CT Image Registration Neural Network Using Deformation Field Decomposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for CT image registration, particularly for lung CT images, are computationally intensive, require additional information that is difficult to obtain, or are expensive to train, due to the large deformations caused by breathing, making them inefficient and error-prone.
Innovation Solution
A training method for a deep neural network model that decomposes the deformation field into two parts using a uniform distribution, generates intermediate images, and updates the model by minimizing a total loss function, incorporating surrogate supervision to improve registration accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a multi-stage strategy is used to optimize coarse-to-fine registration, then registration accuracy is improved, but computational resource consumption increases significantly
Solution Approach 1:
The deformation field is segmented into multiple scales using a pyramid structure, where coarse-to-fine registration is achieved by processing images at different resolution levels. This segmentation allows the system to capture both global and local deformation characteristics without requiring computationally expensive multi-stage processing at full resolution throughout.
Solution Approach 2:
The patent introduces a scale dimension by constructing image pyramids with multiple resolution levels. Instead of performing iterative refinement in the spatial domain only, the system adds a resolution scale dimension, allowing efficient coarse-to-fine registration by processing from coarse to fine scales, thereby reducing computational burden while maintaining accuracy.
2Measurement precision
If external supervision such as lung masks and landmarks is used to enhance registration, then registration accuracy is improved, but the difficulty of obtaining additional information increases
Solution Approach 1:
The system performs self-service by automatically extracting anatomical structures and generating supervision signals from the input images themselves, without requiring external annotations. The network learns to identify lung boundaries and internal structures autonomously, converting the input images into their own supervision signals through self-supervised learning mechanisms.
Solution Approach 2:
The patent introduces an intermediary anatomical structure extraction module that automatically derives supervision signals from the input images. This intermediary component generates pseudo-labels and structural constraints that mediate between the raw images and the registration objective, eliminating the need for direct external annotation while providing necessary guidance for accurate registration.
3Measurement precision
If recursive cascade networks with multiple cascade networks are used, then registration accuracy is improved, but training complexity and computational cost increase
Solution Approach 1:
Multiple cascade networks are merged into a single unified deep neural network architecture. Instead of training separate cascade networks sequentially, the patent combines their functional components into one integrated network that processes all deformation scales simultaneously, thereby reducing training complexity while preserving the hierarchical deformation modeling capability.
Solution Approach 2:
The network performs preliminary deformation estimation at coarse scales during the same forward pass rather than requiring sequential training stages. By pre-computing deformation fields at multiple scales within a single training iteration, the system eliminates the need for multi-stage training procedures, reducing overall training complexity while maintaining accurate coarse-to-fine registration performance.
Data Source
AI summary
A training method of a deep neural network model for CT image registration, comprising the following steps: providing features of a source image and features of a corresponding target image; generating a deformation field from the source image to the target image; decomposing the generated deformation field into a first deformation field part and a second deformation field part; obtaining a first registration result and a second registration result; generating a surrogate image; obtaining a total loss function, and inputting the total loss function into the deep neural network model, and adjusting the parameters of the deep neural network model until a predetermined number of iterations is achieved, so as to obtain the final deep neural network model.