المفاهيم الأساسية
A score-based diffusion model trained on motion-free CT images can be used to estimate the likelihood of motion-affected images, enabling gradient-based optimization of motion parameters to reduce artifacts.
الملخص
The paper presents a method for CT motion compensation that is trained solely on clean, motion-free CT images. The key idea is to train a score-based diffusion model to learn the distribution of motion-free head CT images. This trained model can then be used to estimate the likelihood of a given, potentially motion-affected CT image. The likelihood value serves as a surrogate metric for motion artifact severity, allowing for gradient-based optimization of the underlying motion parameters to bring the image closer to the distribution of motion-free scans.
The method consists of the following steps:
- Train a score-based diffusion model on a dataset of clean, motion-free head CT images.
- Construct a differentiable likelihood function using the trained score model, a neural ODE solver, and the Hutchinson trace estimator.
- Optimize the motion parameters (translations and rotation) by iteratively reconstructing the CT image, evaluating the likelihood of the reconstruction, and updating the motion parameters to maximize the likelihood.
The proposed approach achieves comparable performance to state-of-the-art methods that require a representative dataset of motion-affected images for training. By only using clean images, the method is more robust to unforeseen motion patterns in real-world applications.
الإحصائيات
The method is evaluated on publicly available head CT scans from the CQ500 dataset.
The score network is trained on slices from 200 subjects and evaluated on slices from 40 subjects.
Motion compensation experiments are performed on slices from another 40 disjunct subjects.
اقتباسات
"Motion artifacts can compromise the diagnostic value of computed tomography (CT) images. Motion correction approaches require a per-scan estimation of patient-specific motion patterns."
"We aim to emulate the human observer with a network that identifies motion artifacts after training on clean images only."
"Our approach achieves comparable performance to state-of-the-art methods while eliminating the need for a representative data set of motion-affected samples."