CNN Artifact Reduction Using Pseudo Ground Truth for CT Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CT image artifact reduction techniques, such as interpolation and normalization-based methods, often fail to achieve satisfactory results, particularly in clinical applications like radiation and proton therapy planning, due to residual or introduced artifacts, especially with metal objects.
Innovation Solution
A method using a convolutional neural network (CNN) trained with an estimated ground truth image, generated by reducing artifacts in initial CT images, to create simulated training images with added features, allowing the CNN to learn artifact reduction without actual artifact-free data, thereby improving image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If interpolation and normalization-based artifact reduction methods are used, then processing speed is maintained, but image quality deteriorates due to residual and introduced artifacts
Solution Approach 1:
The patent replaces traditional mechanical interpolation and normalization methods with a deep learning-based CNN system. The network learns artifact patterns and removal strategies from training data, substituting deterministic algorithms with adaptive neural network processing that can generalize across different artifact types and imaging conditions.
Solution Approach 2:
The patent employs preliminary action by generating synthetic training data with known ground truth before actual artifact reduction is needed. The CNN is pre-trained on simulated data with controlled artifacts, allowing it to learn optimal removal strategies in advance rather than relying on post-hoc correction methods.
2Measurement precision
If traditional artifact reduction techniques are applied, then computational complexity remains low, but diagnostic accuracy deteriorates in challenging cases
Solution Approach 1:
The patent substitutes simple interpolation algorithms with a sophisticated CNN architecture that can capture complex artifact patterns. The network uses multiple convolutional layers with ReLU activation and batch normalization to learn hierarchical features, enabling accurate artifact removal in challenging clinical scenarios despite increased computational requirements.
Solution Approach 2:
The patent changes key parameters of the artifact reduction process by introducing learnable weights and biases in the CNN, replacing fixed algorithmic parameters. The network adapts its internal parameters during training to optimize performance for specific artifact types and imaging conditions, achieving superior diagnostic accuracy.
3Manufacturing precision
If more aggressive artifact reduction is applied, then image quality improves, but new artifacts are introduced by the correction method
Solution Approach 1:
The patent implements feedback mechanisms through the CNN's learned prediction of artifact-free regions. The network continuously refines its artifact removal predictions by comparing processed outputs with ground truth during training, learning to avoid over-correction that introduces new artifacts. The feedback loop enables adaptive adjustment of correction intensity.
Solution Approach 2:
The patent introduces an intermediary CNN model that acts as a mediator between the original artifact-containing image and the final corrected image. Rather than directly applying aggressive correction, the network gradually transforms the image through learned operations, preventing the introduction of new artifacts while maintaining effective artifact removal.
Data Source
AI summary
Training a CNN with pseudo ground truth for CT artifact reduction is described. An estimated ground truth apparatus is configured to generate an estimated ground truth image based, at least in part, on an initial CT image that includes an artifact. Feature addition circuitry is configured to add a respective feature to each of a number, N, copies of the estimated ground truth image to create the number, N, initial training images. A computed tomography (CT) simulation circuitry is configured to generate a plurality of simulated training CT images based, at least in part, on at least some of the N initial training images. An artifact reduction circuitry is configured to generate a plurality of input training CT images based, at least in part, on the simulated training CT images. A CNN training circuitry is configured to train the CNN based, at least in part, on the input training CT images and based, at least in part, on the initial training images.