方法。
所述无监督学习方法采用降噪自动编码生成器或限制伯尔曼机进行自学习。
对所述参数值Ai进行优化的方法为有监督学习方法。
所述数学方程为参数数学方程或非参数数学方程，其中，参数数学方程可为线性模型、神经元模型或卷积运算，非参数数学方程可为极值运算方程，数学模型设定方式如下：
y＝g(X)＝fn○fn-1○fn-2○…○f1(X)
其中y是所述输出层中的医疗诊断数据特征，维度为Mn，X是训练素材数据，维度为M0，f1到fn为设定的每一层运算方程，而每一层方程fi的维度为Mi－1→Mi,如第一层f1就是将维度为M0的X转换成维度为M1的输出Z1，而Z1则成为第二层方程f2的输入，以此类推，其中,每一层模型fi有与之相匹配的参数组Ai。
所述医疗素材数据包括临床和医技阶段医生对患者诊断、检查和治疗过程进行的相关信息记录；所述诊断数据包括临床和医技阶段医生对患者初诊判断、出院结果、疾病治疗效果进行的相关信息记录以及医生撰写的文本出诊数据和 跟踪随访数据。
所述数据特征包括医疗训练数据在时空上的变化值、数据本身的各种数理统计值。比如说随着时间的改变，数据上升或下降的趋势。
将所述医学待分析数据和与之匹配的分析结果涉及的结构化数据反馈到所述深度学习模型中形成新的训练数据。
本发明的基于深度学习对医疗数据进行分析的智能分析仪，其包括可将医疗训练数据和医学待分析数据导入计算机中的输入装置、分别或集中保存所述医疗训练数据和医学待分析数据的存储模块、调用存储模块中的医疗训练数据进行自学习的深度学习模型模块、将与所述医学待分析数据匹配的医学病理分析结果导出的输出装置和包括CPU和/或GPU的处理器，其中，
所述医疗训练数据包括医疗素材数据和与之匹配的医疗诊断数据；
所述医疗训练数据和医学待分析数据为计算机可以理解的结构化数据矩阵；
所述自学习采用包括线性模型、神经元模型、卷积运算和/或寻求最大值运算的参数数学方程；
所述输入装置包括设置在医院、医学机构的计算机装置、与该计算机联网的各种医疗检查装置和病理分析装置；
所述输出装置包括设置在医院、医学机构中并与所述输入装置联网的固定式计算机输出终端和移动式智能终端。
本发明的智能分析仪还设有可与互联网、以太网连接的包括光纤连接、WIFI连接或GPRS模块连接的网络连接模块。
本发明的方法及其智能分析仪的核心内容是应用深度学习中的深度卷积神 经元算法(英文全称：Deep Convolution Neural Network，简称DCNN)在计算机内建立模型。该模型利用海量医疗数据选择及优化模型参数，通过“训练”模型自动学习医生或医学研究人员的病理分析过程，继而帮助其处理大量的医疗或医学数据，最终辅助医生做出针对大量医疗数据的正确判断和有效决策。本发明可大大降低医生或医学研究人员的工作压力，提高其工作效率，本发明可使医生或医学研究人员从繁重的对医疗或医学数据的分析工作中解脱出来，从而将更多的精力用于其它更重要的工作中。
附图说明
图1为本发明智能分析仪工作方框图。
图2为脑部核磁共振所生成的图像数据。
图3为删除图像数据中目标体之外后的图像数据。
图4为以图形数据为基础的DNN基本数学构造示意图。
图5为卷积方块运算示意图。
图6为互联多层感知器逻辑运算示意图。
图7为自动编码生成器的工作流程示意图。
具体实施方式
如图1所示，本发明的基于深度学习对医疗数据进行分析的方法是利用海量医疗数据选择及优化模型参数，通过“训练”模型自动学习医生或医学研究人员的病理分析过程，继而帮助其处理大量的医疗或医学数据，最终辅助医生做出针对大量医疗数据的正确判断和有效决策。
通常，医疗数据智能分析系统是医疗科技非常重要的领域。比如医疗影像数据方面，较多人研究的领域是肺部CT结节的分析，主要分成两大技术模块： 图像分割(segmentation)与智能分析(detection)。图像分割的主要目的在于将肺部的关键部位如气管、肺叶、血管等关键部位进行智能分割，并通过3D图像的方式进行建模展示，以帮助临床医生和影像科医生更好地分析肺部结构和做术前准备。图像分割目前已经有非常成熟的技术与算法。不过主要都是使用在非常传统的如cascade模型算法，并不能充分发挥智能分析仪的用途。其次针对图形分割的分析系统只是针对医疗数据处理当中很小的部分，对于医生的价值也有限。
Deep Learning深度学习是目前人工智能领域公认的革命性技术，在图像识别、语音识别等领域都颠覆了传统的应用方法，并成功带来了很多突破性的技术应用：谷歌图片内容分析，谷歌无人驾驶车、Google Book、Google Brain等。
但如今，在医疗数据分析领域，绝大部分的方法都还是使用非常传统的Support Vector Machine等分类方法，并不能代表目前人工智能领域最先进的技术。比如同类功能专利CN201110376737.X当中所使用的就是Gradient Boosting方法，是机器学习领域过去1995到2005年间最为广泛应用的方法，其现已不能代表人工智能领域最先进的方法。
2D和3D影像识别算法当中目前最先进的公认为Deep Neural Network(DNN)深度神经网络算法(详见论文Bengio-2009，引用：Yoshua Bengio，“Learning Deep Architectures for AI”，Foundations and
Figure PCTCN2016084000-appb-000001
in Machine Learning2(1),1-127)，在某些拥有海量训练数据的领域如手写数字识别、红绿灯识别甚至可以达到和超过人为识别的准确率。
本发明将最先进的深度学习算法应用到医疗数据分析当中，配合海量数据进行建模，构建医疗数据分析系统。其可大大降低医生工作压力，增加医生工 作效率。
其主要包括模型训练模块(pre-training)和模型改进模块(fine-tuning)。
模型训练模块主要使用医疗训练数据寻找最能够代表医疗分析过程的数学表达方式。模型应用模块是智能分析仪系统中的主要应用模块，其将医学待分析数据输入到模型训练模块中并由该模块自动输出与所述医学待分析数据相匹配的医学病理分析结果。
以下对本发明进行详细说明。
本发明的方法包括如下步骤：
一、采集海量已备案的同类型的医疗素材数据及与该医疗素材数据匹配的医疗诊断数据作为医疗训练数据通过输入装置存储于计算机中。
医疗训练的目的是让计算机能够从医疗素材数据中自动推算出相对应的医疗诊断分析数据。
所述医疗素材数据包括临床和医技阶段医生对患者诊断、检查和治疗过程进行的相关信息记录；所述诊断数据包括临床和医技阶段医生对患者初诊判断、出院结果、疾病治疗效果进行的相关信息记录以及医生撰写的文本出诊数据和跟踪随访数据。
例如在临床方面(外科、内科等)：
所述医疗素材数据包括：医生撰写输入的患者信息，如现病史、既往病史、体格检查、实验室及器械检查、入院后的治疗过程等记录。
所述医疗诊断数据(又称目标数据)包括：医生对患者的入院初诊及出院结果、疾病治疗效果等记录。
临床举例：
接诊患者，输入患者的相关信息，如年龄、性别、体重、现病史、既往病史、体格检查信息等，整合分析数据，提供该患者疾病种类分析、接诊建议及拟治疗方案。例如输入一名患者的相关信息，65岁男性患者，咳嗽、胸闷、近期消瘦、长期吸烟史、既往未做过检查等。
医技方面(病理科、检验科、放射科、核医学科等)
所述医疗素材数据包括：原始图像数据、病理种类、疾病相关检验数据、病灶具体位置、有无转移或多发等。
所述医疗诊断数据：医生撰写的文本出诊数据，跟踪随访数据。
医技举例：
放射科：通过对不同身体部位、不同影像检查手段的原始图像数据的分析训练，使智能分析仪对于病变具有识别、分析功能，并给出下一步诊疗建议。如肺部单发结节的CT智能诊断，智能分析仪可在极短时间内检索所有原始图像，判断病变所在位置、大小、内部密度、边缘形态、图像内其它部位是否正常等数据。
二、将所述医疗训练数据中不小于二维的影像数据与文本数据中随时间和空间的变化值与对应的数据相关联。或者说，将同一案例的医疗素材数据和医疗诊断数据互相关联。
三、在采集的海量医疗训练数据中，将与每一个个体对应的医疗训练数据和所述变化值汇总为一条单元数据。
即将与某个人或某系列病例相关联的医疗训练数据和所述变化值汇总为一条单元数据。
四、将所述医疗训练数据采用图像分割、关联或文本分析方法整合或格式 化为计算机可以理解的结构化数据矩阵并从每个单元数据中提取数据特征。
