JP7408325B2 - Information processing equipment, learning methods and programs - Google Patents

Description

The present disclosure relates to an information processing device, a learning method, and a program.

2. Description of the Related Art Computer Aided Diagnosis (CAD) systems are known in which a computer analyzes a medical image and presents information to a doctor to help with image interpretation. Among these CAD systems, a CAD system that classifies and presents diagnostic names that are candidates for differential diagnosis from medical images uses a classifier that uses medical data in which a medical image and a correct diagnosis name are paired as training data. There are things that can be achieved through machine learning. In machine learning, a classifier must be trained appropriately for the problem being solved.

The learning device disclosed in Patent Document 1 performs appropriate semi-supervised learning by changing the contribution ratio of errors to labeled data and errors to unlabeled data to learning during the learning process. . Specifically, when an overlearning state is determined, the mixing ratio of unsupervised errors is increased, and if not, it is determined that the learning is appropriate or insufficient, and the unsupervised error ratio is lowered.

The learning device disclosed in Patent Document 2 calculates accuracy using evaluation data when updating parameters of a neural network using a value obtained by multiplying the slope of a loss function by a learning rate, and stores the parameters if the accuracy is improved. However, if there is no improvement, the learning rate is lowered and learning is restarted.

JP2018-181071A Unexamined Japanese Patent Publication No. 2018-160200

However, if the classifier's ability is not appropriate for the problem being solved by machine learning, appropriate learning cannot be performed. Specifically, if the complexity of the distribution that can be expressed by the classifier is high relative to the complexity of the distribution of the problem to be solved, overfitting will occur, and if it is low, underlearning will occur.

The present invention aims to construct a classifier having an appropriate model structure for the problem to be solved.

The first aspect of the present invention is
A determining means for determining the learning state of a classifier that classifies medical data;
Setting means for setting the structure of the classifier based on the state determined by the determining means;
learning means for learning the classifier based on the structure of the classifier set by the setting means;
Equipped with
The setting means sets a degree of change in the structure settings based on a period in which the over-learning state continues or a period in which the under-learning state continues.
It is an information processing device.
The second aspect of the present invention is
A determining means for determining the learning state of a classifier that classifies medical data;
Setting means for setting the structure of the classifier based on the state determined by the determining means;
learning means for learning the classifier based on the structure of the classifier set by the setting means;
Equipped with
The setting means changes the degree of change in the structure setting when a state of overfitting changes to a state of underlearning, or when a state of underlearning changes to a state of overfitting.
It is an information processing device.

The third aspect of the present invention is
A learning method performed by a computer to learn a classifier for classifying medical data, the method comprising:
a determination step of determining the learning state of the classifier;
a setting step of setting a structure of the classifier based on the state determined in the determining step;
a learning step of learning the classifier using teacher data based on the structure set in the setting step;
including;
In the setting step, a degree of change in the structure settings is set based on a period during which the over-learning state continued or a period during which the under-learning state continued.
It is a learning method.
The fourth aspect of the present invention is
A learning method performed by a computer to learn a classifier for classifying medical data, the method comprising:
a determination step of determining the learning state of the classifier;
a setting step of setting a structure of the classifier based on the state determined in the determining step;
a learning step of learning the classifier using teacher data based on the structure set in the setting step;
including;
In the setting step, when the state of overfitting changes to the state of underlearning, or when the state of underlearning changes to the state of overfitting, the degree of change in the setting of the structure is changed,
It is a learning method.

According to the present invention, it is possible to construct a classifier having an appropriate model structure for the problem to be solved.

System configuration diagram of an information processing system including the information processing apparatuses of Embodiments 1 and 2 Hardware configuration diagram of information processing apparatuses of embodiments 1 and 2 Functional block diagram of information processing apparatus of embodiments 1 and 2 Conceptual diagram showing the configuration of the medical image DB of embodiments 1 and 2 Conceptual diagram showing the model structure of the classifier of Embodiments 1 and 2 Flow diagram of the learning process of the classifier in Embodiment 1 Conceptual diagram of learning state determination in Embodiments 1 and 2 Flowchart of learning process of classifier in Embodiment 2

Hereinafter, the present invention will be described in detail based on preferred embodiments thereof with reference to the accompanying drawings. Incidentally, unless otherwise specified, items explained in other embodiments etc. are given the same numbers and their explanations are omitted. Furthermore, the configurations shown in the following embodiments are merely examples, and the present invention is not limited to the illustrated configurations.

