Device and method for producing a numerical classifier of images, so as to determine the viewing quality of the endoscopic videocapsule images of a segment of the digestive tube

Patent metadata

Field	Value
Publication number	US-10980404-B2
Application number	US-201916979252-A
Country	US
Kind code	B2
Filing date	Mar 13, 2019
Priority date	Mar 14, 2018
Publication date	Apr 20, 2021
Grant date	Apr 20, 2021

Abstract

The invention relates to a method for producing an image numerical classifier, to automatically determine the visualization quality of endoscopy videocapsule images of segments of the digestive tract,
comprising a step of acquiring a video in the digestive tract by a videocapsule; a step of extracting images from the video; a so-called “ground truth” step of clinically evaluating the visualization quality of the images based on medical criteria, a step of selecting an initial set of “adequate” visualization images and “inadequate” visualization images, a step of calculating at least one numerical parameter relating to at least one of the medical criteria, a statistical machine learning step to produce the numerical classifier.

First claim

The invention claimed is:

1. A device for producing a numerical classifier to determine the visualization quality of endoscopy videocapsule images of a segment of a digestive tract, comprising:
a videocapsule for acquiring a video of segments of the digestive tract;
a memory, coupled to the videocapsule;
a database with videocapsule-extracted images classified during a ground truth step, said extracted images being classified as adequate visualization images and inadequate visualization images, whereby said extracted images are classified according to a score, said score being determined by a visual analysis of the extracted images based on one or more medical criteria, said medical criteria comprising:
(a) a percentage of mucous membrane visualized,
(b) image brightness,
(c) a presence of bubbles,
(d) a presence of bile/chyme, and
(e) a presence of liquids and undigested debris; and
a processor connected to the memory and incorporating the database, said processor being configured to:
calculate at least two numerical parameters, each relating to at least one of said medical criteria used for determining the score, and extracted from the images of the database, the at least two calculated numerical parameters comprising:
a global colorimetric parameter,
a parameter reflecting an abundance of bubbles, and
a parameter reflecting brightness;

perform a statistical machine learning according to a random forest classifier, comprising the steps of:
a) selecting learning images from the extracted images in the database,
b) performing—random drawings of the learning images, wherein each random drawing includes the same number of the learning images,
c) automatically producing a numerical classifier, said numerical classifier being produced with a succession of automatic thresholds applied to the at least two calculated numerical parameters, by numerical analysis of these learning images to construct N binary decision trees, wherein each random drawing produces one binary decision tree,
wherein each binary decision tree is constructed using the at least two calculated numerical parameters, said automatic thresholds being calculated automatically at each node of each binary decision tree based on the at least two calculated numerical parameters that allows for the distribution of said learning images into a first subgroup and into a second subgroup closest to the distribution of the adequate visualization images and the inadequate visualization images, respectively, performed during the ground truth step with the resulting set of binary decision trees constituting said numerical classifier, said numerical classifier being determined to automatically distribute said learning images into the first subgroup of learning images including the largest number of the adequate visualization images and the second subgroup of learning images including the largest number of the inadequate visualization images.

2. The device according to claim 1 , wherein the processor is configured to perform a step of numerically deciding the visualization quality of the test images, which is a system for voting on all numerical decisions of the binary decision trees, with each test image having been tested in all the binary decision trees, said test images being the remaining images of the database minus the learning images in the numerical classifiers.

3. The device according to claim 1 , wherein the processor is configured to repeat steps b) and c) x times to obtain a final classifier consisting of a plurality of the numerical classifiers resulting from the statistical machine learning and thus having x*N binary decision trees, with N greater than or equal to 100 and x greater than or equal to 10.

4. The device according to claim 1 , wherein
the global colorimetric parameter is a red/green ratio of an image when the segment of the digestive tract is a small bowel, or a red/(green+blue) ratio when the segment of the digestive tract is a colon;
the parameter reflecting the abundance of bubbles is:
a textural parameter from a gray-level co-occurrence matrix (GLCM) of a processed image, or
a bubble occupying surface; and

the parameter reflecting brightness is gray-level contrast of an image.

5. The device according to claim 1 , wherein the numerical classifier is produced based on the 3 following numerical parameters:
an overall colorimetric parameter, said overall colorimetric parameter being a red/green ratio of an image when the segment of the digestive tract is a small bowel, or a red/(green+blue) ratio when the segment of the digestive tract is a colon;
the parameter reflecting the abundance of bubbles which is:
a textural parameter from a gray-level co-occurrence matrix (GLCM) of a processed image, or
a bubble occupying surface; and

the parameter reflecting the brightness which is a gray-level contrast of an image.

6. A control method applied to a video made by a videocapsule, in at least one segment of the digestive tract of a person, to automatically determine the visualization quality of images of the video,
using the numerical classifier of the device according to claim 1 , applied to the images of the video,
to automatically determine, during an automatic control examination, the images with adequate visualization, the images with inadequate visualization, and a rate of adequate visualization images in the video according to a decision of the numerical classifier.

7. The control method according to claim 6 , applied to different persons and further comprising:
a preliminary step of intestinal preparation for the control examination, different for each person;
an automatic examination control step,
a step of comparing the efficacy of the different intestinal preparations under examination depending on the rate of the adequate visualization images determined for each different intestinal preparation by the control method.

8. A control device for automatically determining the visualization quality of a video of one or more segments of a person's digestive tract, performed by the device for producing a numerical classifier of claim 1
the control device configured to:
calculate at least two numerical parameters, each relating to one of the following medical criteria, in images of the video, said at least two numerical parameters comprising:
(a) a percentage of mucous membrane visualized,
(b) a brightness of the image,
(c) a presence of bubbles,
(d) a presence of bile/chyme,
(e) a presence of liquids and undigested debris; and

numerically test the images in the video, and numerically decide the images with adequate visualization, the images with inadequate visualization, and a rate of adequate visualization images in the video according to a decision of the numerical classifier.

Assignees

• Hopitaux Paris Assist Publique
• Univ Sorbonne
• Ecole Nat Superieure De Lelectronique Et De Ses Applications
• Centre Nat Rech Scient
• Cy Cergy Paris Univ

Inventors

• Dray Xavier
• Histace Aymeric

Classifications

• A61B1/000096

  using artificial intelligence · CPC title

• A61B1/000095

  for image enhancement · CPC title

• A61B1/000094

  extracting biological structures · CPC title

• A61B1/041 Primary

  Capsule endoscopes for imaging · CPC title

• A61B1/0002

  provided with data storages · CPC title