Detection model training method and apparatus, and terminal device

Patent metadata

Field	Value
Publication number	US-11315677-B2
Application number	US-202017084475-A
Country	US
Kind code	B2
Filing date	Oct 29, 2020
Priority date	Oct 25, 2018
Publication date	Apr 26, 2022
Grant date	Apr 26, 2022

Abstract

The present application discloses a detection model training method and apparatus. The method includes determining an initial training model; determining a training sample; determining whether a lesion target is present in a first user body organ image through the initial detection model according to a feature of the each first user body organ image, to obtain a detection result; and determining a domain that each user body organ image in the training sample belongs to through the adaptive model according to a feature of the each user body organ image, to obtain a domain classification result; calculating, a loss function value related to the initial training model according to the detection result, the domain classification result, a first identifier, a second identifier, and a third identifier; and adjusting a parameter value in the initial training model according to the loss function value, to obtain a final detection model.

First claim

What is claimed is:

1. A detection model training method, comprising:
determining, by a training apparatus, an initial training model, the initial training model comprising an initial detection model and an adaptive model;
determining, by the training apparatus, a training sample, the training sample comprising source domain data and target domain data, the source domain data comprising a plurality of first user body organ images, each first user body organ image comprising: a first identifier of whether a lesion target is present, and a second identifier of a domain that the each first user body organ image belongs to; the target domain data comprising: a plurality of second user body organ images, and a third identifier of a domain that each second user body organ image belongs to;
determining, by the training apparatus, whether a lesion target is present in the each first user body organ image through the initial detection model according to a feature of the each first user body organ image, to obtain a detection result; and determining, by the training apparatus, a domain that each user body organ image in the training sample belongs to through the adaptive model according to a feature of the each user body organ image, to obtain a domain classification result;
calculating, by the training apparatus, a loss function value related to the initial training model according to the detection result, the domain classification result, the first identifier, the second identifier, and the third identifier; and
adjusting, by the training apparatus, a parameter value in the initial training model according to the loss function value, to obtain a final detection model.

2. The method according to claim 1 , wherein some of the first user body organ images comprise lesion targets, and the source domain data further comprises types and locations of the lesion targets comprised in the some of the user body organ images;
before the calculating a loss function value related to the initial training model, the method further comprises: determining, by the training apparatus, the types and the locations of the lesion targets comprised in the some of the user body organ images through the initial detection model, to obtain a target detection result; and
the calculating a loss function value related to the initial training model further comprises: calculating the loss function value related to the initial training model according to the detection result, the domain classification result, the target detection result, the first identifier, the second identifier, and the third identifier.

3. The method according to claim 2 , wherein the loss function value related to the initial training model further comprises: a function calculated value of a detection loss function and an adaptive loss function;
the detection loss function comprises: a first error between information that is determined according to the initial detection model and that is about whether the lesion targets are comprised in the first user body organ images and the first identifier in the training sample; and a third error between the types and the locations that are of the lesion targets in the some of the user body organ images and that are determined according to the initial detection model and the types and the locations of the lesion targets in the some of the user body organ images in the training sample; and
the adaptive loss function comprises: a second error between information about whether the each user body organ image in the training sample belongs to a source domain or a target domain, which is determined according to the adaptive model, and the second identifier and the third identifier in the training sample.

4. The method according to claim 1 , wherein the determining, by a training apparatus, an initial training model further comprises:
determining, by the training apparatus, that the initial detection model comprises a feature extraction module and a detection and classification module, and that the adaptive model comprises a domain classification module, the detection and classification module being configured to perform classification of whether a lesion target is present according to a feature extracted by the feature extraction module, the domain classification module being configured to perform domain classification according to the feature extracted by the feature extraction module;
determining, by the training apparatus, an initial value of a parameter in the initial detection model and the adaptive model; and
the adjusting, by the training apparatus, a parameter value in the initial training model according to the loss function value further comprises: adjusting, by the training apparatus, the initial value of the parameter according to the loss function value.

5. The method according to claim 4 , wherein the adaptive model further comprises a gradient inversion module; and
the gradient inversion module is configured to transfer the feature extracted by the feature extraction module to the domain classification module, and is further configured to reverse, if the parameter value is adjusted, an error of the domain classification module, to adjust the parameter value in the feature extraction module according to the reversed error.

6. The method according to claim 4 , wherein
the feature extraction module comprises an intermediate feature extraction module and a final feature extraction module, then the domain classification module is configured to perform the domain classification according to a feature extracted by the intermediate feature extraction module or the final feature extraction module; and
the adaptive model comprises one or more domain classification modules.

7. The method according to claim 1 , wherein the training apparatus stops adjusting the parameter value in a case that the adjustment to the parameter value meets any one of the following stop conditions:
a quantity of times of the adjustment to the parameter value reaches a preset quantity of times; and
a difference between a currently adjusted parameter value and a last adjusted parameter value is less than a threshold.

8. The method according to claim 1 , comprising:
determining, by the training apparatus, a to-be-examined user body organ image, and determining, according to the final detection model, whether a lesion target is present in the to-be-examined user body organ image.

9. A non-transitory storage medium, the storage medium storing a plurality of instructions, the instructions being adapted to be loaded by a processor and causing the processor to
determine an initial training model, the initial training model comprising an initial detection model and an adaptive model;
determine a training sample, the training sample comprising source domain data and target domain data, the source domain data comprising a plurality of first user body organ images, each first user body organ image comprising: a first identifier of whether a lesion target is present, and a second identifier of a domain that the each first user body organ image belongs to; the target domain data comprising: a plurality of second user body organ images, and a third identifier of a domain that each second user body organ image belongs to;
separately determine whether a lesion target is present in the each first user body organ image through the initial detection model according to a feature of the each first user body organ image, to obtain a detection result; and separately determine a domain that each user body organ image in the training sample belongs to through the adaptive model according to a feature of the each user body organ image, to obtain a domain classification result;
calculate a loss function va

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Cheng Chen
• Sun Zhongqian
• Chen Hao
• Yang Wei

Classifications

• G06V10/82

  using neural networks · CPC title

• G06V10/454

  Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN] · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/24

  Classification techniques · CPC title

• G06F18/214 Primary

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title