<Embodiment 1>
In the first embodiment, an information processing apparatus that performs machine learning of a classifier of a CAD system that classifies diagnostic names for pulmonary nodule shadows on chest X-ray CT (Computed Tomography) images will be described. The information processing device of this embodiment determines the learning state of the classifier when learning the classifier, and sets the model structure of the classifier based on the determined learning state.

(System configuration)
FIG. 1 is a system configuration diagram of an information processing system 1 including an information processing apparatus 101 of this embodiment. In FIG. 1, an information processing system 1 includes a medical image database (hereinafter referred to as medical image DB) 102, an information processing device 101, and a LAN (Local Area Network) 103.

The medical image DB 102 stores medical data including medical images captured by a medical image capturing device such as a CT device and a diagnosis name of the medical images. It also provides known database functionality for searching and retrieving medical data via the LAN 103. The structure of medical data stored in the medical image DB 102 will be explained using FIG. 4.

(Composition of medical data)
FIG. 4 is a conceptual diagram showing the structure of medical data stored in the medical image DB 102.

In FIG. 4, medical data stored in the medical image DB 102 includes a learning data set 410 and a verification data set 420. The training data set 410 is a data set of teacher data used for learning the classifier 301, and the verification data set 420 is a data set used to verify the learning state of the classifier 301. The learning dataset 410 consists of a plurality of learning medical data 411-i (i=1,...,Nt), and the verification dataset 420 consists of a plurality of validation medical data 421-i (i=1,..., Nt). .., Nv). The learning medical data 411-i (i=1,...,Nt) and the verification medical data 421-i (i=1,...,Nv) have similar configurations, including patient information 401 and diagnosis name. 402 and an image 403. Here, the patient information 401 includes information regarding the patient, such as a patient ID and patient name. The diagnosis name 402 is a diagnosis name related to the image 403, and in this embodiment is one of "primary", "metastasis", and "benign". Image 403 is a three-dimensional partial region image containing a pulmonary nodule extracted from a CT image. The diagnosis name 402 may be "malignant" or "benign," or may be a diagnosis name further subdivided into primary, metastasis, and benign. Further, the image 403 may be a combination of a CT image and coordinate information indicating a three-dimensional partial region including a pulmonary nodule. Further, the training data set 410 and the verification data set 420 may be dynamically generated from a single data set.

(Hardware configuration)
FIG. 2 is a hardware configuration diagram of the information processing apparatus 101 of this embodiment. Information processing device 101 is connected to LAN 103, display 207, keyboard 209, and mouse 210.

In FIG. 2, a storage medium 201 is a storage medium such as an OS (Operating System), a processing program for performing various processes related to this embodiment, and a HDD (Hard Disk Drive) that stores various information. A ROM (Read Only Memory) 202 stores a program such as a BIOS (Basic Input Output System) for initializing hardware and starting an OS. A CPU (Central Processing Unit) 203 performs arithmetic processing when executing the BIOS, OS, and processing programs. A RAM (Random Access Memory) 204 temporarily stores information when the CPU 203 executes a program. The LAN interface 205 is an interface for communicating via the LAN 103, which complies with standards such as IEEE (Institute of Electrical and Electronics Engineers) 802.3ab. The display interface 206 is an interface that converts screen information displayed on the display 207 into a signal and outputs the signal. Input interface 208 is an interface for receiving signals from keyboard 209 and mouse 210. The internal bus 211 is a communication path for each block to communicate.

The display 207 is a display device such as a liquid crystal display. The keyboard 209 is an input device for performing key input, and the mouse 210 is an input device for specifying coordinate positions on the screen and inputting button operations.

