US 12,444,169 B2
Wasserstein distance and difference metric-combined chest radiograph anomaly identification domain adaptation method and system
Bishi He, Hangzhou (CN); Zhe Xu, Hangzhou (CN); Yuanjiao Chen, Hangzhou (CN); Diao Wang, Hangzhou (CN); and Hui Chen, Hangzhou (CN)
Assigned to Hangzhou Dianzi University, Hangzhou (CN)
Filed by Hangzhou Dianzi University, Hangzhou (CN)
Filed on Aug. 10, 2023, as Appl. No. 18/447,328.
Claims priority of application No. 202211393538.4 (CN), filed on Nov. 8, 2022.
Prior Publication US 2024/0153243 A1, May 9, 2024
Int. Cl. G06V 10/764 (2022.01); G06T 7/00 (2017.01)
CPC G06V 10/765 (2022.01) [G06T 7/0012 (2013.01); G06T 2207/10081 (2013.01); G06T 2207/20076 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/30061 (2013.01); G06V 2201/031 (2022.01)] 10 Claims
OG exemplary drawing
 
1. A Wasserstein distance and difference metric-combined chest radiograph anomaly identification domain adaptation method, comprising the following steps:
step 1, data preparation and data pre-processing for chest radiographs:
obtaining computed tomography (CT) images of the chest radiographs and performing the data pre-processing on the CT images;
step 2, multi-scale feature extraction based on a swin transformer network:
inputting the CT images into a patch partition module to perform partitioning, setting 4×4=16 adjacent pixels as a patch, flattening the pixels in a channel direction, performing linear transformation on channel data of each of the pixels through a linear embedding layer, and then constructing feature maps with different sizes through four stages;
step 3, loss minimization based on a Wasserstein distance and a contrastive domain discrepancy:
selecting source domain samples closest to target domain samples and calculating the Wasserstein distance; performing cross-domain similar class approaching and dissimilar class splitting on the target domain samples to obtain the contrastive domain discrepancy; constructing a total objective function through the Wasserstein distance and the contrastive domain discrepancy; and performing optimization and parameter update on the swin transformer network based on the total objective function to obtain a model; and
step 4, using the model to perform chest radiograph prediction after verifying the model:
verifying the optimized and updated swin transformer network and performing a classification prediction task for the chest radiographs.