Machine learning device and method
Provided are a machine learning device and a method capable of performing machine learning of labeling for accurately attaching a plurality of labels to volume data at once by using learning data with mixed inconsistent labeling. A neural network ( 14 ) receives an input of multi-slice images of learning data Di (i=1, 2, . . . n) of which a class to be labeled is n types, and creates a prediction mask of n anatomical structures i by a convolutional neural network (CNN) or the like (S 1 ). A machine learning unit ( 13 ) calculates a prediction accuracy acc(i) of the class corresponding to the learning data Di for each learning data Di (S 2 ). The machine learning unit ( 13 ) calculates a weighted average M of an error di between the prediction accuracy acc(i) and a ground truth mask Gi. The machine learning unit ( 13 ) calculates a learning loss by a loss function Loss (S 4 ). The machine learning unit ( 13 ) changes each coupling load of the neural network ( 14 ) from an output layer side to an input layer side according to a value of the learning loss calculated by the loss function Loss (S 5 ).
1. A machine learning device comprising:
a first learning data input unit that receives an input of first learning data including volume data of a tomographic image and a ground truth label of a first class in the volume data;
a second learning data input unit that receives an input of second learning data including the volume data and a ground truth label of a second class different from the first class in the volume data;
a prediction label creation unit that, by a neural network, creates a prediction label of the first class from the volume data of the first learning data whose input is received by the first learning data input unit, and creates a prediction label of the second class independent of the prediction label of the first class from the volume data of the second learning data whose input is received by the second learning data input unit;
an integrated error calculation unit that calculates an integrated error between an error between the prediction label of the first class and the ground truth label of the first class and an error between the prediction label of the second class and the ground truth label of the second class by weighted averaging; and
a machine learning unit that causes the neural network to perform machine learning to create the prediction labels of both a label of the first class and a label of the second class in the volume data based on the integrated error calculated by the integrated error calculation unit.
2. The machine learning device according to claim 1 ,
wherein the integrated error calculation unit calculates the integrated error based on an intersection over union (IoU) between the prediction label of the first class and the ground truth label of the first class and an intersection over union (IoU) between the prediction label of the second class and the ground truth label of the second class.
3. The machine learning device according to claim 1 ,
wherein the integrated error calculation unit calculates a detection accuracy based on a Dice coefficient between the prediction label of the first class and the ground truth label of the first class and a Dice coefficient between the prediction label of the second class and the ground truth label of the second class.
4. The machine learning device according to claim 1 ,
wherein the prediction label of the first class and the prediction label of the second class are created based on a sigmoid function.
5. The machine learning device according to claim 1 ,
wherein the tomographic image is a three-dimensional medical tomographic image, and
the first class and the second class include an anatomical structure.
6. The machine learning device according to claim 5 ,
wherein the integrated error calculation unit creates the ground truth label of the second class from the ground truth label of the first class based on relevance data indicating a relationship on an anatomical system between the first class and the second class, and then calculates the integrated error between the error between the prediction label of the first class and the ground truth label of the first class and the error between the prediction label of the second class and the ground truth label of the second class.
7. The machine learning device according to claim 6 ,
wherein the second class is in an anatomically upper rank than the first class in the relevance data.
8. The machine learning device according to claim 1 ,
wherein the integrated error calculation unit calculates an integrated error ignoring the error between the prediction label of the first class and the ground truth label of the first class for a region of the volume data in which the ground truth label of the first class does not exist.
9. A machine learning method executed by a computer, the method comprising:
a step of receiving an input of first learning data including volume data of a tomographic image and a ground truth label of a first class in the volume data;
a step of receiving an input of second learning data including the volume data and a ground truth label of a second class different from the first class in the volume data;
a step of, by a neural network, creating a prediction label of the first class from the volume data of the first learning data and creating a prediction label of the second class independent of the prediction label of the first class from the volume data of the second learning data;
a step of calculating an integrated error between an error between the prediction label of the first class and the ground truth label of the first class and an error between the prediction label of the second class and the ground truth label of the second class by weighted averaging; and
a step of causing the neural network to perform machine learning to create the prediction labels of both a label of the first class and a label of the second class in the volume data based on the integrated error.
10. A machine-learned model that is machine-learned by the machine learning method according to claim 9 .
11. A non-transitory computer-readable recording medium for causing a computer to perform the machine learning method according to claim 9 in a case where instructions stored in the recording medium are read by the computer.