(Functional configuration)
FIG. 3 is a functional block diagram of the information processing device 101 of this embodiment. When the CPU 203 executes the program, the information processing apparatus 101 functions as a classifier 301, a learning unit 302, a learning state determining unit 303, and a model structure setting unit 304, as shown in FIG.

The classifier 301 inputs the image 403 and outputs the classification result of the diagnosis name. Specifically, the classifier 301 outputs the likelihood that the image 403 is "nuclear," "metastasis," and "benign" as the classification results of the diagnosis name.

Here, the model structure of the classifier 301 will be explained in detail with reference to FIG. 5. FIG. 5 is a conceptual diagram of the model structure of the classifier 301 of the information processing apparatus 101 of this embodiment. The model structure of the classifier 301 of this embodiment is a known CNN (Convolutional Neural Network), which is a type of deep learning.

In FIG. 5, the model structure includes an input layer 501, an intermediate layer 502-i (i=1, . . . , Nl, Nl=2 in FIG. 5), and an output layer 503. Each of the input layer 501, intermediate layer 502-i (i=1, . . . , Nl), and output layer 503 is composed of a plurality of nodes 510, and inputs and outputs of nodes in adjacent layers are connected to each other. An image 403 is input to a node of the input layer 501 . The nodes of the intermediate layer 502-i (i=1, ..., Nl) perform a convolution operation on the output value of the node in front (on the left side in the figure) using the kernel value of each node, and add the weights. The resulting value is output to the subsequent node (on the right side in the figure) (hereinafter referred to as the convolution layer). In this embodiment, the output layer 503 includes three nodes corresponding to "nuclear," "metastasis," and "benign." Each node in the output layer 503 multiplies the output value of the previous node by a kernel value, adds weight to the summation value (hereinafter referred to as a fully connected layer), and adds a known operation called Softmax to the output. Calculate the likelihood for each classification.

In this embodiment, a CNN is used as an example of the model structure, but it may be a DNN (Deep Neural Network) in which the intermediate layer is replaced with a fully connected layer, and is not limited to this.

The learning unit 302 performs machine learning on the classifier 301 using the learning data set 410 acquired from the medical image DB 102. Specifically, the image 403 of the learning data set 410 is input to the classifier 301, and the loss is calculated from the classification result (likelihood) of the diagnosis name that is the output of the classifier and the correct diagnosis name. Furthermore, the learning parameters of the classifier are updated with a value obtained by multiplying the loss gradient by a prespecified learning rate. This update is repeated a predetermined number of times (one update is called an "epoch"). Here, the loss in this embodiment is based on the known multi-class cross entropy (Category
cal Cross Entropy), and the learning parameters to be updated are the kernel values and weights of each node configuring the CNN.

The learning state determination unit 303 determines the learning state of the classifier 301. The learning state is one in which the classifier is over-trained, under-trained, or properly trained. Here, the overfitting state is a state in which the fit is too high for learning data and the fit for other data such as verification data is reduced. The state of insufficient learning is a state in which the learning parameters are not sufficiently adapted to the learning data even if the learning parameters are repeatedly updated. An appropriate learning state is a state in which adaptation to both learning data and verification data is progressing by repeatedly updating the learning parameters. This learning state includes a loss for the verification data set obtained using the verification data set 420 (hereinafter referred to as verification loss) and a loss for the learning data set obtained from the learning unit 302 (hereinafter referred to as learning loss). Judgment is made based on. The determination method will be explained using FIG. 7.

The model structure setting unit 304 sets a model structure regarding the complexity of the classifier 301 based on the learning state determined by the learning state determining unit 303. Here, the model structure related to complexity is a parameter related to the ability of the classifier 301 to express the distribution of a problem to be learned, and in the case of CNN, is the number of convolutional layers and the number of convolutional filters. Note that the number of convolutional layers and the number of filters are examples of the complexity of the model structure, and are not limited thereto. For example, if the classifier 301 is a DNN, this is the number of layers and the number of nodes of a fully connected layer that is an intermediate layer.

