KR102632773B1 - System for split training of medical data, method therefor, and computer readable medium for performing the method - Google Patents

Abstract

A central server requesting the medical data with the disease name; and at least one client server provided at the medical institution that selects original medical data with a disease name according to a medical data request signal received from the central server, extracts feature maps of the selected original medical data, and generates at least one training medical data. Provides a medical data split learning system including;

Description

Medical data split learning system, its control method, and recording medium for performing the method {SYSTEM FOR SPLIT TRAINING OF MEDICAL DATA, METHOD THEREFOR, AND COMPUTER READABLE MEDIUM FOR PERFORMING THE METHOD}

The present invention relates to a medical data split learning system, a control method thereof, and a recording medium for performing the method. More specifically, a medical data split learning system that learns to determine the name of a disease from medical data, a control method thereof, and a method thereof. It relates to a recording medium for performing.

Deep learning is a technology used to cluster or classify objects or data.

In deep learning technology, the number of data samples used for training has a significant impact on the accuracy of the results. Therefore, for deep learning, securing a sufficient amount of data and data mining work must be done first. Data mining is a series of operations that find rules or patterns in data using mathematical models and statistical techniques. The purpose is to classify and cluster data, find correlation and continuity in data, and predict the future.

Deep learning calculates probability using the weight of each data, and repeats the process several times to increase the accuracy of the results. The accumulated data is learned using a convolution neural network (CNN).

However, there is a problem that personal information leakage problems may occur with the data used for deep learning learning, and overfitting of the deep learning model may occur if a sufficient amount of data is not secured.

KR 10-2279376

The technical problem to be solved by the present invention is to provide a medical data split learning system that extracts feature maps of original medical data, generates and learns training medical data, a control method thereof, and a recording medium for performing the method.

However, the technical challenge that this embodiment aims to achieve is not limited to the technical challenges described above, and other technical challenges may exist.

One aspect of the present invention is a medical data split learning system that learns to determine a disease name from medical data, comprising: a central server that requests medical data related to the disease name; and at least one client server provided at the medical institution, wherein the client server selects original medical data related to the disease name according to a medical data request signal received from the central server, extracts a feature map of the selected original medical data, and extracts at least a feature map of the selected original medical data. The client server generates one or more training medical data; wherein the central server collects and learns the at least one training medical data from the client server to generate a model for determining the name of the disease from the medical data. .

In addition, the client server and the central server perform a CNN algorithm, wherein the client server includes a hiding unit provided as one Hidden Layer that generates the training medical data, and the central server generates the at least one training medical data. It includes a control unit that collects and learns, and the control unit may be provided with at least one hidden layer.

In addition, the control unit includes a feature map operation unit that receives feature maps from at least one client server, adds them, and performs a convolution operation on the summed values to regenerate the feature map; an activation function calculation unit that operates by inputting the feature map regenerated by the feature map calculation unit as an input value of an activation function; a pooling operation unit that reduces the dimension of the operation value of the activation function operation unit using a Max Pooling operation; a loss value calculation unit that calculates a loss value using the calculation value of the pooling calculation unit and a predetermined target output value; and a correction value acquisition unit that obtains correction values for the parameters of each calculation unit through calculation of the loss value.

In addition, the central server may further include a parameter update unit that updates the parameter using a correction value for the parameter obtained from the correction value acquisition unit.

Additionally, the activation function calculation unit may calculate the feature vector using the Sigmoid function or the ReLu function.

Another aspect of the present invention is a control method of a medical data split learning system that learns to determine a disease name from medical data, wherein the central server requests medical data related to the disease name, and at least one client server provided at the medical institution Selects original medical data related to the disease name according to a medical data request signal received from the central server, extracts a feature map of the selected original medical data, and generates at least one training medical data, and the central server The at least one training medical data can be collected and learned from one or more client servers to create a model that determines the name of the disease from the medical data.

In addition, the central server collects and learns the at least one training medical data from the at least one client server, where the feature map operation unit receives the feature maps from the at least one client server, adds them, and performs a convolution operation on the summed value. The feature map is regenerated, the activation function operation unit inputs the feature map regenerated by the feature map operation unit as the input value of the activation function and operates, and the pooling operation unit uses Max Pooling operation for the operation value of the activation function operation unit. The dimensionality is reduced, the loss calculation unit calculates the loss value using the calculation value of the pooling calculation unit and a predetermined target output value, and the correction value acquisition unit calculates the correction value for the parameters of each calculation unit through the calculation of the loss value. It can be obtained.

