Denoising Diffusion Model for Metal Artifact Reduction in CT Sinograms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for metal artifact reduction in medical images, such as CT scans, are limited in effectively removing artifacts caused by metal objects, which can obscure anatomical details and reduce diagnostic confidence and accuracy, especially when severe artifacts are present, and existing AI models like CNNs and GANs require specific training data and may not generalize well to unknown metal objects or geometries.
Innovation Solution
A denoising diffusion probabilistic model (DDPM) is used to generate a corrected sinogram by delineating corrupted regions and inpainting them with uncorrupted data, allowing for the reduction of artifacts in CT images, and is trained on data without metal objects to enhance generalization across various metal types and geometries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional MAR techniques or AI models (CNNs, GANs) are used to reduce metal artifacts, then artifact removal capability is improved, but the model requires specific training data and fails to generalize well to unknown metal objects or geometries
Solution Approach 1:
The patent uses a denoising diffusion probabilistic model that learns the general distribution of clean CT images and metal artifacts during training. During inference, the model generates corrupted sinograms by adding noise to clean images and then denoises them, effectively copying the pattern of metal artifacts without requiring specific training data for each metal type. This allows the model to generalize to unknown metal objects and geometries by learning from the statistical characteristics of artifacts rather than specific examples.
Solution Approach 2:
The patent transforms the problem from image domain to sinogram domain, changing the parameters of the data representation. By working with sinograms instead of directly processing CT images, the model can effectively handle metal artifacts through parameter transformation. The diffusion process operates on these transformed parameters, enabling the model to remove artifacts by reversing the forward diffusion process that generated the corruption.
2Reliability
If deep learning models are trained on data with metal objects to remove artifacts, then artifact reduction is improved, but training complexity and data requirements increase
Solution Approach 1:
Instead of training the model on corrupted images with metal artifacts, the patent inverts the approach by training on clean images without metal objects. The model learns the distribution of clean CT images and then uses this learned distribution to generate and denoise corrupted sinograms during inference. This inversion eliminates the need for complex training data with metal artifacts, simplifying the training process while maintaining effective artifact removal capability.
Solution Approach 2:
The model performs self-service by generating its own training data through the diffusion process. During training, the model learns to denoise random noise patterns into clean images. During inference, it automatically generates appropriate corrupted sinograms by adding noise to clean images and then denoises them, eliminating the need for external training data with metal artifacts. The model serves its own training needs by generating synthetic corrupted data from clean data.
Data Source
AI summary
Various systems and methods are provided for MAR in CT images. A corrupted CT image, of a region of interest (ROI) of a subject, including artifacts caused by a metal object in the subject may be acquired. A corrupted sinogram including a corrupted region of corrupted data caused by the metal object and an uncorrupted region of uncorrupted data may be generated. A mask sinogram that delineates the corrupted region of the corrupted data may be generated. A corrected sinogram including the uncorrupted region of the uncorrupted data and an inpainted region of inpainted data corresponding to the corrupted region may be generated using a denoising diffusion probabilistic model, the corrupted sinogram, and the mask sinogram. A corrected CT image, of the ROI of the subject, that includes reduced artifacts relative to the artifacts in the corrupted CT image may be generated based on the corrected sinogram.