(Learning status determination)
FIG. 7 is a conceptual diagram of learning state determination performed by the learning state determination unit 303. In FIG. 7, 701-i (i=1, 2, 3) is a graph (hereinafter referred to as a learning curve) representing the relationship between epochs, learning loss, and verification loss. In the learning curve 701-i (i = 1, 2, 3), 702-i (i = 1, 2, 3) is the learning loss curve (hereinafter referred to as the learning loss curve), 703-i (i = 1 , 2, 3) is a verification loss curve (hereinafter referred to as verification loss curve).

A learning curve 701-1 shows a proper learning state, a learning curve 701-2 shows an overlearning state, and a learning curve 701-3 shows an insufficient learning state.

In the learning curve 701-1 in a proper learning state, as the number of epochs increases, both the learning loss curve 702-1 and the verification loss curve 703-1 decrease. Therefore, the learning state determination unit 303 can determine that the learning state is appropriate when both the learning loss and the verification loss decrease as the number of epochs increases.

In the learning curve 701-2 in the overfitting state, as the number of epochs increases, the learning loss curve 702-1 decreases, but the verification loss curve 703-1 increases. Therefore, the learning state determination unit 303 can determine that the overlearning state is present when the learning loss decreases as the number of epochs increases, and the verification loss increases as the number of epochs increases.

In the learning curve 701-3 due to insufficient learning, even if the number of epochs increases, both the learning loss curve 702-1 and the verification loss curve 703-1 decrease less. Therefore, the learning state determination unit 303 can determine that the learning state is insufficient when both the learning loss and the verification loss decrease as the number of epochs increases, but the degree of decrease is less than or equal to the threshold value. The threshold value of the degree of decrease for determining that learning is insufficient is specified in advance for each problem to be studied.

Here, we have explained an example in which the learning state is determined based on the change in learning loss and validation loss between epochs. Alternatively, the learning state may be determined based on the difference between these values. For example, the learning state determination unit 303 may determine that the learning state is in an overlearning state when the difference between the learning loss and the verification loss is large, that is, when it is greater than or equal to the first threshold. Further, the learning state determining unit 303 may determine that the learning state is an unlearning state when the learning loss is large, that is, when it is equal to or greater than the second threshold value. The learning state determination unit 303 specifies the first threshold value and the second threshold value in advance for each problem to be studied. In cases other than the above, that is, when the learning loss is small and the difference between the learning loss and the verification loss is small, it may be determined that the learning is in an appropriate learning state.

In this embodiment, the learning state is determined based on a change in loss, but the present invention is not limited to this, and the determination may be made based on a change in accuracy. Specifically, the accuracy for the training dataset 410 (hereinafter referred to as learning accuracy) and the accuracy for the verification dataset 420 (hereinafter referred to as validation accuracy) both increase as the epochs increase. If so, it is determined that the learning state is appropriate. In addition, if the verification accuracy decreases despite the increase in learning accuracy, it is determined that there is overfitting, and if neither the learning accuracy nor the verification accuracy rises to a predetermined level, it is determined that learning is insufficient. Furthermore, the learning state may be determined based on the respective values of learning accuracy and verification accuracy, or the difference between these values, instead of the change in accuracy between epochs.

(Learning process)
FIG. 6 is a flow diagram showing the flow of the learning method of the classifier 301 of this embodiment. The learning process is executed based on instructions from the user after the information processing apparatus 101 is started. When instructing the learning process, the user specifies the training data set 410, the verification data set 420, and the initial model structure of the classifier 301 to be used for learning. Specifically, the data set is specified by a URL (Uniform Resource Locator) of a CSV (Comma-Separated Values) file, and the model structure of the classifier 301 is specified by the number of convolution layers and the number of convolution filters.

In step S601, the learning unit 302 creates an initial model based on the specified number of convolutional layers and the specified number of convolutional filters. The initial model is a model in which initial values are set for the kernel value and weight of the convolution filter, which are learning parameters. The initial value may be a fixed value or a random value.