所述数据特征包括医疗训练数据在时空上的变化值、数据本身的各种数理统计值。
数据特征包括医疗训练数据随时间的改变，如数据上升或下降的趋势；空间的变化，如一个图像数据从其中的一个像素到下一个像素之间的关系。数据特征还包括数据本身的各种数理统计值，如个体数据与其他个体数据对比值。这些数据特征将会以矢量、矩阵或数列的形式格式化为计算机理解的结构。数据特征的采集也包括图像处理或初期数据统计处理。在图像处理当中，分割出与医疗诊治数据有关的图像内容是寻找图像数据特征的第一步。在文档文件处理当中，TF-IDF(term frequency–inverse document frequency)，即一种量化资料检索和文本挖掘的方式，也可被应用。以上初期图像文本处理会大大方便计算机对数据特征的采集。
五、将已形成结构化数据矩阵的医疗训练数据导入设置于计算机内对应深度学习模型的存储模块中；
六、通过计算机对所述深度学习模型进行优化运算，优化方法如下：
1、设定深度学习基本框架，将所述医疗训练数据按照数据特征建立包括输入层、至少一层隐层和输出层的数据模型，输入层包含若干个具有数据特征的节点，输出层包含若干个具有医疗诊断数据特征的节点，每个隐层包含若干个与上一层输出值具有映射对应关系的节点；
2、每个节点采用数学方程建立该节点的数据模型，采用人工或随机方法预设所述数学方程中的相关参数值，输入层中各节点的输入值为所述的数据特征，各隐层及输出层中各节点的输入值为上层的输出值，每层中各节点的输出值为 本节点经所述数学方程运算后所得的值；
3、初始化所述参数值Ai，将所述输出层中各节点的输出值与对应节点的医疗诊断数据特征比对，反复修正各节点的所述参数值Ai，依次循环，最终获得使所述输出层中各节点的输出值生成与所述医疗诊断数据特征相似度为局部最大时的输出值对应的各节点中的参数值Ai；
对所述参数值Ai进行优化的方法为无监督学习方法和有监督学习方法。
所述无监督学习方法采用降噪自动编码生成器或限制伯尔曼机进行自学习。
所述数学方程为参数数学方程或非参数数学方程，其中，参数数学方程可为线性模型、神经元模型或卷积运算，非参数数学方程可为极值运算方程，数学模型设定方式如下：
y＝g(X)＝fn○fn-1○fn-2○…○f1(X)
其中y是所述输出层中的医疗诊断数据特征，维度为Mn，X是训练素材数据，维度为M0，f1到fn为设定的每一层运算方程，而每一层方程fi的维度为Mi－1→Mi,如第一层f1就是将维度为M0的X转换成维度为M1的输出Z1，而Z1则成为第二层方程f2的输入，以此类推，其中,每一层模型fi有与之相匹配的参数组Ai。
如逻辑方程的表现形式如下：
Figure PCTCN2016084000-appb-000002
又如线性方程的表现形式如下：
Figure PCTCN2016084000-appb-000003
其中，xm是方程的输入值，y则是方程的输出值，am则是方程的基础参数。
初始化深度学习模型参数A1至An，可随意设定模型参数,模型深度等，亦可以某种方式选择初始化参数模型。
运算方法解释如下
运算的核心为有监督学习算法深度学习，其是过去五到十年人工智能和机器学习领域革命性的技术。本发明的在DNN算法的基础上独创性地加入病变扫描时的空间时序变化，充分考虑到病变体在三维空间当中的成像规律，提高识别概率；同时模型可引入医生人为判断因素，融合纯智能判断因素与医生的专业判断进行综合建模计算病变概率。
本发明隶属于人工智能技术，数据运算的最终目的是要“训练”模型能够自动在医疗影像当中识别病变，给出概率并进行标示，辅助医生的诊疗工作。因此在模型构建的过程当中，海量数据相当于教学素材，而模型框架则是算法将具体信息进行抽象化总结的具体流程，因此在智能运算的过程当中，海量数据和智能算法必不可少。
下文将分别介绍：
1.医疗训练数据
a)客观存在的自然数据：即医院在实际业务当中所形成的自然诊疗数据，训练数据与实际运用场景所产生的医疗诊断数据越接近则训练效果越好。而数据作为教材也有相应要求。两种数据源(即医疗素材数据和医疗诊断数据)对于本专利的实现都是必不可少的：首先需原始医疗数据，比如一般从医疗器械成像之后的数据格式有很多，如.nii.gz，.dcm等格式，所有格式在进入训练 模型之前都可被表示为多维矩阵数据信息。比如MRI(英文全称：Magnetic Resonance Imaging，中文名为核磁共振成像)图形可为三维矩阵(灰阶)，即二维灰阶指数与一维横切面；或者四维矩阵(rgb)，即二维颜色指数外加三种颜色指数，最后加一维横切面。任何医疗数据均可以被抽象地简化成为此类矩阵。而此类矩阵构成模型读取的原始数据源。如图2、3所示，显示的是脑部MRI所生成的图像数据，假设系统将一张MRI切片生成为像素512×512的图，而一次脑部扫描为200张切面，则一次灰阶脑部扫描的数据源可被总结为512×512×200＝52,428,800维度的数据行。对于模型来说，52,428,800个数字当中涵盖了所有该脑部扫描的所有被理解和概括的信息。
对于无监督学习部分来说，此类医疗原始矩阵即建模的基础数据。对于有监督学习部分来说，还需要与图形相匹配的分析目标。最简单的二元分析信息如：测量病变体(稍复杂一些的信息可以是病变的概率)。之后可融入更为复杂的医学信息如：病变种类、治疗效果、病变的具体位置等信息。对于更为复杂的病变发展预测系统，可匹配病人过往体检的纵向时间序列数据，让算法学会预测医学现象的发展规律。
b)模拟形成的模拟数据：此类数据通过电脑自行加工或模拟而成，以模拟数据作为建模的训练数据。
此类模型最典型的例子即微软的Xbox Kinect系统，开发阶段手部姿势识别模型的基础数据全部有3D建模完成。
在本发明当中，模拟数据可被理解为基于原有医疗数据通过变形、扭曲和噪音叠加而构建出的新数据。使用模拟数据的原因有两层：第一，加入变形后数据有利于教会算法更加稳定地识别医疗数据当中的核心变化规律；第二，一 般的DNN模型需要求导几百万以上的参数，在数据量有限的情况下，很容易导致过拟合现象，即模型过度学习已有历史数据，而无法很好地对于核心变化规律进行归纳和抽象总结，加入模拟的变形数据相当于在训练过程当中加入了噪音，逼迫算法能够更好地区分噪音和有效信息，有助于解决过拟合问题。
2.机器学习算法模型
机器学习算法模型为本发明用以归纳和对信息进行抽象总结的基本数学框架，主要目的是将模式识别的过程以电脑可以被理解的数学结构表达出来。训练的过程则是估算模型当中的参数，参数估算完毕以后模型将成为本发明方法的核心部分。根据不同目的机器学习算法可以被归为有监督学习和无监督学习两种类别，本发明涵盖两种类别的算法。
a)有监督学习：有监督学习算法强调人为设立模型所寻找的目标规律。正如前篇所描述，除去原始图形矩阵数据以外，有监督学习算法还需要相匹配的分析结论数据(如所述的医疗诊断数据)。
本专利主要包括以下有监督学习算法
i.深度神经元算法(Deep Neural Network)
此算法的基本原理模仿人脑辨别过程。DNN算法的输入为原始的医疗数据和医生的历史分析结果，最终能自动完成分析过程。而DNN的抽象概括即为
f(x)＝y
x为原始医疗矩阵数据，y为智能系统分析结果，而DNN即是方程f，x至y的数学映射表达方式。
不同的有监督学习对于f有不同的假设，DNN算法模拟人脑的神经元结构进行建模，以图形数据为基础的DNN基本数学构造如图4所示。
【注：这只是我们上述DNN模型的其中一种做法而已，第一层为卷积层、第二层为最大池层以此循环】DNN的结构从左边的图形原始数据到最右边的分析结果可以被分为多层，每一层完成不同的数学运算。模型总共有多层神经元结构。第一层对于医疗数据完成多个并行内积运算。在第一层中最常用的算法是卷积，卷积算法通过在输入的一系列的值上滑动一个新的方程，从而输出这个新的方程与原系列数据的内积。如在3D医疗图形当中，算法构建多个卷积方块，每个卷积方块是一个三围的矩阵。卷积方块x与y轴涵盖了一个图像在空间变化的方程，而卷积方块的z轴则涵盖了图像在空间变化的方程。每个卷积方块矩阵随着数据维度本身滑动，计算3D图形各个维度的值与卷积方块的内积，内积运算所得数值可被理解为数据维度与卷积方块之间的相似度，而数据当中每一个部分输出的内积值将会成为下一层神经元的输入值。从直观形象的角度来说，并行的卷积矩阵相当于特定形状，而卷积矩阵外内计算相当于在数据当中判断不同区域是否与特定形态相近。
