PET-CT Registration via Deep Learning Deformation Field
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Misalignment between PET and CT images in PET/CT scans leads to inaccurate quantification and artifacts, as existing registration techniques are inefficient and require significant computational resources, especially when dealing with respiratory motion and anatomical specificity.
Innovation Solution
A machine-learning model is employed to generate a deformation field for spatial alignment of CT and PET data, using a generative adversarial network trained on attenuation-corrected PET data to improve registration accuracy and efficiency, allowing for accurate attenuation correction in PET reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If classic mutual information registration approaches are used, then registration is performed, but the effect is limited and computational efficiency is poor
Solution Approach 1:
The patent replaces traditional mechanical registration approaches (mutual information, optimization-based methods) with a deep learning-based system. A neural network model is trained to directly predict transformation parameters that align PET and CT images, substituting iterative computational mechanics with a trained predictive model that achieves both high accuracy and efficiency.
Solution Approach 2:
The system performs preliminary training of the deep learning model using paired PET-CT data with known transformations. This pre-computed knowledge is stored in the trained model weights, allowing rapid inference during actual registration without repeating complex optimization calculations, thus improving computational efficiency while maintaining accuracy.
2Measurement precision
If integrated reconstruction techniques such as MLRR are used, then attenuation correction is improved, but image noise increases and reconstruction time increases
Solution Approach 1:
The patent separates the attenuation correction process into distinct stages: first, a deep learning model predicts transformation parameters to align PET and CT; second, the aligned CT is used to generate an attenuation map; third, the attenuation map is applied to correct PET data. This segmentation avoids the iterative joint optimization of MLRR, reducing reconstruction time while maintaining correction accuracy.
Solution Approach 2:
The system uses a pre-trained deep learning model that has learned optimal transformation patterns from training data. During reconstruction, instead of performing complex iterative optimization, the system copies the learned transformation parameters from the trained model to achieve accurate alignment and attenuation correction more efficiently.
3Measurement precision
If consistency conditions optimization approaches are used, then attenuation correction is attempted, but computation is inefficient and requires very powerful hardware
Solution Approach 1:
The patent replaces complex optimization-based consistency condition enforcement with a deep learning approach. The neural network is trained to directly output transformation parameters that satisfy alignment consistency, eliminating the need for powerful hardware to solve complex optimization problems during reconstruction.
Solution Approach 2:
The system uses a trained deep learning model that can be deployed on standard hardware. Instead of requiring powerful computing resources during reconstruction to enforce consistency conditions, the system uses the pre-trained model's learned knowledge, which can be applied efficiently on less powerful hardware, reducing device complexity requirements.
4Stability of the object's composition
If respiratory motion compensation techniques are applied to both PET and CT separately, then motion artifacts are reduced, but registration between PET and CT remains problematic
Solution Approach 1:
The patent merges the motion compensation and registration processes into a unified deep learning framework. The model simultaneously learns to compensate for respiratory motion in both PET and CT data and to compute the transformation parameters that align the two modalities, ensuring anatomical consistency and alignment accuracy work together rather than conflicting.
Solution Approach 2:
The system changes the approach from separate motion compensation with subsequent registration to a joint learning approach where transformation parameters are learned directly from paired PET-CT data. The model learns optimal parameter transformations that account for respiratory motion while maintaining alignment, adapting to anatomical variations more effectively.
Data Source
AI summary
CT and PET are registered, providing a spatial alignment to be used in attenuation correction for PET reconstruction. A model for machine learning is defined to generate a deformation field. The model is trained with loss based, in part, on the attenuation corrected PET data rather than or in addition to loss based on the uncorrected PET or the generated deformation field. Due to the nature of the mapping from CT to attenuation, a separate, pre-trained network is used to form the attenuation corrected PET data in training the model.