PET-CT Image Reconstruction Using DVF Motion Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Nuclear imaging systems face challenges such as subject movement during scans leading to misalignment and inaccurate attenuation correction, resulting in lower quality medical images due to approximations used to compensate for detection loss.
Innovation Solution
Employing machine learning processes, specifically deep learning neural networks, to generate displacement vector field (DVF) data that characterizes offsets between PET and modality measurement data, allowing for accurate image alignment and correction, including attenuation and scatter corrections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image reconstruction methods are used, then the processing is simpler and faster, but the image alignment accuracy deteriorates due to subject movement
Solution Approach 1:
The patent replaces conventional mathematical algorithms with deep learning-based neural networks to perform image alignment. The neural network automatically learns complex non-linear relationships between PET and CT images, substituting traditional mechanical/mathematical registration methods with an intelligent system that achieves superior alignment accuracy while handling subject movement dynamically.
Solution Approach 2:
The patent transforms the image registration problem by changing the parameter space from traditional geometric transformations to a deep learning representation. The neural network processes images through multiple layers, converting pixel data into feature representations that capture complex spatial relationships, enabling accurate alignment without explicit geometric parameter manipulation.
2Reliability
If approximations are used to compensate for detection loss, then the processing is more robust, but the image quality deteriorates
Solution Approach 1:
The patent replaces traditional approximation-based correction methods with deep learning neural networks that directly model the relationship between PET measurements and anatomical structures. This substitution eliminates the need for compensatory approximations while maintaining robustness through the network's ability to learn from training data and handle various imaging conditions.
Solution Approach 2:
The neural network is trained using synthetic data that copies and augments the complex relationships between PET and CT images. During inference, the trained network reproduces accurate anatomical alignments by copying the learned patterns from training data, achieving high image quality without relying on approximate correction formulas.
3Measurement precision
If deep learning-based image alignment is applied, then the image alignment accuracy and quality improve, but the processing time and computational resources increase
Solution Approach 1:
The neural network is pre-trained using extensive synthetic training data generated before actual imaging sessions. This preliminary training allows the network to capture complex alignment relationships in advance, so that during actual PET/CT processing, the network can quickly apply learned patterns without performing time-consuming computations on each new image pair.
Solution Approach 2:
The trained neural network copies pre-learned feature representations and spatial relationships from training data to process new PET/CT image pairs. This copying mechanism enables fast inference where the network simply applies learned transformations rather than performing complex real-time calculations, significantly reducing processing time while maintaining high alignment accuracy.
Data Source
AI summary
Systems and methods for reconstructing medical images based on motion estimation are disclosed. Measurement data from positron emission tomography (PET) measurement data, and modality measurement data from an anatomy modality, such as computed tomography (CT) data, is received from an image scanning system. A trained deep learning process is applied to the PET measurement data and the modality measurement data to generate displacement vector field (DVF) data characterizing motion between the PET measurement data and the modality measurement data. A modality image is reconstructed from the modality measurement data, and the modality image is adjusted based on the DVF data. A PET image is then reconstructed from the PET measurement data and the adjusted modality image, and the PET image is adjusted based on a computed inverse of the DVF data. The adjusted PET image and the modality image spatially match, and are displayed.