见图5所示，DNN模型第二个常用层是进行池化运算(英文名为Max Pooling，简称MP运算)的层次。MP运算过程将维度信息合成为更大范围的方块，每一个方块内进行Max(寻求最大值)运算。MP运算主要模仿视觉神经网络当中的神经元活跃经济特性。在一定范围内的信息框架内，只保留最活跃的信息单元进入下一层。从图形的角度来说，MP运算使运算结果不再因为数据本身的旋转而改变。从运算的角度来说，MP运算相当于降维处理，结合第一层神经元运算，MP运算去掉了与第一层卷积方块相似度比较低的区域信息，降低每个区域内无效信息的含量。
DNN的结构往往由卷积层和池化层两层结合并不断重复来提取与医疗诊断 相关的数据特征。直观说来，比如说在图4中，中间层算法使用了第一层与第二层所构成的非线性元素，以此类推。以此构建更加抽象的框架元素。通过本专利的方法，可以构建无数层神经元。
理论上来说，在模型能够得到大量的数据来进行训练的前提下，卷积层与池化层的组合重复越多越好，虽然人脑的运作机制还未被完全理解，但是已知道人脑属于非常深层次的神经元结构。因此越深层次的神经元模型功率越强。可是越深层次的神经元结构参数越多，也越难训练和估参，很容易出现导数消亡以及过拟合问题。
DNN在多次卷积运算和MP层运算之后，剩下的信息进入最后的完全互联多层感知器(Multi-Layer-Perceptron MLP)当中。该感知器的基本结构即为两层的逻辑回归运算，相当于为不同抽象图形元素对于最终评判结果的贡献进行附值，而MLP运算最终的输出值即为模型的医疗分析结果。如图6所示，感知器一般为一层隐性感知层，每单元的变量与上层所有变量完全互联，每一层进行逻辑运算，以求导下一层数值。
b)无监督学习：神经网络的概念很多年前就已经存在，可是受限于可以使用的数据量、处理器的运算能力，使得导数消亡问题非常严重，无法用来解决实际问题。在完整的模型训练过程当中，模型的输出预测值与实际值之间的误差构成了模型参数优化的基础，而过分深层的神经网络结构无法将参数优化信息反推到底层的网络当中，即表层信息无法一层层传递到深层网络结构当中，给模型训练带来了很大的困难。特别是在医疗数据分析领域当中，医疗数据体量往往非常大，要进行完整优化搜寻不现实，运算难度高的问题比起别的领域更加严重。本发明通过无监督学习对于模型中的参数进行初步优化，使得模型 优化的过程当中参数的初始条件变得非常有利，让模型优化过程更快能寻找到局部极小值。
其中效果最好的两种无监督学习方法为降噪自动编码生成器(Denoising Autoencoders，dAE)以及限制伯尔曼机(Restricted Boltzmann Machine，RBM)
i.降噪自动编码生成器(以下简称dAE)
自动编码生成器的原理为寻找某一数据变量的有效隐性变量。如图7所示，完整地呈现了自动编码生成器的工作原理，根据原始医疗数据输入x，自动编码生成器寻找隐性要素代表y与参数W来映射出新的数据z，而自动编码生成器最终的目的是要找到参数W来最小化z与x之间的差别，换句话说就是要找到在有限信息之内最能够完整代表数据变量信息的参数。而这些参数可被视为涵盖了模型范围内最大量的原始数据信息。而降噪自动编码生成器则在简单编码生成器的工作原理上人为引入大量噪音，直观来说是通过大量噪音来强迫模型寻找更有价值的潜在规律，而不受噪音当中无效规律的影响。最终训练出来的参数将会成为有监督学习的初步参数起始值，相当于为模型第一步寻找一个良好的起点，大大加快受训模型参数优化的速度。
ii.限制伯尔曼机(以下简称RBM)
限制伯尔曼机起到的作用与dAE非常类似，都为无监督学习当中的参数优化步骤做准备。与dAE非常类似，RBM亦是想要寻找最能体现原有数据变量的因素方向，来作为后期有监督学习的起始参数。但是与dAE不同，RBM使用的是能量概率生成式模型，并非如dAE一样使用确定性数学表达式。从运算求解的角度来说叫dAE稍更复杂。但因其概率生成模型的本质对于某些偏重事件概率性 的建模场景更加适用。对于我们的建模过程来说，为进一步降低过拟合现象的产生，我们RBM和dAE两种无监督学习都将在我们集成学习过程当中使用。
优化数据的进一步改进是建立损失方程或目标方程，根据损失方程进行有监督学习数据模型参数优化。如损失方程可以设定为训练数据当中深度学习模型输出的分析结果和实际目标变量之间的区别，运算中根据损失方程的变化、通过优化方法来调整模型内参数值。比如在医疗影像心血管横截面积测量这一应用案例当中，损失方程可以被设定为模型生成测量值和实际值之间的区别(如方差区别)，通过梯度降低(gradient descent)的参数优化方法在每一个循环当中移动参数值，在参数优化循环符合特定条件之后停止参数优化的过程(如设定循环与循环之前损失方程的值小于特定阀值、或循环数超过某数则停止优化运算)，保留最佳值。
目标方程的优化过程可以通过如下方式表达:
Figure PCTCN2016084000-appb-000004
其中g(x，A)是基本深度学习的分析输出,Y是分析目标的实际值,L[Y；g(X；A)]主要用以计算深度学习的分析输出和实际值之间的差别所造成的成本。
较常见的L[Y，g(X，A)]＝Σi|yi-gi(xi，A)|或者
L[Y，g(X，A)]＝Σi[yi-gi(xi，A)]2
R(A)主要为正则化表达式,主要功能用以避免模型的过度拟合。
最常见的正则化表达式有L1正则化：R(A)＝Σa∈A|a|和L2正则化：R(A)＝Σa∈Aa2，但使用者也可2以使用任意挑选的正则化表达式。
参数优化的方法可以任意选择,最常见的为梯度降低法,步骤数学表达式如下：
Figure PCTCN2016084000-appb-000005
在j轮将参数往目标方程的微分反方向移动。反复多轮移动之后,在满足特定停止条件的情况下停止移动。
优化数据的又一步的改进是数据增噪方法，可以人为地为模型和数据增加噪音，来起到稳定模型和抗击数据模型的过度拟合问题。比如在医疗影像智能分析仪当中，可以对原始数据进行变形和扭曲，强迫模型识别噪音以外的有效信息。
优化数据的再一改进是测试样本分割，可以将训练数据进一步分割出测试样本，利用剩余的训练数据建立模型，通过测试样本来测试模型的有效性，可以根据结果自动或者手动调整深度学习模型的核心框架。
七、将获取的已形成结构化矩阵数据的医学待分析数据导入该深度学习模型中进行与之匹配的医学病理分析；
八、由该深度学习模型通过输出装置输出与所述医学待分析数据相匹配的医学病理分析结果。
九、将所述医学待分析数据和与之匹配的分析结果涉及的结构化数据反馈到所述深度学习模型中形成新的训练数据对该深度学习模型进一步优化。
本发明的基于深度学习对医疗数据进行分析的智能分析仪，其包括可将医疗训练数据和医学待分析数据导入计算机中的输入装置、分别或集中保存所述医疗训练数据和医学待分析数据的存储模块、调用存储模块中的医疗训练数据进行自学习的本发明方法中的深度学习模型模块、将与所述医学待分析数据匹配的医学病理分析结果导出的输出装置和包括CPU和/或GPU的处理器，其中，
所述医疗训练数据包括医疗素材数据和与之匹配的医疗诊断数据；
所述医疗训练数据和医学待分析数据为计算机可以理解的结构化数据矩阵；
所述自学习采用包括线性模型、神经元模型、卷积运算和/或寻求最大值运算的参数数学方程；
所述输入装置包括设置在医院、医学机构的计算机装置、与该计算机联网的各种医疗检查装置和病理分析装置；如电脑、彩超仪器、X光、同步心电仪、生化分析仪、免疫分析仪、纤维内窥镜、核磁共振、CT多普勒诊断仪、血压计、体重计等等
所述输出装置包括设置在医院、医学机构中并与所述输入装置联网的固定式计算机输出终端和移动式智能终端。如电脑、医疗器械终端、手机终端等等。
为了实现医院或医学机构共享资源，可在本发明的智能分析仪上安装可与互联网、以太网连接的包括光纤连接、WIFI连接或GPRS模块连接的网络连接模块。
本发明的智能分析仪将训练所得深度学习模型应用到实际当中，为完整集成系统。新的医疗数据(即所述的待分析医疗数据)产生之后与模型参数相结合，将得出分析预测值。实际应用过程当中，智能分析仪在分析流程当中成为一个额外插件。根据应用场景的不同，可以成为医疗器械当中的额外插件，也可以成为常用PACS(中文名为影像归档和通信系统)系统或者HIS(中文名为医院信息)系统当中的插入界面，也可以通过互联网接口将其生成的分析报告导入其它系统中。