In step S602, the learning unit 302 acquires the specified learning data set 410 and verification data set 420 from the medical image DB 102.

In step S603, the learning unit 302 calculates a learning loss. More specifically, the image 403 is input to the classifier for each learning medical data 411-i (i=1,...,Nt), the likelihood for each classification is calculated, and the calculated likelihood and the correct answer are The known categorical cross entropy is calculated from the diagnosis name 402 and the average is taken.

In step S604, the learning unit 302 updates the learning parameters of the model based on the learning loss calculated in step S603. The parameters are updated using a known method called the gradient method. The gradient method is a method in which the gradient of the function for which the loss is determined is calculated, and the value obtained by multiplying the gradient by a constant called the learning rate is subtracted from the learning parameter. The learning rate may be fixed or variable over the period of learning. If it is variable, it is decreased as learning progresses, but is not limited to this.

In step S605, the learning state determination unit 303 calculates a verification loss. More specifically, the image 403 is input to the classifier for each verification medical data 421-i (i=1,...,Nv), the likelihood for each classification is calculated, and the calculated likelihood and the correct answer are Cate known from the diagnosis name 402
Calculate the gorical Cross Entropy and take the average.

In step S606, the learning state determining unit 303 determines the learning state based on the learning loss and the verification loss. Since the learning state determination has been explained using FIG. 7, repeated explanation will be omitted. The learning state obtained is one of overlearning, underlearning, and appropriate learning.

In step S607, the model structure setting unit 304 sets the increase/decrease number δ (>0) used in step S621 and step S631 according to the change in the learning state determined in step S606. Note that in this embodiment, the model structure setting unit 304 sets the model structure by changing the complexity of the current model structure by a predetermined increase/decrease number δ. The increase/decrease number δ is a parameter representing the degree of change when changing the complexity of the model structure. In step S607, the model structure setting unit 304 specifically decreases the increase/decrease number δ when there is a change from an under-learning or appropriate state to an over-fitting state, or from an over-fitting state to an under-learning or appropriate state, or Reset the increase/decrease number δ to the initial value. Further, the model structure setting unit 304 increases the increase/decrease number δ by a predetermined value when the overfitting state or the underlearning state continues for a certain period of time (a certain number of epochs). The longer the overlearning state or the underlearning state continues, the larger the value of the increase/decrease number δ may be. By doing this, when the state of overfitting continues and then changes to the state of underlearning, and when the state of underlearning continues and then changes to the state of overfitting, the number of increases and decreases (degree of change) δ is It will be smaller than the value in the most recent change.

As will be described later, when changing the model structure, either the number of layers or the number of filters is changed. Therefore, in step S607, the model structure setting unit 304 may determine the increase/decrease number δ for the number of layers and the number of filters, or determine one increase/decrease number δ in step S607, and then determine the increase/decrease number δ in a later step. The number of layers or filters to be increased or decreased may be determined based on the number δ.

In step S608, the model structure setting unit 304 determines whether the learning state is an insufficient learning state. If the learning state is a state of insufficient learning (Yes in S608), the process advances to step S621, and if the learning state is not a state of insufficient learning (No in S608), the process advances to step S609.

In step S621, the model structure setting unit 304 increases the complexity of the model by the increase/decrease number δ. Specifically, either the number of convolutional layers or the number of convolutional filters of the CNN is randomly selected and increased by a predetermined number. Here, in the case of a model in which the size of data output from the convolution layer is smaller than the size of input data, the number of layers to be increased is limited so that the output size does not become smaller than the filter size. The increased kernel values and weights of the convolution filter are initialized with fixed values or random values, and parameters such as other kernel values and weights that are being learned are inherited from the previously learned parameters. Note that the object to be increased may be fixedly selected from one of the number of layers and the number of filters, both may be selected, or may be selected based on a predetermined order. Furthermore, the positions of layers and filters to be added within the model structure may be selected randomly, fixedly, or in a predetermined order. Upon completion of the process in step S621, the process advances to step S632.

