Medical image learning method and medical image processing device

1. A medical image learning method comprising:

generating a first model through first learning using a first learning image group including a normal image which is a medical image having no region of interest;

inputting an input image group to the first model, the input image group including at least a medical image different from the first learning image group, to perform abnormality detection based on a difference from a reference of the first model;

performing sorting of an extracted image from the input image group, the extracted image being used for learning to prevent erroneous recognition of the region of interest, according to a result of the abnormality detection; and

generating a second model through second learning using a second learning image group including at least the extracted image, the second model detecting a medical image having the region of interest from input medical images.

2. The medical image learning method according to claim 1 ,

wherein each medical image of the input first learning image group is generated or restored in the first learning.

3. The medical image learning method according to claim 1 , wherein

the input image group includes the normal image, and

the sorting is a first sorting where the medical image determined to be abnormal as the result of the abnormality detection is sorted into the extracted image.

4. The medical image learning method according to claim 1 , wherein

the input image group includes an abnormal image which is a medical image having the region of interest, and

the sorting is a second sorting where the medical image determined to be normal as the result of the abnormality detection is sorted into the extracted image.

5. The medical image learning method according to claim 1 , wherein

there are a case where the input image group includes the normal image and a case where the input image group includes an abnormal image which is a medical image having the region of interest,

in the case where the input image group includes the normal image, the sorting is a first sorting where the medical image determined to be abnormal as the result of the abnormality detection is sorted into the extracted image,

in the case where the input image group includes the abnormal image, the sorting is a second sorting where the medical image determined to be normal as the result of the abnormality detection is sorted into the extracted image, and

the first model switches between the first sorting and the second sorting depending on a type of the input image group.

6. The medical image learning method according to claim 1 ,

wherein the second learning image group includes the extracted image and an medical image different from the extracted image.

7. The medical image learning method according to claim 1 ,

wherein a degree of abnormality of each medical image of the input image group is evaluated through the abnormality detection, and

the second model is generated through the second learning using the second learning image group weighted on each medical image according to the degree of abnormality.

8. The medical image learning method according to claim 7 ,

wherein in the weighting,

weight is set small for the extracted image, and

weight is set large for a non-extracted image, which is not sorted as the extracted image from the input image group.

9. The medical image learning method according to claim 1 ,

wherein the first model is an autoencoder.

10. The medical image learning method according to claim 9 ,

wherein the first model learns autoencoding and calculates a reference intermediate feature amount of the normal image, in the first learning,

acquires an autoencoded image and an intermediate feature amount on the basis of each medical image of the input image group input to the first model, and

performs the abnormality detection by using at least one of a restoration error obtained by comparing an image before the autoencoding with the autoencoded image or a comparison result between the reference intermediate feature amount and the intermediate feature amount.

11. The medical image learning method according to claim 1 ,

wherein the first model is a GAN, and performs the abnormality detection by using at least one of a trained discriminator or a trained generator.

12. The medical image learning method according to claim 1 ,

wherein the first model learns image interpolation through the first learning,

makes each medical image constituting the input image group, which is input to the first model, deficient and interpolates the deficiency by using a reference calculated through the first learning, and

performs the abnormality detection by using a restoration error calculated by comparing the medical image before the deficiency with the medical image after the interpolation.

13. A medical image processing device comprising:

a processor configured to:

generate a first model through first learning using a first learning image group including a normal image which is a medical image having no region of interest;

input an input image group to the first model, the input image group including at least a medical image different from the first learning image group, to perform abnormality detection based on a difference from a reference of the first model;

perform sorting of an extracted image from the input image group, the extracted image being used for learning to prevent erroneous recognition of the region of interest, according to a result of the abnormality detection, and

generate a second model through second learning using a second learning image group including at least the extracted image, the second model detecting a medical image having the region of interest from input medical images.