Claims (10)

  1. 一种基于深度学习对医疗数据进行分析的方法，其特征在于：其包括如下步骤：
    1)采集海量已备案的同类型的医疗素材数据及与该医疗素材数据匹配的医疗诊断数据作为医疗训练数据通过输入装置存储于计算机中；
    2)将所述医疗训练数据中不小于二维的影像数据与文本数据中随时间和空间的变化值与对应的数据相关联；
    3)在采集的海量医疗训练数据中，将与每一个个体对应的医疗训练数据和所述变化值汇总为一条单元数据；
    4)将所述医疗训练数据采用分割、关联或文本数据挖掘方法整合或格式化为计算机可以理解的结构化数据矩阵并从每个单元数据中提取数据特征；
    5)将已形成结构化数据矩阵的医疗训练数据导入设置于计算机内对应深度学习模型的存储模块中；
    6)通过计算机对所述深度学习模型进行优化运算，优化方法如下：
    a.设定深度学习基本框架，将所述医疗训练数据按照数据特征建立包括输入层、至少一层隐层和输出层的数据模型，输入层包含若干个具有数据特征的节点，输出层包含若干个具有医疗诊断数据特征的节点，每个隐层包含若干个与上一层输出值具有映射对应关系的节点；
    b.每个节点采用数学方程建立该节点的数据模型，采用人工或随机方法预设所述数学方程中的相关参数值，输入层中各节点的输入值为所述的数据特征，各隐层及输出层中各节点的输入值为上层的输出值，每层中各节点的输出值为 本节点经所述数学方程运算后所得的值；
    c.初始化所述参数值Ai，将所述输出层中各节点的输出值与对应节点的医疗诊断数据特征比对，反复修正各节点的所述参数值Ai，依次循环，最终获得使所述输出层中各节点的输出值生成与所述医疗诊断数据特征相似度为局部最大时的输出值对应的各节点中的参数值Ai；
    7)将获取的已形成结构化矩阵数据的医学待分析数据导入该深度学习模型中进行与之匹配的医学病理分析；
    8)由该深度学习模型通过输出装置输出与所述医学待分析数据相匹配的医学病理分析结果。
  2. 根据权利要求1所述的方法，其特征在于：对所述参数值Ai进行优化的方法为无监督学习方法。
  3. 根据权利要求2所述的方法，其特征在于：所述无监督学习方法采用降噪自动编码生成器或限制伯尔曼机进行自学习。
  4. 根据权利要求1所述的方法，其特征在于：对所述参数值Ai进行优化的方法为有监督学习方法。
  5. 根据权利要求1－4中任一项所述的方法，其特征在于：所述数学方程为参数数学方程或非参数数学方程，其中，参数数学方程可为线性模型、神经元模型或卷积运算，非参数数学方程可为极值运算方程，数学模型设定方式如下：
    y＝g(X)＝fn○fn-1○fn-2○…○f1(X)
    其中y是所述输出层中的医疗诊断数据特征，维度为Mn，X是训练素材数据，维度为M0，f1到fn为设定的每一层运算方程，而每一层方程fi的维度为 Mi－1→Mi,如第一层f1就是将维度为M0的X转换成维度为M1的输出Z1，而Z1则成为第二层方程f2的输入，以此类推，其中,每一层模型fi有与之相匹配的参数组Ai。
  6. 根据权利要求1所述的方法，其特征在于：所述医疗素材数据包括临床和医技阶段医生对患者诊断、检查和治疗过程进行的相关信息记录；所述诊断数据包括临床和医技阶段医生对患者初诊判断、出院结果、疾病治疗效果进行的相关信息记录以及医生撰写的文本出诊数据和跟踪随访数据。
  7. 根据权利要求1所述的方法，其特征在于：所述数据特征包括医疗训练数据在时空上的变化值、数据本身的各种数理统计值。比如说随着时间的改变，数据上升或下降的趋势。
  8. 根据权利要求1所述的方法，其特征在于：将所述医学待分析数据和与之匹配的分析结果涉及的结构化数据反馈到所述深度学习模型中形成新的训练数据。
  9. 一种基于深度学习对医疗数据进行分析的智能分析仪，其特征在于：其包括可将医疗训练数据和医学待分析数据导入计算机中的输入装置、分别或集中保存所述医疗训练数据和医学待分析数据的存储模块、调用存储模块中的医疗训练数据进行自学习的深度学习模型模块、将与所述医学待分析数据匹配的医学病理分析结果导出的输出装置和包括CPU和/或GPU的处理器，其中，
    所述医疗训练数据包括医疗素材数据和与之匹配的医疗诊断数据；
    所述医疗训练数据和医学待分析数据为计算机可以理解的结构化数据矩阵；
    所述自学习采用包括线性模型、神经元模型、卷积运算和/或寻求最大值运 算的参数数学方程；
    所述输入装置包括设置在医院、医学机构的计算机装置、与该计算机联网的各种医疗检查装置和病理分析装置；
    所述输出装置包括设置在医院、医学机构中并与所述输入装置联网的固定式计算机输出终端和移动式智能终端。
  10. 根据权利要求9所述的基于深度学习对医疗数据进行分析的智能分析仪，其特征在于：其还设有可与互联网、以太网连接的包括光纤连接、WIFI连接或GPRS模块连接的网络连接模块。
PCT/CN2016/084000 2015-06-02 2016-05-31 基于深度学习对医疗数据进行分析的方法及其智能分析仪 WO2016192612A1 (zh)

