Tomographic Image Reconstruction With Sinogram Feedback Regularization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep image prior (DIP) techniques for noise reduction in tomographic imaging suffer from image quality degradation due to overtraining of convolutional neural networks (CNNs), leading to unwanted reconstruction of random noise.
Innovation Solution
An image processing apparatus and method that incorporates a sinogram creation unit, CNN processing unit, forward projection calculation unit, and CNN training unit, utilizing an evaluation function with error and regularization terms to train the CNN based on the difference between calculated and measured sinograms, thereby reducing noise and suppressing image quality degradation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the CNN is trained repeatedly using the DIP technique to reduce noise in the tomographic image, then the noise reduction performance is improved, but the image quality degrades due to overtraining of the CNN
Solution Approach 1:
The patent implements a feedback mechanism by calculating the error between the measured sinogram and the calculated sinogram (obtained through forward projection of the CNN output image) and using this error to iteratively train the CNN. This feedback loop allows the system to continuously improve noise reduction while monitoring image quality through the error metric, preventing overtraining by stopping when the error minimization converges or deteriorates.
Solution Approach 2:
The patent applies partial action by selectively applying the DIP technique only to the sinogram data rather than directly to the image domain, and by using forward projection (Radon transform) to create the calculated sinogram for comparison. This partial application in the sinogram space allows for controlled noise reduction without the excessive action of fully training the CNN on image data, which would cause overtraining and quality degradation.
2Measurement precision
If the forward projection calculation is performed from the CNN output image to create the calculated sinogram, then the error between measured and calculated sinograms can be evaluated, but the computational complexity increases
Solution Approach 1:
The patent replaces direct image-space optimization with sinogram-space optimization. Instead of directly minimizing image error through complex image processing operations, the system substitutes this with forward projection calculations (Radon transform) in the sinogram domain, which are computationally more efficient and provide a clearer error metric for guiding CNN training while reducing overall computational complexity.
Data Source
AI summary
An image processing apparatus includes a sinogram creation unit, a CNN processing unit, a convolution integration unit, a forward projection calculation unit, and a CNN training unit. The forward projection calculation unit performs forward projection calculation on an output image to create a calculated sinogram. The CNN training unit uses an evaluation function including an error evaluation term representing an evaluation value related to an error between a measured sinogram and the calculated sinogram and a regularization term representing an evaluation value related to a difference of pixel values between adjacent pixels in the output image, and trains the CNN based on a value of the evaluation function.