Tissue nodule detection and tissue nodule detection model training method, apparatus, device, and system

Patent metadata

Field	Value
Publication number	US-11880972-B2
Application number	US-202017093022-A
Country	US
Kind code	B2
Filing date	Nov 9, 2020
Priority date	Nov 8, 2018
Publication date	Jan 23, 2024
Grant date	Jan 23, 2024

Abstract

This application relates to a tissue nodule detection and tissue nodule detection model training method, apparatus, device, storage medium and system. The method for training a tissue nodule detection model includes: obtaining source domain data and target domain data, the source domain data comprising a source domain image and an image annotation, the target domain data comprising a target image, and the image annotation being used for indicating location information of a tissue nodule in the source domain image; performing feature extraction on the source domain image using a neural network model to obtain a source domain sampling feature, performing feature extraction on the target image using the neural network model to obtain a target sampling feature, and determining a model result according to the source domain sampling feature using the neural network model; determining a distance parameter between the source domain data and the target domain data according to the source domain sampling feature and the target sampling feature, the distance parameter being a parameter describing a magnitude of a data difference between the source domain data and the target domain data; determining, according to the model result and the image annotation, a loss function value corresponding to the source domain image; and training the neural network model to obtain a tissue nodule detection model by iteratively reducing a combination of the loss function value and the distance parameter. In this way, the detection accuracy can be improved.

First claim

What is claimed is:

1. A method for training a tissue nodule detection model, performed by a computer device, the method comprising:
obtaining source domain data and target domain data, the source domain data comprising a source domain image and an image annotation, the target domain data comprising a target image with no annotation, and the image annotation being used for indicating location information of a tissue nodule in the source domain image, wherein the source domain data is collected by a first type of device that is different from a second type of device that collects the target domain data, and wherein an image to be detected by the tissue nodule detection model is collected by the second type of device, wherein the first type of device and the second type of device are based on a same radiology technology and are different in at least one of following aspects: a brand name; a model; a sampling distance; a noise level; or a nodule diameter distribution;
performing feature extraction on the source domain image using a neural network model to obtain a source domain sampling feature, performing feature extraction on the target image using the neural network model to obtain a target sampling feature, and determining a model result according to the source domain sampling feature using the neural network model;
determining a distance parameter between the source domain data and the target domain data according to the source domain sampling feature and the target sampling feature, the distance parameter being a parameter describing a magnitude of a data difference between the source domain data and the target domain data;
determining, according to the model result and the image annotation, a loss function value corresponding to the source domain image; and
training the neural network model to obtain a tissue nodule detection model by iteratively reducing a combination of the loss function value and the distance parameter.

2. The method according to claim 1 , wherein the distance parameter comprises a maximum mean discrepancy based (MMD-based) discrepancy loss, and wherein determining the distance parameter between the source domain data and the target domain data according to the source domain sampling feature and the target sampling feature comprises determining the MMD-based discrepancy loss between the source domain data and the target domain data according to the source domain sampling feature and the target sampling feature.

3. The method according to claim 2 , wherein determining the MMD-based discrepancy loss between the source domain data and the target domain data according to the source domain sampling feature and the target sampling feature comprises:
determining, based on a Gaussian kernel function, the MMD-based discrepancy loss between the source domain data and the target domain data according to the source domain sampling feature and the target sampling feature.

4. The method according to claim 2 , wherein determining the MMD-based discrepancy loss between the source domain data and the target domain data according to the source domain sampling feature and the target sampling feature comprises:
determining a first MMD-based discrepancy loss between the source domain data and the target domain data according to the source domain sampling feature and the target sampling feature;
performing target region extraction on the source domain sampling feature, to obtain a source domain candidate region, and performing target region extraction on the target sampling feature, to obtain a target candidate region;
performing, after performing pooling processing on the source domain sampling feature and the source domain candidate region, mapping to obtain a source domain mapping result, and performing, after performing pooling processing on the target sampling feature and the target candidate region, mapping to obtain a target mapping result;
determining a second MMD-based discrepancy loss between the source domain data and the target domain data according to the source domain mapping result and the target mapping result; and
determining the MMD-based discrepancy loss between the source domain data and the target domain data according to the first MMD-based discrepancy loss and the second MMD-based discrepancy loss.

5. The method according to claim 1 , wherein training the neural network model to obtain the tissue nodule detection model comprises:
modifying the loss function value based on the distance parameter to generate a modified loss function value; and
training the neural network model to obtain the tissue nodule detection model based on iteratively reducing the modified loss function value.

6. The method according to claim 5 , wherein the distance parameter comprises a square of a maximum mean discrepancy (MMD) between the source domain data and the target domain data, and wherein modifying the loss function value based on the distance parameter to generate the modified loss function value comprises performing linear summation of the square of the MMD and the loss function value to obtain the modified loss function value.

7. The method according to claim 1 , wherein performing feature extraction on the source domain image using the neural network model to obtain the source domain sampling feature, and performing feature extraction on the target image using the neural network model to obtain the target sampling feature comprises:
segmenting the source domain image, to obtain a source domain tissue region, and segmenting the target image, to obtain a target tissue region; and
performing feature extraction on the source domain tissue region using the neural network model, to obtain the source domain sampling feature, and performing feature extraction on the target tissue region using the neural network model, to obtain the target sampling feature.

8. The method according to claim 1 , wherein the source domain image in the source domain data and the target image in the target domain data meet a predetermined quantity relationship.

9. The method according to claim 8 , wherein the source domain image in the source domain data and the target image in the target domain data are equal in quantity.

10. The method according to claim 1 , wherein performing feature extraction on the source domain image using the neural network model to obtain the source domain sampling feature, performing feature extraction on the target image using the neural network model to obtain the target sampling feature, and determining the model result according to the source domain sampling feature comprises:
performing feature extraction on the source domain image using a first neural network model, to obtain the source domain sampling feature, and determining the model result according to the source domain sampling feature; and
performing feature extraction on the target image using a second neural network model, to obtain the target sampling feature, the second neural network model and the first neural network model sharing a same weight.

11. A tissue nodule detection model training apparatus, comprising a memory for storing computer instructions and a processor in communication with the memory, wherein, when the processor executes the computer instructions, the processor is configured to cause the apparatus to:
obtain source domain data and target domain data, the source domain data comprising a source domain image and an image annotation, the target domain data comprising a target image with no annotation, and the image annotation being used for indicating location information of a tissue nodule in the source domain image, wherein the source domain data is collected by a first type of device that is differ

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Cheng Chen
• Sun Zhongqian
• Chen Zhao
• Yang Wei

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/096

  Transfer learning · CPC title

• G06T7/0012 Primary

  Biomedical image inspection · CPC title

• G06F18/213

  Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods · CPC title