In step S609, the model structure setting unit 304 determines whether the learning state is an overfitting state. If the learning state is an overlearning state (Yes in S609), the process advances to step S631, and if the learning state is not an overlearning state (No in S609), the process advances to step S610. If the learning state is neither overfitting nor underlearning (No in S608 and No in S609), that is, in a proper learning state, the model structure is not changed.

In step S631, the model structure setting unit 304 reduces the complexity of the model by the increase/decrease number δ. Specifically, either the number of convolutional layers or the number of convolutional filters of the CNN is randomly selected and reduced by a predetermined number. Here, a limit is set so that the number of intermediate layers is at least one. The kernel values and weights during learning inherit and use the values learned so far. Note that the object to be reduced may be fixedly selected from one of the number of layers and the number of filters, both may be selected, or may be selected based on a predetermined order. Furthermore, the positions of layers and filters to be reduced within the model structure may be selected randomly, fixedly, or in a predetermined order. Upon completion of the process in step S631, the process advances to step S632.

In step S632, the model structure setting unit 304 resets the number of epochs. Specifically, the epoch number counter is reset to zero or a predetermined value. Since this step is a process aimed at performing learning for at least a predetermined number of epochs after changing the model structure, the number of epochs for end determination used in step S611 may be increased by a predetermined number. When the process in step S632 is finished, the process advances to step S611.

In step S610, the learning unit 302 increments the epoch number counter, and proceeds to step S611. In step S611, the learning unit 302 determines whether or not a predetermined number of epochs have ended, and if the predetermined epoch has not ended (No in S611), repeats the process from step S603, and if the predetermined epoch has ended, (Yes in S611) ends the learning process.

As described above, according to the present embodiment, the learning state of a classifier that classifies diagnoses from medical images is determined, and the model structure of the classifier is set based on the learning state. This makes it possible to construct a classifier with an appropriate model structure for the problem being solved.

Furthermore, when changing the model structure in step S631, by inheriting the kernel values and weights that have been learned up to that point, it is possible to omit the initial update of learning parameters for the inherited model structure, and the learning process It also becomes possible to improve efficiency.

(Modification 1 of Embodiment 1)
Although the model structure of the classifier 301 in the first embodiment is a CNN with a convolutional layer as an intermediate layer, it may also have a model structure in which a process called dropout is performed between the convolutional layer and the next layer. Dropout is a process that reduces the output of a node to zero at a fixed rate during learning. The model structure setting unit 304 may increase the dropout rate when determining overfitting, and may decrease the dropout rate when determining underlearning.

Further, the model structure may be a model structure in which a plurality of small networks (hereinafter referred to as modules) are combined to form a large network, such as a known residual network. Modules include known residual modules and inception modules, but are not limited thereto. The model structure setting unit 304 may change the number of modules or the structure of the module itself based on the state of learning.

Further, the model structure may be an SVM (Support Vector Machine) using an RBF (Radial Basis Function Kernel) kernel. The model structure setting unit 304 may change a parameter (referred to as γ) regarding the complexity of the decision boundary based on the state of learning.

(Modification 2 of Embodiment 1)
Embodiment 1 may be semi-supervised learning, in which case, in addition to changing the number of layers and filters, loss for labeled data (hereinafter referred to as supervised loss) and loss for unlabeled data (hereinafter referred to as supervised loss) is required. The contribution ratio of the loss (referred to as "no loss") to learning may be changed. Here, the contribution ratio is a weight for the supervised loss when updating the learning parameters based on the loss that is the sum of the supervised loss and the unsupervised loss. The model structure setting unit 304 sets the contribution ratio when the increase/decrease number δ of the number of layers or the number of filters becomes smaller than a predetermined standard in step S607. That is, in step S621, the contribution ratio to the labeled data is increased, and in step S631, the contribution ratio is decreased.

<Embodiment 2>
In Embodiment 2, similarly to Embodiment 1, an information processing apparatus that is a CAD system that performs diagnostic inference regarding pulmonary nodule shadows on chest X-ray CT images will be described.

