GroupRegNet: a groupwise one-shot deep learning-based 4D image registration method

Abstract

Accurate deformable four-dimensional (4D) (three-dimensional in space and time) medical images registration is essential in a variety of medical applications. Deep learning-based methods have recently gained popularity in this area for the significantly lower inference time. However, they suffer from drawbacks of non-optimal accuracy and the requirement of a large amount of training data. A new method named GroupRegNet is proposed to address both limitations. The deformation fields to warp all images in the group into a common template is obtained through one-shot learning. The use of the implicit template reduces bias and accumulated error associated with the specified reference image. The one-shot learning strategy is similar to the conventional iterative optimization method but the motion model and parameters are replaced with a convolutional neural network and the weights of the network. GroupRegNet also features a simpler network design and a more straightforward registration process, which eliminates the need to break up the input image into patches. The proposed method was quantitatively evaluated on two public respiratory-binned 4D-computed tomography datasets. The results suggest that GroupRegNet outperforms the latest published deep learning-based methods and is comparable to the top conventional method pTVreg. To facilitate future research, the source code is available at https://github.com/vincentme/GroupRegNet.

End-to-End Deep Learning of Non-rigid Groupwise Registration and Reconstruction of Dynamic MRI

“…There are in general three different types of groupwise registration: sum-of-pairs approach that attempts to reduce the registration loss among all image pairs; reference-based approach that requires the designation of one image as reference; implicit template approach that implicitly determines the template image during registration, and can avoid the bias caused by selecting one particular image as reference while being more computationally efficient than the sum-of-pairs method ( 41 ). Deep learning groupwise registration has been adopted in several recent studies ( 36 , 37 , 42 ) and has demonstrated superior performance over pairwise registration. For example, Zhang et al propose an one-shot learning groupwise registration network to register respiratory motion-resolved 3D CT images ( 36 ).…”

“…Deep learning groupwise registration has been adopted in several recent studies ( 36 , 37 , 42 ) and has demonstrated superior performance over pairwise registration. For example, Zhang et al propose an one-shot learning groupwise registration network to register respiratory motion-resolved 3D CT images ( 36 ). Martín-González et al ( 42 ) develop a deep learning framework to achieve groupwise registration of 2D dynamic sequence, in which the implicit template deep learning groupwise registration approach is adopted to estimate the nonrigid motion across the dynamic sequence.…”

“…The network is optimized via the intensity-based similarity (mean-squared-error, MSE) and velocity field smoothness loss: where the first term enforces the similarity between the implicit reference image and the warped dynamic images; the second term imposes the intensity similarity between the original dynamic images and the generated dynamic images by transforming the template with the inverse transformations ( ); the third term is to encourage the smoothness of the estimated velocity fields and α is the regularization weight. The implicit template image is defined as the average of the warped dynamic images ( 36 , 37 ): In addition, to enforce the reference image to lie in the geometric center of the group, as proposed in Li et al ( 37 ), the average velocity field is subtracted from each of the estimated velocity fields: , and the sum of all the velocity fields is consequently zero.…”

Lung‐CRNet: A convolutional recurrent neural network for lung 4DCT image registration

“…Some studies have proposed to use the convolutional neural network as a feature extractor for registering multiple 3DCT images. 20,21 In essence, their one-shot methods are traditional iterative optimization methods since their networks still need to be optimized when registering new images. The way they use the network is equivalent to taking the input as the training sample as well as the test sample instead of learning from other samples, which is prone to overfitting and lacks stability.…”

An AI Approach for Reducing Residual Motion and Noise from Myocardial Perfusion for Assessment of Coronary Artery Disease

“…Image registration is the process of aligning the images together that are misaligned due to organ or patient movement. We have studied a few learning-based denoising and registration methods where individuals are trying to solve these two problems independently [3][4][5][6][7][8]. But it is essential to jointly look at these two issues together instead of dealing with each one separately.…”