Priority Applications (5)

Application Number Priority Date Filing Date Title
EP16802535.1A EP3306500A4 (en) 2015-06-02 2016-05-31 MEDICAL PROCESSING DATA ANALYSIS METHOD BASED ON DEEP LEARNING, AND INTELLIGENT ANALYZER THEREFOR
JP2017559611A JP6522161B2 (ja) 2015-06-02 2016-05-31 ディープラーニングに基づく医療データ分析方法及びそのインテリジェントアナライザー
US15/579,212 US11200982B2 (en) 2015-06-02 2016-05-31 Method for analysing medical treatment data based on deep learning and intelligence analyser thereof
IL255856A IL255856B (en) 2015-06-02 2017-11-22 A method and analysis for analyzing a database of medical treatment using deep learning
US17/519,873 US20220059229A1 (en) 2015-06-02 2021-11-05 Method and apparatus for analyzing medical treatment data based on deep learning

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510294286.3 2015-06-02
CN201510294286.3A CN104866727A (zh) 2015-06-02 2015-06-02 基于深度学习对医疗数据进行分析的方法及其智能分析仪

Related Child Applications (2)

Application Number Title Priority Date Filing Date
US15/579,212 A-371-Of-International US11200982B2 (en) 2015-06-02 2016-05-31 Method for analysing medical treatment data based on deep learning and intelligence analyser thereof
US17/519,873 Continuation US20220059229A1 (en) 2015-06-02 2021-11-05 Method and apparatus for analyzing medical treatment data based on deep learning

Publications (1)

Publication Number Publication Date
WO2016192612A1 true WO2016192612A1 (zh) 2016-12-08

Family

ID=53912551

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/084000 WO2016192612A1 (zh) 2015-06-02 2016-05-31 基于深度学习对医疗数据进行分析的方法及其智能分析仪

Country Status (6)

Country Link
US (2) US11200982B2 (zh)
EP (1) EP3306500A4 (zh)
JP (1) JP6522161B2 (zh)
CN (1) CN104866727A (zh)
IL (1) IL255856B (zh)
WO (1) WO2016192612A1 (zh)

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2019028072A (ja) * 2017-07-25 2019-02-21 清華大学Ｔｓｉｎｇｈｕａ Ｕｎｉｖｅｒｓｉｔｙ Ｃｔ画像を再構成する方法及びデバイス、並びに記憶媒体
WO2019051007A1 (en) * 2017-09-07 2019-03-14 Butterfly Network, Inc. ULTRASONIC DEVICE ON A CHIP RELATED TO THE WRIST
CN109671499A (zh) * 2018-10-22 2019-04-23 南方医科大学 一种直肠毒性预测系统构建方法
WO2019080502A1 (zh) * 2017-10-23 2019-05-02 平安科技（深圳）有限公司 利用语音进行疾病预测的方法、应用服务器和计算机可读存储介质
CN109816140A (zh) * 2018-12-12 2019-05-28 哈尔滨工业大学（深圳） 基于跨市场影响的股价预测方法、装置、设备及存储介质
CN110164524A (zh) * 2019-04-29 2019-08-23 北京国润健康医学投资有限公司 一种偏瘫患者康复训练任务自适应匹配方法及其系统
JP2020013529A (ja) * 2018-07-12 2020-01-23 國立臺灣科技大學 機械学習を適用した医療画像分析方法及びそのシステム
CN111180076A (zh) * 2018-11-13 2020-05-19 零氪科技（北京）有限公司 一种基于多层语义分析的医疗信息提取方法
CN111192680A (zh) * 2019-12-25 2020-05-22 山东众阳健康科技集团有限公司 一种基于深度学习和集成分类的智能辅助诊断方法
CN111341435A (zh) * 2019-07-01 2020-06-26 郑州大学第一附属医院 一种基于分布式深度学习的智能病理诊断方法
CN111370098A (zh) * 2020-04-27 2020-07-03 北京贝叶科技有限公司 一种基于边缘侧计算和服务装置的病理诊断系统及方法
CN111582493A (zh) * 2020-04-15 2020-08-25 马鞍山师范高等专科学校 一种深度学习模型训练的管理系统
CN111696674A (zh) * 2020-06-12 2020-09-22 电子科技大学 一种电子病历的深度学习方法及系统
US11468998B2 (en) * 2018-10-09 2022-10-11 Radect Inc. Methods and systems for software clinical guidance

