US 12,437,426 B2
Image registration method and model training method thereof
Chen Zhang, Guangzhou (CN); Zhoushe Zhao, Beijing (CN); and Yingbin Nie, Guangzhou (CN)
Assigned to GE Precision Healthcare LLC, Waukesha, WI (US)
Filed by GE Precision Healthcare LLC, Waukesha, WI (US)
Filed on Dec. 13, 2023, as Appl. No. 18/538,489.
Application 18/538,489 is a division of application No. 17/343,017, filed on Jun. 9, 2021, granted, now 11,880,988.
Claims priority of application No. 202010530181.4 (CN), filed on Jun. 11, 2020.
Prior Publication US 2024/0119611 A1, Apr. 11, 2024
Int. Cl. G06T 7/33 (2017.01); G06T 3/4007 (2024.01); G06T 7/32 (2017.01)
CPC G06T 7/337 (2017.01) [G06T 3/4007 (2013.01); G06T 7/32 (2017.01); G06T 2207/20081 (2013.01)] 3 Claims
OG exemplary drawing
 
1. A method for training a non-rigid registration model for images, comprising:
(a) inputting a preprocessed reference image and a preprocessed floating image to a U-Net to obtain spatial transformation parameters, wherein the preprocessed reference image is obtained using rigid registration based on an iterative closest point registration and a mutual information registration, wherein rigid registration includes performing coarse registration about contour point data sets on the preprocessed reference image and the preprocessed floating image to be registered by using iterative closest point registration, and a contour point data set of each of the preprocessed reference image and the preprocessed floating image is extracted through a marching cubes algorithm;
(b) inputting the spatial transformation parameters to a spatial transformation network, and performing spatial transformation and an interpolation operation on the preprocessed floating image, so as to obtain a registration result image;
(c) calculating a loss function value between the preprocessed reference image and the registration result image by using a loss function, wherein the loss function comprises both a correlation coefficient and a mean squared error between the preprocessed reference image and the registration result image; and
(d) repeating steps (a) to (c) a predetermined number of times for iterative training or until the non-rigid registration model converges.