Anatomical object tracking device, anatomical object tracking method, anatomical object tracking program, learning method and learning program for anatomical object classifier

Abstract

To provide means capable of further reliably and easily identifying an outline of a structure from temporally continuous medical images.SOLUTION: An anatomical object tracking device comprises: an image acquisition part for acquiring temporally continuous medical images; a setting acquisition part for acquiring area information representing a range set as an area of a prescribed structure in at least one reference image out of a plurality of frame images included in the medical images; and an identification part using a discriminator to acquire the area information of an object image in the medical image on the basis of the area information of the reference image. The discriminator has learned learning data including both of a set of area information for two frames in which a difference in imaging timing is within a reference range (Tn1) range in the temporally continuous medical images and a set of area information for two frames in which the difference in imaging timing is outside the reference range (Tn1) therein.SELECTED DRAWING: Figure 3

Claims

1. An image acquisition unit that acquires medical images in a continuous time sequence, A setting acquisition unit that acquires region information indicating the range set as the region of a predetermined structure in at least one reference image among a plurality of frame images included in the medical image, A distinguishing unit that uses a discriminator to obtain the region information of a target image other than the reference image in the medical image based on the region information of the reference image, Equipped with, The classifier is trained using both a set of region information from two frames in a sequence of medical images where the difference in shooting timing is within a reference range, and a set of region information from two frames where the difference in shooting timing is outside the reference range, as training data. Based on the region information from at least one reference image, it outputs the region information from a target image other than the reference image. Anatomical object tracking device.

2. The anatomical object tracking device according to claim 1, wherein the ratio of the number of sets of region information within the reference range used for learning to the number of sets of region information outside the reference range is higher than the ratio of time within the reference range to time outside the reference range in the medical image.

3. The anatomical object tracking device according to claim 2, wherein the time within the reference range in the medical image is shorter than the time outside the reference range in the medical image, a portion of the set of region information within the reference range is selectively set, and the remaining set other than the portion is randomly determined.

4. The anatomical object tracking device according to claim 1, wherein the classifier outputs the region information of the target image in response to input data including the reference image, the region information of the reference image, and the target image.

5. The anatomical object tracking device according to claim 1, wherein the reference range is less than one second.

6. The anatomical object tracking device according to claim 1, wherein the reference range is less than one-quarter of the typical time of the phenomenon including the change in the structure of the object detected by the medical image.

7. The anatomical object tracking device according to claim 6, wherein the reference range is less than 1 / 8 of the representative time.

8. The anatomical object tracking device according to claim 6 or 7, wherein the phenomenon to be detected is a periodic phenomenon, and the representative time corresponds to one period of the phenomenon.

9. The anatomical object tracking device according to any one of claims 1 to 7, wherein the aforementioned reference image is one image.

10. The anatomical object tracking device according to any one of claims 1 to 7, wherein the classifier is trained based on deep learning.

11. The anatomical object tracking device according to any one of claims 1 to 7, wherein the medical images are a group of X-ray images.

12. The anatomical object tracking device according to any one of claims 1 to 7, wherein the predetermined structure is a lung field region.

13. The anatomical object tracking device according to any one of claims 1 to 7, wherein the predetermined structure is the tracheal region.

14. Image acquisition step to acquire sequential medical images over time, A setting acquisition step of acquiring region information indicating the range set as the region of a predetermined structure in at least one reference image among a plurality of frame images included in the medical image, A selection step of using a classifier to obtain the region information of a target image other than the reference image in the medical image based on the region information of the reference image, Includes, The classifier is trained using both a set of region information from two frames in a sequence of medical images where the difference in shooting timing is within a reference range, and a set of region information from two frames where the difference in shooting timing is outside the reference range, as training data. Based on the region information from at least one reference image, it outputs the region information from a target image other than the reference image. Methods for tracking anatomical objects.

15. Computers Image acquisition means for acquiring sequential medical images over time. A setting acquisition means for acquiring region information indicating the range set as the region of a predetermined structure in at least one reference image among a plurality of frame images included in the medical image. A means for obtaining the region information of a target image other than the reference image in the medical image based on the region information of the reference image using a classifier, To make it function as, The classifier is trained to include both a set of region information for two frames in a sequence of medical images where the difference in shooting timing is within a reference range, and a set of region information for two frames where the difference in shooting timing is outside the reference range. Based on the region information of at least one reference image, it outputs the region information of a target image in a frame other than the reference image. A program for tracking anatomical objects.

16. A method for training a classifier, wherein the classifier is trained to output the region information of a target image other than the reference image based on the region information of at least one reference image, using region information indicating the extent of a predetermined structure included in each of multiple frame images in a series of time-sequential medical images as ground truth data, The learning process includes both a set of region information for two frames in temporally consecutive medical images where the difference in acquisition timing is within the reference range, and a set of region information for two frames where the difference in acquisition timing is outside the reference range. How to train a classifier.

17. The method for learning a classifier according to claim 16, wherein the ratio of the number of sets of region information within the reference range to the number of sets of region information outside the reference range is higher than the ratio of time within the reference range to time outside the reference range in the medical image.

18. The learning method for a classifier according to claim 17, wherein the time within the reference range in the medical image is shorter than the time outside the reference range in the medical image, a portion of the set of region information within the reference range is selectively set, and the remaining set other than the portion is randomly determined.

19. A method for training a classifier according to claim 16, wherein the classifier is trained to output the region information of a target image in response to input data including a reference image, the region information of the reference image, and a target image different from the reference image.

20. The learning method for a classifier according to claim 16, wherein the reference range is less than one second.

21. The method for learning a classifier according to claim 16, wherein the reference range is less than one-quarter of the typical time of the phenomenon including the change in the structure of the object to be detected by the medical image.

22. The learning method for a classifier according to claim 21, wherein the reference range is less than 1 / 8 of the representative time.

23. The learning method for a classifier according to claim 21 or 22, wherein the phenomenon to be detected is a periodic phenomenon, and the representative time corresponds to one period of the phenomenon.

24. The method for learning a classifier according to any one of claims 16 to 22, wherein the reference image is one image.

25. A method for training a classifier according to any one of claims 16 to 22, wherein training is performed based on deep learning.

26. The method for learning a classifier according to any one of claims 16 to 22, wherein the medical images are a group of X-ray images.

27. The method for learning a classifier according to any one of claims 16 to 22, wherein the predetermined structure is a lung field region.

28. The learning method for a classifier according to any one of claims 16 to 22, wherein the predetermined structure is the tracheal region.

29. Computers This learning means functions to train a classifier to output the region information of a target image in a frame other than the reference image, based on the region information of at least one reference image, using region information indicating the extent of a predetermined structure included in each of multiple frame images in a temporally continuous medical image as ground truth data. In the learning means, the learning is performed by including both a set of region information for two frames in a time-sequential medical image where the difference in acquisition timing is within the reference range, and a set of region information for two frames where the difference in acquisition timing is outside the reference range. Learning program.

Images