Machine learning method of neural network predicting mechanism of action of drug, and method of predicting mechanism of action of drug by using the neural network
A method for a computing device to predict an action mechanism of a drug from medical images of a subject is disclosed. The method includes, from a plurality of medical images obtained in time series, outputting first compressed data corresponding to the plurality of medical images, each of the first compressed data having a smaller size than a corresponding medical image, estimating second compressed data corresponding to a medical image at a next time point to time points at which the plurality of medical images have been captured, based on the first compressed data, and predicting the action mechanism of the drug for the subject by inputting the second compressed data to a neural network predicting the action mechanism of the drug.
1 . A method of predicting an action mechanism of a drug from medical images of a subject by a computing device, the method comprising:
from a plurality of medical images X 0 , . . . X t obtained in time series, outputting first compressed data Z 0 , . . . Z t corresponding to the plurality of medical images X 0 , . . . X t , each of the first compressed data having a smaller size than a corresponding medical image,
wherein the plurality of medical images X 0 , . . . X t are the medical images related to a brain of the subject;
estimating second compressed data h 0 by inputting first compressed data Z 0 to a recurrent neural network;
estimating second compressed data h i corresponding to a predicted medical image X′ i+1 at a time point i+1 after a time point i at which a medical image X i was captured by inputting first compressed data Z; and second compressed data h i−1 to the recurrent neural network, for a period satisfying 0<i≤t, while increasing a value of i; and
predicting the action mechanism of the drug for the subject by inputting second compressed data h t to a neural network learned to predict the action mechanism of the drug,
wherein the second compressed data h t corresponds to a predicted medical image X′ t+2 which reflects a change in a blood flow of the brain or an activity of the brain predicted to occur at a time point t+1 after a time point t.
2 . The method according to claim 1 , wherein the neural network predicting the action mechanism of the drug includes a fully connected layer, and outputs one of predetermined result classes for the action mechanism of the drug by receiving the second compressed data h t .
3 . The method according to claim 1 , wherein the outputting of first compressed data Z 0 , . . . Z t comprises:
performing a pre-treatment algorithm for each of the plurality of medical images X 0 , . . . X t ; and
outputting the first compressed data Z 0 , . . . Z t so as to satisfy a condition that each of pre-treated medical images Y 0 , . . . Y t can be restored within a predetermined error range.
4 . The method according to claim 3 , wherein each of the first compressed data Z 0 , . . . Z t includes a parameter for a probability distribution of a latent variable that enables restoration of a respective one of the pre-treated medical images Y 0 , . . . Y t within the predetermined error range.
5 . The method according to claim 1 , wherein the estimation of second compressed data h t comprises:
estimating the second compressed data h t by inputting the first compressed data Z 0 , . . . Z t to a recurrent neural network in an order corresponding to a time series order of the plurality of medical images X 0 , . . . X t .
6 . The method according to claim 1 , wherein the neural network predicting the action mechanism of the drug is in a state of being trained based on (i) a first training database including a plurality of second compressed data h t of another subjects and (ii) a second training database including action mechanism results of the drug for the another subjects.
7 . A computing device comprising:
a communication unit; and
a processor coupled to the communication unit,
wherein the processor is configured to perform processes including:
from a plurality of medical images X 0 , . . . X t obtained in time series, outputting first compressed data Z 0 , . . . Z t corresponding to the plurality of medical images X 0 , . . . X t , each of the first compressed data having a smaller capacity than a corresponding medical image,
wherein the plurality of medical images X 0 , . . . X t are the medical images related to a brain of the subject;
estimating second compressed data h 0 by inputting first compressed data Z 0 to a recurrent neural network;
estimating second compressed data h 1 corresponding to a predicted medical image X′ i+1 at a time point i+1 after a time point i at which a medical image X i was captured by inputting first compressed data Z; and second compressed data h i−1 to the recurrent neural network, for a period satisfying 0<i≤t, while increasing a value of i; and
predicting the action mechanism of a drug for a subject by inputting second compressed data h t to a neural network learned to predict the action mechanism of the drug,
wherein the second compressed data h t corresponds to a predicted medical image X′ t+1 , which reflects a change in a blood flow of the brain or an activity of the brain predicted to occur at a time point t+1 after a time point t.
