Unsupervised Deformable Image Registration via Cycle-Consistent Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image registration methods, especially AI-based techniques, face challenges in efficiently registering large three-dimensional medical images like CT images due to limited topology preservation and high computational requirements, leading to suboptimal registration performance.
Innovation Solution
An unsupervised learning-based neural network with cycle consistency is employed to generate a registration image by outputting a deformation field for image registration, using a spatial deformation function to transform moving images into a fixed coordinate system, trained on cyclic loss, registration loss, and identity loss without requiring ground-truth deformation fields.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If diffeomorphic image registration method is used to preserve topology and ensure one-to-one mapping, then topology preservation is improved, but computational time and resources increase substantially
Solution Approach 1:
The patent replaces the mechanical iterative optimization process of traditional diffeomorphic registration with a neural network-based system. The neural network learns the deformation field directly from image pairs, substituting the iterative mechanical optimization with a learned mapping that achieves topology preservation without the computational burden of iterative methods.
Solution Approach 2:
The patent performs preliminary training of the neural network on large datasets of image pairs before actual registration. This preliminary learning phase allows the network to pre-learn deformation patterns and topology-preserving transformations, so that during actual registration, the computation is significantly reduced while maintaining topology preservation.
2Productivity
If AI-based image registration technique is used to reduce computational time, then registration speed is improved, but topology preservation constraint becomes unclear
Solution Approach 1:
The patent incorporates a feedback mechanism through the loss function that explicitly penalizes topology violations. The network receives feedback during training about whether its predicted deformations preserve topology, allowing it to learn to maintain one-to-one mappings while achieving fast registration speeds.
Solution Approach 2:
The patent changes the parameter space by representing deformations as displacement fields predicted by the neural network rather than as transformation matrices. This parameterization, combined with appropriate regularization terms in the loss function, enables the network to learn topology-preserving deformations efficiently.
3Adaptability or versatility
If AI-based image registration is applied to large three-dimensional CT images, then registration capability is improved, but registration performance becomes limited
Solution Approach 1:
The patent segments the large 3D CT volume into smaller sub-volumes or slices that can be processed independently or in a hierarchical manner. The neural network processes these smaller units and the results are combined to achieve full-volume registration, maintaining both adaptability to large images and high registration performance.
Solution Approach 2:
The patent processes 3D volumetric data by transforming it into a sequence of 2D slices or by using 3D convolutions that operate layer-by-layer through the volume. This dimensional decomposition allows the neural network to handle large 3D images effectively while maintaining registration accuracy through the preservation of spatial relationships across dimensions.
Data Source
AI summary
Disclosed are an unsupervised learning-based image registration method using a neural network with cycle consistency and an apparatus therefor. An image registration method includes receiving a first image and a second image for image registration, outputting a deformation field for the first image and the second image using an unsupervised learning-based neural network with cycle consistency for the deformation field, and generating a registration image for the first image and the second image based on a spatial deformation function using the output deformation field. The outputting of the deformation field includes outputting the deformation field for the first image for registering the first image to the second image may be output, when the first image is a moving image and the second image is a fixed image, and the generating of the registration image includes generating the registration image by applying the deformation field for the first image to the first image using the spatial deformation function.