Method for training a decision system for segmenting medical images

Patent metadata

Field	Value
Publication number	US-12511747-B2
Application number	US-202218550020-A
Country	US
Kind code	B2
Filing date	Mar 11, 2022
Priority date	Mar 11, 2021
Publication date	Dec 30, 2025
Grant date	Dec 30, 2025

Abstract

A method implemented by computer means for training a decision system for segmenting medical images from a training set of annotated medical images, the segments belonging to at least one class, each annotation of the medical images including quantitative information about a number of pixels of the image that belongs to each of the classes, the method using weakly-supervised algorithm based on a percentage of the pixels of the image belonging to a concerned class.

First claim

The invention claimed is:

1 . A method implemented by computer means for training a decision system for segmenting medical images from a training set of annotated medical images, said segments belonging to at least one class, each annotation of said medical images including quantitative information about a number of pixels of the image that belongs to each of said classes, said method comprising, for at least a part of the images of the training set, the following iterative steps:
(a) dividing an image of said part of the training set into a set of sub-images or deriving features from said image, so as to obtain image instances, (b) predicting, using the decision system, for each pixel of at least part of said sub-images or for each image feature, the probability that said pixel or said feature belongs to each of the above classes, the prediction from the decision system being in the form of a prediction tensor, (c) calculating, from the prediction tensor, a pseudo ground-truth tensor using a pseudo ground-truth generator, said pseudo ground-truth generator assigning, for each instance and for each of said classes, a single probability value to at least part of the pixels or features having the highest probability of belonging to the concerned class, the number of pixels to which the high probability value is assigned being dependent on said quantitative information of the annotation, (d) calculating, for each instance, the result of a cost function based on the prediction tensor and the pseudo ground-truth tensor, and (e) updating the parameters of the decision system based on the result of the cost function, said method further comprising a final step of outputting a pre-trained model with updated parameters.

2 . The method according to claim 1 , wherein step consists in calculating, from the prediction tensor, a pseudo ground-truth tensor using a pseudo ground-truth generator, said pseudo ground-truth generator assigning, for each instance and for each of said classes, a single high probability value to at least part of the pixels or features having the highest probability of belonging to the concerned class and a single low probability value to at least part of the other pixels or features, the number of pixels to which the high probability value is assigned being dependent on said quantitative information of the annotation.

3 . The method according to claim 2 , wherein the single high value is a probability value equal to 1, the single low value being a probability value equal to 0.

4 . The method according to claim 1 , wherein each image of the training set is randomly divided into a set of sub-images.

5 . The method according to claim 1 , wherein a mask is applied to at least one image of the training set, discarding a part of said image from being divided into sub-images.

6 . The method according to claim 1 , wherein the decision system implement non-linear decisions algorithms, such as neural networks, including convolutional neural network or transformer-based algorithm.

7 . A computer software, comprising instructions to implement at least a part of the method according to claim 1 when the software is executed by a processor.

8 . A computer device comprising:
an input interface to receive a training set of annotated medical images, a memory for storing at least instructions of a computer program comprising instructions to implement at least a part of the method according to claim 1 , a processor accessing to the memory for reading the aforesaid instructions and executing then the method, and an output interface to provide the pre-trained model with updated parameters.

9 . A computer-readable non-transient recording medium on which a computer software is registered to implement the method according to claim 1 , when the computer software is executed by a processor.

Assignees

• Roussy Inst Gustave
• Univ Paris Saclay
• Centralesupelec
• Centrale Supelec

Inventors

• Lerousseau Marvin
• Deutsch Eric

Classifications

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G16H30/40

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

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0895

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

• G06N3/045

  Combinations of networks · CPC title