Families Citing this family (94)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104866727A (zh) 2015-06-02 2015-08-26 陈宽 基于深度学习对医疗数据进行分析的方法及其智能分析仪
US10387464B2 (en) * 2015-08-25 2019-08-20 Facebook, Inc. Predicting labels using a deep-learning model
CN105117611B (zh) * 2015-09-23 2018-06-12 北京科技大学 基于卷积神经元网络的中医舌诊模型的确定方法及系统
JP2019513315A (ja) * 2016-02-29 2019-05-23 ウルグス ソシエダード アノニマ 惑星規模解析のためのシステム
CN105808712A (zh) * 2016-03-07 2016-07-27 陈宽 将文本类医疗报告转换为结构化数据的智能系统及方法
US20170262583A1 (en) 2016-03-11 2017-09-14 International Business Machines Corporation Image processing and text analysis to determine medical condition
CN106264460B (zh) * 2016-07-29 2019-11-19 北京医拍智能科技有限公司 基于自学习的脑活动多维时间序列信号的解码方法及装置
CN106156530A (zh) * 2016-08-03 2016-11-23 北京好运到信息科技有限公司 基于栈式自编码器的体检数据分析方法及装置
CN107025369B (zh) * 2016-08-03 2020-03-10 北京推想科技有限公司 一种对医疗图像进行转换学习的方法和装置
CN106326640A (zh) * 2016-08-12 2017-01-11 上海交通大学医学院附属瑞金医院卢湾分院 一种医疗语音控制系统及其控制方法
CN106599530B (zh) * 2016-10-31 2019-08-02 北京千安哲信息技术有限公司 一种检测数据的处理方法和装置
CN107358014B (zh) * 2016-11-02 2021-01-26 华南师范大学 一种生理数据的临床前处理方法及系统
CN108198625B (zh) * 2016-12-08 2021-07-20 推想医疗科技股份有限公司 一种分析高维医疗数据的深度学习方法和装置
CN106777953A (zh) * 2016-12-09 2017-05-31 江西中科九峰智慧医疗科技有限公司 医学影像数据的分析方法及系统
CN106934216B (zh) * 2017-02-16 2021-09-14 山东大学齐鲁医院 基于多目标的医疗器械临床评价方法
CN106934235B (zh) * 2017-03-09 2019-06-11 中国科学院软件研究所 一种基于迁移学习的疾病领域间病人相似性度量迁移系统
CN107833605A (zh) * 2017-03-14 2018-03-23 北京大瑞集思技术有限公司 一种医院病历信息的编码方法、装置、服务器及系统
US10261903B2 (en) 2017-04-17 2019-04-16 Intel Corporation Extend GPU/CPU coherency to multi-GPU cores
CN114796891A (zh) * 2017-06-05 2022-07-29 西安大医集团股份有限公司 放疗系统
CN107239767A (zh) * 2017-06-08 2017-10-10 北京纽伦智能科技有限公司 小鼠行为识别方法及其系统
EP3270308B9 (en) 2017-06-14 2022-05-18 Siemens Healthcare GmbH Method for providing a secondary parameter, decision support system, computer-readable medium and computer program product
WO2019022779A1 (en) * 2017-07-28 2019-01-31 Google Llc SYSTEM AND METHOD FOR PREDICTING AND SUMMING MEDICAL EVENTS FROM ELECTRONIC HEALTH RECORDINGS
CN107616796B (zh) * 2017-08-31 2020-09-11 北京医拍智能科技有限公司 基于深度神经网络的肺结节良恶性检测方法及装置
CN107767935A (zh) * 2017-09-15 2018-03-06 深圳市前海安测信息技术有限公司 基于人工智能的医学影像分类处理系统及方法
CN109947782A (zh) * 2017-11-03 2019-06-28 中国移动通信有限公司研究院 一种大数据实时应用系统的更新方法、装置及系统
TWI649698B (zh) * 2017-12-21 2019-02-01 財團法人工業技術研究院 物件偵測裝置、物件偵測方法及電腦可讀取媒體
JP7006296B2 (ja) * 2018-01-19 2022-01-24 富士通株式会社 学習プログラム、学習方法および学習装置
CN108538390A (zh) * 2018-04-28 2018-09-14 中南大学 一种面向医学数据的增量式处理方法
CN112055878B (zh) * 2018-04-30 2024-04-02 皇家飞利浦有限公司 基于第二组训练数据调整机器学习模型
US20190370956A1 (en) 2018-05-30 2019-12-05 General Electric Company Contrast imaging system and method
WO2019235335A1 (ja) * 2018-06-05 2019-12-12 住友化学株式会社 診断支援システム、診断支援方法及び診断支援プログラム
KR102095959B1 (ko) * 2018-07-11 2020-04-01 주식회사 아이센스 인공지능 딥러닝 학습을 이용한 생체측정물 농도 측정방법
CN108986913A (zh) * 2018-07-13 2018-12-11 希蓝科技（北京）有限公司 一种优化人工智能心电诊断方法及系统
US11756691B2 (en) 2018-08-01 2023-09-12 Martin Reimann Brain health comparison system
US12008462B2 (en) 2018-08-09 2024-06-11 Board Of Trustees Of Michigan State University Systems and methods for providing flexible, multi-capacity models for use of deep neural networks in mobile devices
CN110827988B (zh) * 2018-08-14 2022-10-21 上海明品医学数据科技有限公司 一种基于移动终端进行医学数据研究的控制方法
TWI701680B (zh) * 2018-08-19 2020-08-11 長庚醫療財團法人林口長庚紀念醫院 分析醫學影像之方法及系統
CN109191446B (zh) * 2018-08-30 2020-12-29 杭州深睿博联科技有限公司 用于肺结节分割的图像处理方法及装置
CA3111650A1 (en) * 2018-09-07 2020-03-12 Minas Chrysopoulo System to provide shared decision making for patient treatment options
KR102150673B1 (ko) * 2018-10-02 2020-09-01 (주)지엘테크 외관불량 검사방법 및 외관불량 검사 시스템
KR102180135B1 (ko) * 2018-10-12 2020-11-17 계명대학교 산학협력단 심혈관 질환 종류에 따른 심전도 패턴 시뮬레이션 생체신호 구현 시스템 및 방법
CN111091914B (zh) * 2018-10-23 2023-11-21 百度在线网络技术（北京）有限公司 基于病历的癌症分型分期方法及其装置
CN109378064B (zh) * 2018-10-29 2021-02-02 南京医基云医疗数据研究院有限公司 医疗数据处理方法、装置电子设备及计算机可读介质
US11507822B2 (en) * 2018-10-31 2022-11-22 General Electric Company Scalable artificial intelligence model generation systems and methods for healthcare
CN109247923B (zh) * 2018-11-15 2020-12-15 中国科学院自动化研究所 基于视频的非接触式脉搏实时估计方法及设备
CN109583358A (zh) * 2018-11-26 2019-04-05 广东智源信息技术有限公司 一种医疗卫生监督快速精准执法方法
CN111261278A (zh) * 2018-11-30 2020-06-09 上海图灵医疗科技有限公司 一种基于三维图像的深度学习模型的心脏疾病检测方法
US10957442B2 (en) * 2018-12-31 2021-03-23 GE Precision Healthcare, LLC Facilitating artificial intelligence integration into systems using a distributed learning platform
KR102190756B1 (ko) * 2018-12-31 2020-12-14 월드버텍 주식회사 딥러닝 기반의 식물 인식 시스템
US11687691B2 (en) * 2019-01-03 2023-06-27 International Business Machines Corporation Forward and reverse transformations of a model of functional units of a biological system trained by a global model of the systems
CN109800294B (zh) * 2019-01-08 2020-10-13 中国科学院自动化研究所 基于物理环境博弈的自主进化智能对话方法、系统、装置
CN110119432B (zh) * 2019-03-29 2023-05-05 中国人民解放军总医院 一种用于医疗平台的数据处理方法
CN110121056B (zh) * 2019-04-03 2022-01-14 视联动力信息技术股份有限公司 跨区域视联网监控视频获取方法及装置
CN110085314A (zh) * 2019-04-11 2019-08-02 上海翼依信息技术有限公司 医学检验数据的智能分析方法、系统以及设备
CN110033863B (zh) * 2019-04-23 2021-06-04 科大讯飞股份有限公司 应用于临床决策支持系统的不合理疾病诊断检测方法、装置
US20200342968A1 (en) * 2019-04-24 2020-10-29 GE Precision Healthcare LLC Visualization of medical device event processing
US11416653B2 (en) * 2019-05-15 2022-08-16 The Mitre Corporation Numerical model of the human head
CN110322959B (zh) * 2019-05-24 2021-09-28 山东大学 一种基于知识的深度医疗问题路由方法及系统
US11449793B2 (en) * 2019-07-03 2022-09-20 Kpn Innovations, Llc. Methods and systems for medical record searching with transmittable machine learning
US11443137B2 (en) 2019-07-31 2022-09-13 Rohde & Schwarz Gmbh & Co. Kg Method and apparatus for detecting signal features
CN110706803B (zh) * 2019-08-20 2023-06-27 南京医基云医疗数据研究院有限公司 一种确定心肌纤维化的方法、装置、可读介质及电子设备
US11783189B2 (en) * 2019-08-29 2023-10-10 Nec Corporation Adversarial cooperative imitation learning for dynamic treatment
JP7158355B2 (ja) * 2019-08-30 2022-10-21 株式会社豊田中央研究所 ノイズ除去装置および距離測定装置
CN110808095B (zh) * 2019-09-18 2023-08-04 平安科技（深圳）有限公司 诊断结果识别、模型训练的方法、计算机设备及存储介质
US10936962B1 (en) 2019-11-01 2021-03-02 Kenneth Neumann Methods and systems for confirming an advisory interaction with an artificial intelligence platform
KR102446775B1 (ko) 2019-11-06 2022-09-22 주식회사 하나금융티아이 컨벌루션 신경망의 입력 데이터 생성 장치 및 방법 그리고 이를 이용한 인공지능 기반 펀드가격 예측 장치 및 방법
CN110879980B (zh) * 2019-11-13 2023-09-05 厦门大学 基于神经网络算法的核磁共振波谱去噪方法
CN110890157A (zh) * 2019-11-18 2020-03-17 京东方科技集团股份有限公司 一种分析模型建立方法、辅助诊断装置及设备、介质
US11901056B2 (en) 2019-11-30 2024-02-13 Kpn Innovations, Llc Methods and systems for informed selection of prescriptive therapies
US10734105B1 (en) 2019-11-30 2020-08-04 Kpn Innovations, Llc Methods and systems for informed selection of prescriptive therapies
CN110767312A (zh) * 2019-12-26 2020-02-07 杭州迪英加科技有限公司 人工智能辅助病理诊断系统及方法
WO2021146748A1 (en) * 2020-01-17 2021-07-22 Acucela Inc. Database of retinal physiology derived from ophthalmic measurements performed by patients
CN113223693B (zh) * 2020-01-20 2023-11-10 深圳市理邦精密仪器股份有限公司 心电图机在线交互方法、心电图机、存储介质
KR102322616B1 (ko) * 2020-02-18 2021-11-05 서울대학교 산학협력단 Mri 스캐너 고유의 장치특성정보를 이용하는 변환 네트워크부를 포함하는 mri 데이터 변환장치 및 이를 이용한 mri 데이터 변환방법
JP6885517B1 (ja) * 2020-03-17 2021-06-16 株式会社村田製作所 診断支援装置及びモデル生成装置
KR102321157B1 (ko) * 2020-04-10 2021-11-04 (주)휴톰 수술 후 수술과정 분석 방법 및 시스템
CN111696660B (zh) * 2020-05-13 2023-07-25 平安科技（深圳）有限公司 基于人工智能的患者分群方法、装置、设备及存储介质
CN111789675B (zh) * 2020-06-29 2022-02-22 首都医科大学附属北京天坛医院 颅内血肿手术定位辅助方法及装置
CN111883247B (zh) * 2020-07-29 2022-03-15 复旦大学 一种行为数据与医疗结局相关性的分析系统
US20220059200A1 (en) * 2020-08-21 2022-02-24 Washington University Deep-learning systems and methods for medical report generation and anomaly detection
CN112231306B (zh) * 2020-08-23 2021-05-28 山东翰林科技有限公司 基于大数据的能源数据分析系统及方法
CN112070731B (zh) * 2020-08-27 2021-05-11 佛山读图科技有限公司 应用人工智能引导人体模型图集与病例ct图像配准的方法
CN112053785A (zh) * 2020-09-02 2020-12-08 北京小白世纪网络科技有限公司 基于冠心病诊断神经网络模型冠心病诊断方法及系统
CN112022124A (zh) * 2020-09-14 2020-12-04 山东省创新设计研究院有限公司 生理监测方法、装置、计算机设备和存储介质
CN112164448B (zh) * 2020-09-25 2021-06-22 上海市胸科医院 免疫治疗疗效预测模型训练方法、预测系统及方法和介质
CN112164446B (zh) * 2020-10-13 2022-04-22 电子科技大学 一种基于多网络融合的医疗影像报告生成方法
KR102272545B1 (ko) * 2020-11-26 2021-07-05 웰트 주식회사 건강 진단을 위한 사용자 장치 제어 방법 및 이러한 방법을 수행하는 장치
CN112786190B (zh) * 2021-01-14 2024-02-13 金陵科技学院 一种多维数据融合的医疗健康诊疗方法
CN112869691B (zh) * 2021-02-03 2021-11-02 清华大学 双波长增强型拉曼内窥式无创病理检测装置及检测方法
US11275903B1 (en) * 2021-05-13 2022-03-15 Retain Health, Inc System and method for text-based conversation with a user, using machine learning
CN114582494B (zh) * 2022-03-03 2022-11-15 数坤(北京)网络科技股份有限公司 诊断结果分析方法、装置、存储介质及电子设备
CN114373511B (zh) * 2022-03-15 2022-08-30 南方医科大学南方医院 基于5hmC分子标志物检测的肠癌模型及肠癌模型构建方法
CN115115620B (zh) * 2022-08-23 2022-12-13 安徽中医药大学 一种基于深度学习的肺炎病变模拟方法及系统
KR20240041722A (ko) * 2022-09-23 2024-04-01 삼성전자주식회사 시뮬레이션된 생체 신호를 출력하는 디바이스 및 그 동작 방법