The information processing device in Embodiment 1 sets the model structure based on learning states such as overfitting and underlearning, but in this embodiment, in addition to these learning states, a state called vanishing gradient is determined and the model structure is set. I do.

Note that the system configuration, hardware configuration, functional blocks, configuration of the medical image DB 102, model structure, and learning state determination method of the information processing apparatus according to this embodiment are the same as those in Embodiment 1, so explanations thereof will be omitted.

The gradient disappearance determined in this embodiment refers to a state in which the gradient of the loss function calculated when updating the learning parameters becomes close to zero, and the learning parameters are not updated. Vanishing gradient becomes more likely to occur when the number of layers increases.

In the functional block of FIG. 3, the learning state determination unit 303 determines gradient disappearance based on whether the gradient of the loss function becomes equal to or less than a predetermined value close to zero by increasing the number of layers. When gradient vanishing is determined in a state of insufficient learning , the model structure setting unit 304 operates to increase the number of filters, which is a complexity other than the number of layers. That is, an upper limit is set on the number of layers based on the evaluation result of gradient vanishing.

FIG. 8 is a flow diagram of the learning method of the classifier 301 of this embodiment. In FIG. 8, processes similar to those in Embodiment 1 are given the same numbers, and description thereof will be omitted.

If the learning state is insufficient learning in the process of step S608 described in the first embodiment (Yes in S608), the process advances to step S801.

In step S801, the learning state determining unit 303 determines whether gradient vanishing has occurred. If it is determined that the gradient vanishes (Yes in S801), the process proceeds to step S802, and if it is not determined (No in S801), the process proceeds to step S621 described in the first embodiment.

In step S802, the model structure setting unit 304 increases the complexity of the model excluding the number of layers by the increase/decrease number δ. Specifically, if the model structure is CNN, the number of filters is increased. When the process of step S802 is finished, the process advances to step S632 described in the first embodiment.

As described above, according to the present embodiment, the learning state of a classifier that classifies diagnoses from medical images is determined, and the model structure of the classifier is set based on the learning state. It also determines the occurrence of gradient vanishing and sets the model structure. Therefore, it is possible to construct a classifier with an appropriate model structure for the problem to be solved. Furthermore, by not creating layers whose learning parameters are not updated, waste in the model structure can be reduced.

<Other embodiments>
In the above embodiment, the classifier 301 classifies the diagnosis name of pulmonary nodule shadow from the chest X-ray CT image, but the classifier 301 classifies the diagnosis name of any disease from any type of medical image. It may be something that does. Further, the classifier 301 may be a classifier that receives images other than medical images and outputs the classification results. Furthermore, the input to the classifier 301 may be medical data in any data format other than images.

The classifier 301 learned by the information processing device 101 may be used in any way. For example, the information processing apparatus 101 may classify diagnosis names using the classifier 301. Alternatively, the classifier 301 or its learned model data may be copied to another device and the classification may be performed in the other device.

The present invention provides a system or device with a program that implements one or more of the functions of the embodiments described above via a network or a storage medium, and one or more processors in the computer of the system or device reads and executes the program. This can also be achieved by processing. It can also be realized by a circuit (for example, ASIC) that realizes one or more functions.

101: Information processing device 301: Classifier 302: Learning unit 303: Learning state determination unit 304: Model structure setting unit

Claims (22)