In addition, the parameter updater may further include performing an update on the parameter using a correction value for the parameter obtained from the correction value acquisition unit.

In another aspect of the present invention, a computer program for performing a control method of a medical data split learning system may be recorded on a computer-readable storage medium.

According to one aspect of the present invention described above, the central server learns by receiving feature maps of original medical data from a plurality of client servers, thereby preventing personal information leakage and solving overfitting problems of deep learning models.

Figure 1 is a conceptual diagram showing a medical data split learning system according to an embodiment of the present invention.
Figure 2 is a conceptual diagram showing a control block diagram of a client server.
Figure 3 is a conceptual diagram showing a control block diagram of the central server.
Figure 4 is a diagram showing the process in which the CNN algorithm is performed separately on the client server and the central server.
Figure 5 is a diagram showing original data and training data.
Figure 6 is a table showing the accuracy of confirming Covid-19.
Figure 7 is a table showing fracture discrimination accuracy.
8 to 9 are tables showing loss values calculated by the loss value calculation unit.
Figure 10 is an overall flowchart showing the control method of the medical data split learning system according to an embodiment of the present invention.

The detailed description of the invention described below refers to the accompanying drawings, which show by way of example specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that the various embodiments of the invention are different from one another but are not necessarily mutually exclusive. For example, specific shapes, structures and characteristics described herein may be implemented in one embodiment without departing from the spirit and scope of the invention. Additionally, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the invention. Accordingly, the detailed description that follows is not intended to be taken in a limiting sense, and the scope of the invention is limited only by the appended claims, together with all equivalents to what those claims assert, if properly described. Similar reference numbers in the drawings refer to identical or similar functions across various aspects.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

Figure 1 is a conceptual diagram showing a medical data split learning system according to an embodiment of the present invention.

The medical data split learning system 1 may include a client server 100 and a central server 200.

The medical data split learning system 1 may be a system that learns to determine the name of the disease from medical data based on a deep learning learning algorithm.

Medical data may include patient's endoscopic images, CT and MRI images, etc. Determination of the disease name may include confirmation of COVID-19, presence of fractures, cholesterol level, etc. Deep learning learning algorithms include Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Restricted Boltzmann Machine (RBM), and Deep Trust Neural Network. There are algorithms such as Deep Belief Network (DBN) and Deep Q-Networks. At this time, the medical data split learning system 1 may be a system that learns based on a convolutional neural network (CNN).

The central server 200 may request medical data of MRI and CT images indicating disease from at least one client server 100. The client server 100 may select original medical data representing the corresponding disease according to the medical data request signal received from the central server 200. At this time, the original medical data may be stored in the client server 100. The client server 100 may extract a feature map from the original medical data and transmit it to the central server 200. At this time, the extracted feature map can be used as training medical data for the central server 200. At this time, original medical data is not transmitted or received between the plurality of client servers 100 to protect personal information.

The client server 100 can only perform the process of extracting a feature map from the CNN algorithm, and the central server 200 can extract the feature map, transform it using an activation function, and reduce the dimension through a pooling operation. A CNN algorithm that can be performed can be performed.

The CNN algorithm performed in the client server 100 can be performed in one hidden layer, and the CNN algorithm performed in the central server 200 can be performed in at least one hidden layer.

Figure 2 is a conceptual diagram showing a control block diagram of a client server.

The client server 100 may include an input unit 110, a communication unit 120, a control unit 130, a storage unit 140, and an output unit 150, and the control unit 130 includes a medical data selection unit 131. and a hidden portion 132.

The input unit 110 can input medical data of individual patients from the hospital.

The communication unit 120 may communicate with the central server 200 and may be configured to transmit training medical data to the central server 200.

The storage unit 140 may store medical data input from the input unit 110.

The output unit 150 may output a feature map extracted from original medical data.

The medical data selection unit 131 may select original medical data requested according to a request signal received from the central server 200.

A CNN algorithm may be performed on the hidden portion 132. The hidden portion 132 may be provided with an input layer and one hidden layer so that the CNN algorithm can be performed. Original medical data on which the CNN algorithm will be performed can be input into the input layer. The original medical data can be transformed by extracting the feature map for the original medical data input from the Hidden Layer. Through this, it is possible to protect original medical data. The process of extracting the feature map (fc) is done through Equation 1 below.

[Equation 1]

Here, l means the layer, size^l means the size of the layer, In means the number of input data, Ia means the number of labels, O means the Output Convolution Layer, and w means the weight, and b means the bias of the feature map.