8 . The computing device according to claim 7 , wherein the neural network predicting the action mechanism of the drug includes a fully connected layer, and outputs one of predetermined result classes for the action mechanism of the drug by receiving the second compressed data h t .
9 . The computing device according to claim 7 , wherein the outputting of first compressed data Z 0 , . . . Z t comprises:
performing a pre-treatment algorithm for each of the plurality of medical images X 0 , . . . X t ; and
outputting the first compressed data Z 0 , . . . Z t so as to satisfy a condition that each of pre-treated medical images Y 0 , . . . Y t can be restored within a predetermined error range.
10 . The computing device according to claim 9 , wherein each of the first compressed data Z 0 , . . . Z t includes a parameter for a probability distribution of a latent variable that enables restoration of a respective one of the pre-treated medical images Y 0 , . . . Y t within the predetermined error range.
11 . The computing device according to claim 7 , wherein the estimation of second compressed data h t comprises:
estimating the second compressed data h t by inputting the first compressed data Z 0 , . . . Z t to a recurrent neural network in an order corresponding to a time series order of the plurality of medical images X 0 , . . . X t .
12 . The computing device according to claim 7 , wherein the neural network predicting the action mechanism of the drug is in a state of being trained based on (i) a first training database including a plurality of second compressed data h t of another subjects and (ii) a second training database including action mechanism results of the drug for the another subjects.
13 . A non-transitory computer program product for predicting an action mechanism of a drug from medical images of a subject, wherein the computer program product comprises a computer readable storage medium having program code embodied therein, and wherein the program code, when executed, performs operations, the operations comprising:
from a plurality of medical images X 0 , . . . X t obtained in time series, outputting first compressed data Z 0 , . . . Z t corresponding to the plurality of medical images X 0 , . . . X t , each of the first compressed data having a smaller size than a corresponding medical image,
wherein the plurality of medical images X 0 , . . . X t are the medical images related to a brain of the subject;
estimating second compressed data h 0 by inputting first compressed data Z 0 to a recurrent neural network;
estimating second compressed data h i corresponding to a predicted medical image X′ i+1 at a time point i+1 after a time point i at which a medical image X i was captured by inputting first compressed data Z; and second compressed data h i−1 to the recurrent neural network, for a period satisfying 0<i≤t, while increasing a value of i; and
predicting the action mechanism of the drug for the subject by inputting second compressed data h t to a neural network learned to predict the action mechanism of the drug, wherein the second compressed data h t corresponds to a predicted medical image X′ t+1 , which reflects a change in a blood flow of the brain or an activity of the brain predicted to occur at a time point t+1 after a time point t.
14 . The computer program product according to claim 13 , wherein the neural network predicting the action mechanism of the drug includes a fully connected layer, and outputs one of predetermined result classes for the action mechanism of the drug by receiving the second compressed data h t .
15 . The computer program product according to claim 13 , wherein the outputting of first compressed data Z 0 , . . . Z t comprises:
performing a pre-treatment algorithm for each of the plurality of medical images X 0 , . . . X t ; and
outputting the first compressed data Z 0 , . . . Z t so as to satisfy a condition that each of pre-treated medical images Y 0 , . . . Y t can be restored within a predetermined error range.
16 . The computer program product according to claim 15 , wherein each of the first compressed data Z 0 , . . . Z t includes a parameter for a probability distribution of a latent variable that enables restoration of a respective one of the pre-treated medical images Y 0 , . . . Y t within the predetermined error range.
17 . The computer program product according to claim 13 , wherein the estimation of second compressed data h t comprises:
estimating the second compressed data h t by inputting the first compressed data Z 0 , . . . Z t to a recurrent neural network in an order corresponding to a time series order of the plurality of medical images X 0 , . . . X t .