Image processing system, image processing device, endoscope system, interface, image processing method and information storage medium
An image processing system includes an interface to which an annotation result on a learning image captured inside a living body is input and a processor including hardware. The processor acquires metadata including difficulty information indicating difficulty of the annotation of the learning image itself, determines reliability information indicating reliability of the annotation result based on the metadata, and outputs a dataset in which the learning image, the annotation result, and the reliability information are associated with each other, as data used in generating a trained model used in inference based on deep learning on an inference target image captured inside a living body.
1. An image processing system comprising:
a processor including hardware, wherein the processor is configured to:
receive an annotation result, inputted through an interface, on a learning image captured inside a living body;
acquire metadata including difficulty information indicating difficulty of annotating the learning image;
determine reliability information indicating reliability of the annotation result based on the metadata; and
output, as data used in generating a trained model used in inference based on deep learning on an inference target image captured inside a living body, a dataset in which the learning image, the annotation result, and the reliability information are associated with each other.
2. The image processing system as defined in claim 1 ,
wherein the processor is configured to perform image processing on the learning image to acquire the difficulty information.
3. The image processing system as defined in claim 2 ,
wherein the processor is configured to determine, by the image processing, at least one of information indicating identifiability of an annotation target in the learning image, information indicating image quality of the learning image, and information indicating a shielded degree of the annotation target in the learning image, and acquire the difficulty information.
4. The image processing system as defined in claim 3 ,
wherein the processor is configured to determine that the lower the identifiability, the higher the difficulty.
5. The image processing system as defined in claim 3 ,
wherein the processor is configured to determine that the lower the image quality, the higher the difficulty.
6. The image processing system as defined in claim 3 ,
wherein the processor is configured to determine that the higher the shielded degree, the higher the difficulty.
7. The image processing system as defined in claim 1 ,
wherein:
the difficulty information includes operation time information indicating time required to give the annotation result on the learning image; and
the processor is configured to determine that the difficulty is high when operation time indicated by the operation time information is longer than a predetermined time.
8. The image processing system as defined in claim 1 ,
wherein:
the metadata includes capability information indicating capability of a user who inputs the annotation result using the interface; and
the processor is configured to determine the reliability information based on the difficulty information and the capability information.
9. The image processing system as defined in claim 8 ,
wherein the capability information includes fatigue level information indicating a fatigue level of the user.
10. The image processing system as defined in claim 8 ,
wherein the capability information includes proficiency level information indicating a proficiency level of the user.
11. The image processing system as defined in claim 1 , wherein:
the interface accepts, as the metadata, self-assessment information indicating self-assessment of a user who has input the annotation result; and
the processor is configured to determine the reliability information based on the metadata including the self-assessment information.
12. The image processing system as defined in claim 1 , wherein, in a case where a first to N-th (N is an integer of 2 or more) annotation results, which were input by a plurality of users, are associated with an arbitrary one of the learning image, the processor is configured to:
determine first to N-th reliability information based on the metadata assigned to each of the first to N-th annotation results; and
determine the annotation result to be associated with the learning image based on the first to N-th annotation results, and determine the reliability information of the determined annotation result based on the first to N-th reliability information.
13. The image processing system as defined in claim 1 , wherein:
the interface includes a display that displays a display image including the learning image; and
the processor is configured to control the display to display the metadata in the display image.
14. The image processing system as defined in claim 1 ,
wherein the processor is configured to:
output the learning image and the annotation result as teacher data for the deep learning; and
output the reliability information as weight information of an objective function in the deep learning.
15. The image processing system as defined in claim 1 ,
wherein the processor is configured to:
output the learning image and the annotation result as teacher data for the deep learning; and
output, as weight information of an objective function in the deep learning, a value nonlinear to the reliability information.
16. The image processing system as defined in claim 1 ,
wherein the processor is configured to perform a process of requesting re-input of the annotation result when the reliability indicated by the reliability information is lower than a given reference value.
17. An image processing device comprising:
a processor including hardware, wherein the processor is configured to:
receive input of an annotation result on a learning image captured inside a living body;
acquire metadata including difficulty information indicating difficulty of annotating the learning image; and
output the acquired metadata as information for determining reliability of the annotation result in association with the learning image,
wherein the learning image and the annotation result are used in generating a trained model used in inference based on deep learning on an inference target image captured inside a living body, and the reliability being used as weight information of an objective function in the deep learning.
18. An interface comprising:
an input section configured to receive input of an annotation result on a learning image captured inside a living body,
wherein:
the interface is connected to a processor including hardware;
the interface is configured to input, to the processor, metadata including difficulty information indicating difficulty of annotating the learning image;
the processor determines reliability information indicating reliability of the annotation result based on the metadata; and
the processor outputs, as data used in generating a trained model used in inference based on deep learning on an inference target image captured inside a living body, a dataset in which the learning image, the annotation result, and the reliability information are associated with each other.
19. An image processing method comprising
acquiring an annotation result on a learning image captured inside a living body;
acquiring metadata including difficulty information indicating difficulty of annotating of the learning image;
determining reliability information indicating reliability of the annotation result based on the metadata; and
outputting, as data used in generating a trained model used in inference based on deep learning on an inference target image captured inside a living body, a dataset in which the learning image, the annotation result, and the reliability information are associated with each other.
20. A non-transitory information storage medium storing a program, the program causing a computer to at least execute:
acquiring an annotation result on a learning image captured inside a living body;
acquiring metadata including difficulty information indicating difficulty of annotating the learning image;
determining reliability information indicating reliability of the annotation result based on the metadata; and
outputting, as data used in generating a trained model used in inference based on deep learning on an inference target image captured inside a living body, a dataset in which the learning image, the annotation result, and the reliability information are associated with each other.