Methods and devices for grading a medical image

Patent metadata

Field	Value
Publication number	US-11742073-B2
Application number	US-202117469520-A
Country	US
Kind code	B2
Filing date	Sep 8, 2021
Priority date	Dec 27, 2018
Publication date	Aug 29, 2023
Grant date	Aug 29, 2023

Abstract

Method and system for grading a medical image. For example, a system for grading a medical image comprising a grading network configured to provide a grading result corresponding to the medical image based on at least the medical image and/or a list of lesion candidates generated by a lesion identification network.

First claim

What is claimed is:

1. A system for grading a plurality of medical images, the system comprising:
a grading network configured to provide a plurality of grading results in urgency corresponding to the plurality of medical images, each grading result in urgency of the plurality of grading results in urgency corresponding to a respective medical image of the plurality of medical images, the each grading result is determined based at least in part on the respective medical image and a plurality of lesion candidates, the plurality of lesion candidates being identified by a lesion identification network based on the respective medical image;
wherein the grading network is further configured to:
provide one or more outputs to guide scheduling of the plurality of medical images for manual review and diagnosis based at least in part on the plurality of grading results in urgency corresponding to the plurality of medical images.

2. The system of claim 1 , wherein the plurality of lesion candidates includes a first lesion candidate corresponding to a first disease and a second lesion candidate corresponding to a second disease, wherein the first disease is different from the second disease.

3. The system of claim 1 , wherein:
the lesion identification network is configured to provide a score corresponding to the plurality of lesion candidates;
the grading result is provided further based on the score; and
the score is provided based at least in part on a probability value corresponding to the medical image including lesions of multiple lesion categories.

4. The system of claim 1 , wherein:
the respective medical image includes a chest radiograph; and
the plurality of lesion candidates includes at least one selected from a group consisting of a list of lung lesion candidates, a list of cardiac lesion candidates, a list of cardiovascular lesion candidates, a list of hilar lesion candidates, a list of bone tissue lesion candidates, and a list of pleural lesion candidates.

5. The system of claim 1 , wherein the grading network is a student network trained by an attention transfer learning process comprising:
establishing a teacher network and the student network for the grading network;
training the teacher network; and
training the student network based at least in part on:
extracting a feature map from one or more middle layers corresponding to both the student network and the teacher network;
calculating one or more attention transfer learning losses corresponding to the one or more middle layers; and
backpropagating the one or more attention transfer learning losses into the student network.

6. The system of claim 1 , wherein the lesion identification network is configured to identify at least one lesion characteristic selected from a group consisting of color, shape, size, grayscale value, position, and morphology.

7. The system of claim 1 , wherein the grading result is determined based on the medical image as a whole or one or more partial regions of the medical image.

8. The system of claim 1 , wherein the grading result is selected from a group consisting of a first grade, a second grade, and a third grade, wherein:
the first grade corresponds to a first priority for manual review and diagnosis;
the second grade corresponds to a second priority for manual review and diagnosis;
the third grade corresponds to a third priority for manual review and diagnosis;
the first priority being greater than the second priority and the third priority.

9. A computer-implemented method for grading a plurality of medical images, the method comprising:
providing a plurality of grading results in urgency corresponding to the plurality of medical images by a grading network, each grading result in urgency of the plurality of grading results in urgency corresponding to a respective medical image of the plurality of medical images, the each grading result is determined based at least in part on the respective medical image and a plurality of lesion candidates, the plurality of lesion candidates being identified by a lesion identification network based on the respective medical image; and
providing one or more outputs to guide scheduling of the plurality of medical images for manual review and diagnosis based at least in part on the plurality of grading results in urgency corresponding to the plurality of medical images.

10. The computer-implemented method of claim 9 , wherein the plurality of lesion candidates includes a first lesion candidate corresponding to a first disease and a second lesion candidate corresponding to a second disease, wherein the first disease is different from the second disease.

11. The computer-implemented method of claim 9 , wherein:
the lesion identification network is configured to provide a score corresponding to the plurality of lesion candidates;
the grading result is provided further based on the score; and
the score is provided based at least in part on a probability value corresponding to the medical image including lesions of multiple lesion categories.

12. The computer-implemented method of claim 9 , wherein:
the respective medical image includes a chest radiograph; and
the plurality of lesion candidates includes at least one selected from a group consisting of a list of lung lesion candidates, a list of cardiac lesion candidates, a list of cardiovascular lesion candidates, a list of hilar lesion candidates, a list of bone tissue lesion candidates, and a list of pleural lesion candidates.

13. The computer-implemented method of claim 9 , wherein the grading network is a student network trained by an attention transfer learning process, the method further comprising:
establishing a teacher network and the student network for the grading network;
training the teacher network; and
training the student network based at least in part on:
extracting a feature map from one or more middle layers corresponding to both the student network and the teacher network;
calculating one or more attention transfer learning losses corresponding to the one or more middle layers; and
backpropagating the one or more attention transfer learning losses into the student network.

14. The computer-implemented method of claim 9 , wherein the lesion identification network is configured to identify at least one lesion characteristic selected from a group consisting of color, shape, size, grayscale value, position, and morphology.

15. The computer-implemented method of claim 9 , wherein the grading result is determined based on the medical image as a whole or one or more partial regions of the medical image.

16. The computer-implemented method of claim 9 , wherein the grading result is selected from a group consisting of a first grade, a second grade, and a third grade, wherein:
the first grade corresponds to a first priority for manual review and diagnosis;
the second grade corresponds to a second priority for manual review and diagnosis;
the third grade corresponds to a third priority for manual review and diagnosis;
the first priority being greater than the second priority and the third priority.

17. The computer-implemented method of claim 9 , wherein the scheduling of the plurality of radiographs for manual review and diagnosis is based on at least one selected from a group consisting of expertise, pay level, and seniority.

18. A non-transitory computer-readable medium with instructions stored thereon, that when executed by a processor, cause the processor to perform operations comprise:
providing a plurality of grading res

Assignees

• Shanghai United Imaging Intelligence Co Ltd

Inventors

• Cheng Jie-Zhi
• Gong Zaiwen
• He Zhiqiang
• Zhan Yiqiang
• Zhou Xiang Sean

Classifications

• G06N3/096

  Transfer learning · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G16H30/40 Primary

  for processing medical images, e.g. editing · CPC title