Figure 3 is a conceptual diagram showing a control block diagram of the central server.

The central server 200 may include a communication unit 210, a control unit 220, a storage unit 230, and an output unit 240, and the control unit 220 includes a feature map calculation unit 221 and an activation function calculation unit 222. ), a pooling operation unit 223, a loss value calculation unit 224, a correction value acquisition unit 225, and a parameter update unit 226.

The communication unit 210 can communicate with the client server 100 and receive feature maps received from a plurality of client servers 200.

The control unit 220 may collect and learn at least one training medical data from the client server 100 to create a model that determines the name of the disease from the medical data.

The output unit 240 may output the name of the disease determined from the input medical data as a result of learning.

As a result of learning, the storage unit 230 can match medical data with a specific disease name with the specific disease name and store it.

The feature map calculation unit 221 may concatenate feature maps received from a plurality of client servers 200. The feature map calculation unit 221 may perform a convolution operation on the concatenated feature map to regenerate the feature map. The process by which the feature map calculation unit 221 regenerates the feature map can be accomplished through Equation 2 below.

[Equation 2]

(One)

(2)

(1) is an expression that represents the process of concatenating feature maps received from a plurality of client servers 100.

(2) is an equation showing the result of performing a convolution operation using the concatenated feature map. That is, (2) may mean a regenerated feature map.

At this time, In is the number of input data, Ia is the number of labels, O is the Output Convolution Layer, and w means the weight.

The activation function calculation unit 222 may perform calculations by inputting the feature map regenerated by the feature map calculation unit 221 as an input value of the activation function. At this time, the Sigmoid function or ReLu function can be used as the activation function.

In the present invention, the Sigmoid function was used when the medical data was MURA (MUSculoskeletal Radiographs) and CT data related to COVID-19, and the ReLu function was used when the medical data was data related to Cholesterol.

The pooling operator 223 can reduce the dimension of the operation value of the activation function operator 222 using Max Pooling operation. The purpose of the pooling operation unit 223 is to reduce the size of the data dimension. In other words, the pooling operation is an operation that reduces the size of the vertical and horizontal space in data. There are various types of pooling operation methods (e.g., average, median, maximum value, minimum value, etc.), and the pooling method used in the embodiment of the present invention is the maximum value pooling operation. Using the maximum value pooling operation (max polling), the maximum value can be extracted from a limited area of the image, noise in the data can be removed, and over fitting can be prevented during the data reduction process. Pooling operation can be performed through Equation 3 below.

[Equation 3]

Here, x refers to the matrix input from the pooling layer, l refers to the layer, i refers to the row provided in the matrix, j refers to the column provided in the matrix, and size^l refers to the size of the layer. do. Im is the number of input data of the pooling layer, and Ia may mean the number of labels.

The loss value calculation unit 224 may calculate the loss value using the calculation value of the pooling calculation unit 223 and a predetermined target output value. The loss value can be calculated using MSLE in Equation 4 below, RMSLE in Equation 5, or sMAPE in Equation 6 below. At this time, the predetermined target output value may be GT (Ground Truth). At this time, GT may be a value obtained by performing a convolution operation on the original medical data in the hidden layer of the central server 200 and performing Max Pooling based on the convolution operation value.

[Equation 4]

[Equation 5]

[Equation 6]

means the operation value of the pooling operation unit 223, means a predetermined target output value.

The correction value acquisition unit 225 may obtain correction values for the parameters of each calculation unit from the loss value. At this time, the parameter of each operation unit may mean a weight of w.

The parameter update unit 226 may update the parameter using the correction value for the parameter obtained from the correction value acquisition unit 225. Training of the medical data split learning system (1) can be re-performed using the updated parameters.

Figure 4 is a diagram showing the process in which the CNN algorithm is performed separately on the client server and the central server.

Hospital 1 to Hospital N represent a plurality of client servers 100, which represent servers provided at a plurality of hospitals. Centralized Server may refer to the central server 200. The plurality of client servers 100 may store original medical data, which is medical data.

The CNN algorithm can be performed as a series of processes in the client server 100 and the central server 200. To implement the CNN algorithm, it can be prepared with an input layer, multiple hidden layers, and an output layer. At this time, an input layer and one hidden layer may be provided in the client server 100, and a plurality of hidden layers and an output layer may be provided in the central server 200.