A determining means for determining the learning state of a classifier that classifies medical data;
Setting means for setting the structure of the classifier based on the state determined by the determining means;
learning means for learning the classifier based on the structure of the classifier set by the setting means;
Equipped with
The setting means sets a degree of change in the structure settings based on a period in which the over-learning state continues or a period in which the under-learning state continues.
Information processing device. A determining means for determining the learning state of a classifier that classifies medical data;
Setting means for setting the structure of the classifier based on the state determined by the determining means;
learning means for learning the classifier based on the structure of the classifier set by the setting means;
Equipped with
The setting means changes the degree of change in the structure setting when a state of overfitting changes to a state of underlearning, or when a state of underlearning changes to a state of overfitting.
Information processing device. The state determined by the determination means includes at least one of overlearning, insufficient learning, and appropriate learning.
The information processing device according to claim 1 or 2 . The structure set by the setting means is a structure related to the complexity of the classifier,
The information processing device according to any one of claims 1 to 3 . The setting means sets the structure to reduce the complexity when the state is overfitted.
The information processing device according to claim 4 . The setting means sets the structure to increase the complexity when the state is insufficient learning.
The information processing device according to claim 4 or 5 . The classifier is a classifier based on deep learning,
The complexity set by the setting means is the number of layers or nodes of the intermediate layer of the classifier,
The information processing device according to any one of claims 4 to 6 . The classifier is a classifier based on deep learning,
The complexity set by the setting means is the number of convolution layers or the number of filters of the convolution filter of the classifier;
The information processing device according to any one of claims 4 to 7 . The setting means randomly selects which of the number of intermediate layers or convolutional layers, the number of nodes, or the number of filters to be set.
The information processing device according to claim 7 or 8 . The setting means selects in a predetermined order which of the number of intermediate layers or convolutional layers and the number of filters to set.
The information processing device according to claim 7 or 8 . The determining means determines the occurrence of vanishing gradient as a state of the classifier,
The setting means does not increase the number of layers even if learning is insufficient if gradient vanishing occurs in the classifier.
The information processing device according to any one of claims 7 to 10 . The setting means does not change the structure when the state is appropriate learning.
The information processing device according to any one of claims 4 to 11 . The setting means increases the degree of change in the structure settings the longer the overlearning state continues, or the longer the underlearning state continues.
The information processing device according to any one of claims 1 to 12 . When the state of overfitting changes to the state of underlearning, or when the state of underlearning changes to the state of overfitting, the setting means changes the degree of change in the structure settings to the change in the most recent settings. make it smaller than the degree,
The information processing device according to any one of claims 1 to 13 . The learning means performs learning of the classifier by increasing the dropout rate if the state is over-trained and decreasing the dropout rate if the state is under-trained.
The information processing device according to any one of claims 1 to 14 . The learning means performs learning of the classifier by semi-supervised learning,
The setting means determines the degree of change when setting the structure of the classifier based on the change in the state,
If the degree of change is smaller than a predetermined standard, the learning means reduces the contribution ratio of loss to labeled data to learning if the state is overfitted, and reduces the contribution ratio of loss to learning if the state is undertrained. Increasing the contribution ratio of loss to learning for data with
The information processing device according to any one of claims 1 to 15 . The learning means sets the structure of the classifier by the setting means, and then performs learning of the classifier by taking over the values of the parameters of the classifier that have been learned up to that point.
The information processing device according to any one of claims 1 to 16 . The learning means performs learning for at least a predetermined number of epochs after the configuration of the classifier is set by the setting means.
The information processing device according to any one of claims 1 to 17 . A learning method performed by a computer to learn a classifier for classifying medical data, the method comprising:
a determination step of determining the learning state of the classifier;
a setting step of setting a structure of the classifier based on the state determined in the determining step;
a learning step of learning the classifier using teacher data based on the structure set in the setting step;
including;
In the setting step, a degree of change in the structure settings is set based on a period during which the over-learning state continued or a period during which the under-learning state continued.
How to learn. A learning method performed by a computer to learn a classifier for classifying medical data, the method comprising:
a determination step of determining the learning state of the classifier;
a setting step of setting a structure of the classifier based on the state determined in the determining step;
a learning step of learning the classifier using teacher data based on the structure set in the setting step;
including;
In the setting step, when the state of overfitting changes to the state of underlearning, or when the state of underlearning changes to the state of overfitting, the degree of change in the setting of the structure is changed,
How to learn. A program for causing a computer to function as each means of the information processing apparatus according to claim 1. A program for causing a computer to execute each step of the learning method according to claim 19 or 20 .

Images