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103544392A (zh) * 2013-10-23 2014-01-29 电子科技大学 基于深度学习的医学气体识别方法
CN103914841A (zh) * 2014-04-03 2014-07-09 深圳大学 基于超像素和深度学习的细菌分割与分类方法及其应用
CN104523266A (zh) * 2015-01-07 2015-04-22 河北大学 一种心电信号自动分类方法
CN104866727A (zh) * 2015-06-02 2015-08-26 陈宽 基于深度学习对医疗数据进行分析的方法及其智能分析仪

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE69610926T2 (de) * 1995-07-25 2001-06-21 Horus Therapeutics Inc Rechnergestütztes verfahren und anordnung zur diagnose von krankheiten
US5687716A (en) * 1995-11-15 1997-11-18 Kaufmann; Peter Selective differentiating diagnostic process based on broad data bases
US7640051B2 (en) * 2003-06-25 2009-12-29 Siemens Medical Solutions Usa, Inc. Systems and methods for automated diagnosis and decision support for breast imaging
US20070027368A1 (en) * 2005-07-14 2007-02-01 Collins John P 3D anatomical visualization of physiological signals for online monitoring
JP4985480B2 (ja) * 2008-03-05 2012-07-25 国立大学法人山口大学 がん細胞を分類する方法、がん細胞を分類するための装置及びがん細胞を分類するためのプログラム
CN102542562A (zh) 2011-11-23 2012-07-04 首都医科大学 一种基于肺结节三正交位ct图像纹理的提取方法和预测肺癌方法
CN203122364U (zh) * 2012-11-14 2013-08-14 华北电力大学 一种基于神经网络技术的病情自诊断设备
US20140279754A1 (en) * 2013-03-15 2014-09-18 The Cleveland Clinic Foundation Self-evolving predictive model
US20170124269A1 (en) * 2013-08-12 2017-05-04 Cerner Innovation, Inc. Determining new knowledge for clinical decision support
US9700219B2 (en) 2013-10-17 2017-07-11 Siemens Healthcare Gmbh Method and system for machine learning based assessment of fractional flow reserve
WO2015069827A2 (en) * 2013-11-06 2015-05-14 H. Lee Moffitt Cancer Center And Research Institute, Inc. Pathology case review, analysis and prediction
CN104298651B (zh) * 2014-09-09 2017-02-22 大连理工大学 一种基于深度学习的生物医学命名实体识别和蛋白质交互关系抽取在线方法
US9687199B2 (en) * 2014-09-15 2017-06-27 Wisconsin Alumni Research Foundation Medical imaging system providing disease prognosis