Original medical data is selected and input into the input layer provided in the client server 100, and the feature map of the original medical data can be extracted through convolution operation in the hidden layer.

The plurality of client servers 100 transmit the extracted feature maps to the central server 200, and perform convolution operation, activation function operation, pooling operation, loss value operation, and correction value operation in the hidden layer provided in the central server 200. It can be done.

Figure 5 is a diagram showing original data and training data.

Figure 5 shows chest CT data of COVID-19. (a) may refer to the original COVID-19 chest CT data stored in the storage of the client server 100. (b) may refer to training medical data, which is modified COVID-19 chest CT data generated by the client server 100 by extracting the feature map of the original medical data. Therefore, personal information can be protected by generating training medical data.

Figure 6 is a table showing the accuracy of confirming Covid-19.

The solid line may represent the value learned by the medical data split learning system (1) in which the client server 100 and the central server 200, which are an embodiment of the present invention, learn in a spatial and temporal separation, and the dotted line represents the value learned by the client server ( 100) can only mean the value of learning.

The vertical axis represents the accuracy (%) of determining Covid-19 confirmation from CT data, and the horizontal axis represents the number of Epochs, which is the number of learning repetitions. The solid line shows that even with a small number of epochs, the accuracy of confirming Covid-19 is high.

In addition, as the CT data of COVID-19, the batch size was 64, the input size was 64*64*1, and the sigmoid function was used as the activation function.

Figure 7 is a table showing fracture discrimination accuracy.

Spatio-temporal may refer to a value learned in the medical data split learning system (1) in which the client server 100 and the central server 200, which are an embodiment of the present invention, learn separately in space and time, and Single-client may mean a value learned only by the client server 100.

The vertical axis represents the accuracy (%) of determining fracture from MURA data, and the horizontal axis represents the fracture site.

In the case of spatio-temporal, it can be seen that the accuracy of determining fractures in the finger, elbow, forearm, hand, humerus, shoulder, and wrist areas is higher than that of single-client.

Additionally, as MURA data, Epoch was 50, batch size was 128, input size was 224*224*1, and sigmoid function was used as activation function.

8 to 9 are tables showing loss values calculated by the loss value calculation unit.

This table shows the results of calculating loss values using target output values and actual calculation values based on cholesterol levels received from a plurality of client servers 100.

Figure 8 shows a table measuring CDF based on the process of the loss value calculation unit 224 calculating the loss value based on Equation 4 to Equation 6. Referring to FIG. 8, the solid line may indicate a value learned in the medical data split learning system 1, in which the client server 100 and the central server 200, which are an embodiment of the present invention, learn separately in time and space. , the dotted line may mean a value learned only by the client server 100.

CDF (Cumulative Distribution Function) is a cumulative distribution function that indicates the probability that a given random variable is less than or equal to a specific value.

Having a steep slope at the beginning of the graph may mean that the loss distribution of low values is large, which means that it represents the actual operation value close to the target output value. It can be seen that the solid line shown in (a) to (c) represents the actual calculated value close to the target output value.

Figure 9 shows a table measuring PDF based on the process of the loss value calculation unit 224 calculating the loss value based on Equation 4 to Equation 6. Referring to FIG. 9, the solid line may represent a value learned in the medical data split learning system 1, in which the client server 100 and the central server 200, which are an embodiment of the present invention, learn separately in time and space. , the dotted line may mean a value learned only by the client server 100.

PDF (Probability Density Function) is a probability density function and refers to a probability distribution function based on continuous variables. It refers to a function that indicates how relatively high the probability of a specific random variable interval is compared to other intervals.

In (a) to (c), the result value performed by the medical data split learning system (1), in which the client server 100 and the central server 200, which are an embodiment of the present invention, learn separately in space and time, is close to the low loss value. It can be seen that the probability is high.

Additionally, as the cholesterol level data, Epoch was 200, batch size was 2048, input data was 326032, and ReLu function was used as the activation function.

Figure 10 is an overall flowchart showing the control method of the medical data split learning system according to an embodiment of the present invention.

The central server 200 may request medical data with a specific disease name from a plurality of client servers 100. The plurality of client servers 100 may select original medical data with the requested disease name, extract feature maps of the selected original medical data, and transmit the extracted feature maps to the central server 200.

The central server 200 may add up the feature maps received from a plurality of client servers 100 and perform a convolution operation on the summed feature maps to regenerate the feature maps.

The central server 200 may substitute the regenerated feature map into the activation function, perform a pooling operation using the output value obtained from the activation function, and calculate a loss value using the pooling operation value and a predetermined target output value. The central server 200 can obtain a correction value for the parameter using the calculated loss value and perform a re-learning process with the obtained parameter value.

Such a control method of the medical data split learning system may be implemented as an application or in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. A computer-readable recording medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on a computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable by those skilled in the computer software field.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. -optical media), and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc.

Examples of program instructions include not only machine language code such as that created by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. A hardware device may be configured to operate as one or more software modules to perform processing according to the invention, and vice versa.

Although the embodiments of the present invention have been described above, the spirit of the present invention is not limited to the embodiments presented in the present specification, and those skilled in the art who understand the spirit of the present invention can add components within the scope of the same spirit. , other embodiments can be easily proposed by change, deletion, addition, etc., but this will also be said to be within the scope of the present invention.

1: Medical data split learning system
100: Client Server
110: Input unit of client server
120: Communication unit of client server
130: Control unit of client server
131: Medical data selection unit
132: Hidden part
140: Storage unit of client server
150: Output unit of client server
200: central server
210: Communication department of central server
220: Control unit of central server
221: Feature map calculation unit
222: Activation function operation unit
223: Pooling operation unit
224: Loss value calculation unit
225: Correction value acquisition unit
226: Parameter update unit
230: Storage unit of central server
240: Output unit of central server

Claims (9)

In the medical data split learning system that learns to determine the name of the disease from medical data,
A central server that requests medical data related to the disease name; and
At least one client server provided at a medical institution, wherein the client server selects original medical data related to the disease name according to a medical data request signal received from the central server, extracts a feature map of the selected original medical data, and extracts at least one feature map. Including the client server that generates the above training medical data,
The central server collects and learns the at least one training medical data from the client server to create a model that determines the name of the disease from the medical data,
The client server and the central server perform a CNN algorithm,
The client server includes a hidden unit provided as one hidden layer that generates the training medical data,
The central server includes a control unit that collects and learns the at least one training medical data, and the control unit is provided with at least one hidden layer,
The control unit,
a feature map operation unit that receives feature maps from at least one client server, adds them up, and performs a convolution operation on the summed values to regenerate the feature map;
an activation function calculation unit that operates by inputting the feature map regenerated by the feature map calculation unit as an input value of an activation function;
a pooling operation unit that reduces the dimension of the operation value of the activation function operation unit using a Max Pooling operation;
a loss value calculation unit that calculates a loss value using the calculation value of the pooling calculation unit and a predetermined target output value; and
A medical data split learning system including a correction value acquisition unit that obtains correction values for the parameters of each operation unit through calculation of the loss value.
delete delete According to clause 1,
The central server is,
A medical data split learning system further comprising a parameter update unit that updates the parameter using a correction value for the parameter obtained from the correction value acquisition unit.
According to clause 1,
The activation function operator,
A medical data split learning system that calculates the feature map using the Sigmoid function or ReLu function.
In the control method of a medical data split learning system that learns to determine the name of the disease from medical data,
The central server requests medical data related to the disease name,
At least one client server provided at the medical institution selects original medical data related to the disease name according to the medical data request signal received from the central server, extracts the feature map of the selected original medical data, and provides at least one training medical data. create,
The central server collects and learns the at least one training medical data from the at least one client server to create a model that determines the name of the disease from the medical data,
The client server and the central server perform a CNN algorithm,
The client server includes a hidden unit provided as one hidden layer that generates the training medical data,
The central server includes a control unit that collects and learns the at least one training medical data, and the control unit is provided with at least one hidden layer,
The central server learns by collecting the at least one training medical data from the at least one client server,
The feature map operator receives feature maps from at least one client server, adds them, performs a convolution operation on the summed values, and regenerates the feature map,
The activation function operator inputs the feature map regenerated by the feature map operator as an input value of the activation function and performs the calculation,
The pooling operator reduces the dimension of the operation value of the activation function operator using Max Pooling operation,
The loss value calculation unit calculates the loss value using the calculation value of the pooling calculation unit and a predetermined target output value,
A control method of a medical data split learning system in which the correction value acquisition unit obtains correction values for the parameters of each operation unit through calculation of the loss value.
delete According to clause 6,
A control method of a medical data split learning system further comprising a parameter update unit updating the parameter using a correction value for the parameter obtained from the correction value acquisition unit.
A computer-readable storage medium on which a computer program for performing the control method of the medical data split learning system according to claim 6 is recorded.

Images