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103544392A (zh) * 2013-10-23 2014-01-29 电子科技大学 基于深度学习的医学气体识别方法
CN103914841A (zh) * 2014-04-03 2014-07-09 深圳大学 基于超像素和深度学习的细菌分割与分类方法及其应用
CN104523266A (zh) * 2015-01-07 2015-04-22 河北大学 一种心电信号自动分类方法
CN104866727A (zh) * 2015-06-02 2015-08-26 陈宽 基于深度学习对医疗数据进行分析的方法及其智能分析仪

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3306500A4 *

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP7064986B2 (ja) 2017-07-25 2022-05-11 清華大学 Ｃｔ画像を再構成する方法及びデバイス、並びに記憶媒体
JP2019028072A (ja) * 2017-07-25 2019-02-21 清華大学Ｔｓｉｎｇｈｕａ Ｕｎｉｖｅｒｓｉｔｙ Ｃｔ画像を再構成する方法及びデバイス、並びに記憶媒体
WO2019051007A1 (en) * 2017-09-07 2019-03-14 Butterfly Network, Inc. ULTRASONIC DEVICE ON A CHIP RELATED TO THE WRIST
WO2019080502A1 (zh) * 2017-10-23 2019-05-02 平安科技（深圳）有限公司 利用语音进行疾病预测的方法、应用服务器和计算机可读存储介质
JP2020013529A (ja) * 2018-07-12 2020-01-23 國立臺灣科技大學 機械学習を適用した医療画像分析方法及びそのシステム
US11468998B2 (en) * 2018-10-09 2022-10-11 Radect Inc. Methods and systems for software clinical guidance
CN109671499A (zh) * 2018-10-22 2019-04-23 南方医科大学 一种直肠毒性预测系统构建方法
CN109671499B (zh) * 2018-10-22 2023-06-13 南方医科大学 一种直肠毒性预测系统构建方法
CN111180076A (zh) * 2018-11-13 2020-05-19 零氪科技（北京）有限公司 一种基于多层语义分析的医疗信息提取方法
CN111180076B (zh) * 2018-11-13 2023-09-05 零氪科技（北京）有限公司 一种基于多层语义分析的医疗信息提取方法
CN109816140A (zh) * 2018-12-12 2019-05-28 哈尔滨工业大学（深圳） 基于跨市场影响的股价预测方法、装置、设备及存储介质
CN110164524A (zh) * 2019-04-29 2019-08-23 北京国润健康医学投资有限公司 一种偏瘫患者康复训练任务自适应匹配方法及其系统
CN111341435B (zh) * 2019-07-01 2022-11-08 郑州大学第一附属医院 一种用于疾病诊断的用户IoT设备
CN111341435A (zh) * 2019-07-01 2020-06-26 郑州大学第一附属医院 一种基于分布式深度学习的智能病理诊断方法
CN111192680A (zh) * 2019-12-25 2020-05-22 山东众阳健康科技集团有限公司 一种基于深度学习和集成分类的智能辅助诊断方法
CN111582493A (zh) * 2020-04-15 2020-08-25 马鞍山师范高等专科学校 一种深度学习模型训练的管理系统
CN111582493B (zh) * 2020-04-15 2023-07-28 马鞍山师范高等专科学校 一种深度学习模型训练的管理系统
CN111370098A (zh) * 2020-04-27 2020-07-03 北京贝叶科技有限公司 一种基于边缘侧计算和服务装置的病理诊断系统及方法
CN111696674A (zh) * 2020-06-12 2020-09-22 电子科技大学 一种电子病历的深度学习方法及系统
CN111696674B (zh) * 2020-06-12 2023-09-08 电子科技大学 一种电子病历的深度学习方法及系统

Also Published As

Publication number Publication date
JP2018529134A (ja) 2018-10-04
EP3306500A4 (en) 2019-01-23
US20220059229A1 (en) 2022-02-24
IL255856B (en) 2022-01-01
CN104866727A (zh) 2015-08-26
IL255856A (en) 2018-01-31
EP3306500A1 (en) 2018-04-11
US20180137941A1 (en) 2018-05-17
CN113421652A (zh) 2021-09-21
US11200982B2 (en) 2021-12-14
JP6522161B2 (ja) 2019-05-29

Similar Documents

Publication Publication Date Title
WO2016192612A1 (zh) 基于深度学习对医疗数据进行分析的方法及其智能分析仪
Suganyadevi et al. A review on deep learning in medical image analysis
US10706333B2 (en) Medical image analysis method, medical image analysis system and storage medium
CN113040715B (zh) 一种基于卷积神经网络的人脑功能网络分类方法
CN109544518B (zh) 一种应用于骨骼成熟度评估的方法及其系统
Sun et al. Intelligent analysis of medical big data based on deep learning
CN109659033A (zh) 一种基于循环神经网络的慢性疾病病情变化事件预测装置
CN112489769A (zh) 基于深度神经网络的慢性病智慧中医诊断与药物推荐系统
Feng et al. A review of methods for classification and recognition of ASD using fMRI data
Kazemi et al. Classifying tumor brain images using parallel deep learning algorithms
CN117457192A (zh) 智能远程诊断方法及系统
Ju et al. 3D-CNN-SPP: A patient risk prediction system from electronic health records via 3D CNN and spatial pyramid pooling
CN114191665A (zh) 机械通气过程中人机异步现象的分类方法和分类装置
Saputra et al. Implementation of machine learning and deep learning models based on structural MRI for identification autism spectrum disorder
Ahmad et al. Autism spectrum disorder detection using facial images: A performance comparison of pretrained convolutional neural networks
CN109360658A (zh) 一种基于词向量模型的疾病模式挖掘方法及装置
CN112712895A (zh) 一种针对2型糖尿病并发症的多模态大数据的数据分析方法
CN116759076A (zh) 一种基于医疗影像的无监督疾病诊断方法及系统
Bhardwaj et al. Improved healthcare monitoring of coronary heart disease patients in time-series fashion using deep learning model
Kumar et al. Identification and classification of pneumonia using cnn model with chest x-ray image
CN114822734A (zh) 基于循环卷积神经网络的中医病案分析方法
CN113421652B (zh) 对医疗数据进行分析的方法、训练模型的方法及分析仪
Srirangam Artificial Intelligence Based Hyper-parameter Adjustment on Deep Neural Networks: An Application of Detection and Classification of COVID-19 Diseases
Chandrasekar et al. ANALYSIS FOR HEART DISEASE PREDICTION USING DEEP NEURAL NETWORK AND VGG_19 CONVOLUTION NEURAL NETWORK.
Thamizhvani et al. Deep learning interpretation of biomedical data

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: 16802535

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2017559611

Country of ref document: JP

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 255856

Country of ref document: IL

WWE Wipo information: entry into national phase

Ref document number: 15579212

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE