CNN Ringing Correction for MRI via Dimensional Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI ringing correction technologies face challenges in effectively correcting ringing artifacts in both two-dimensional and three-dimensional images, particularly due to the high computational load and long processing times required for training convolutional neural networks (CNNs) with high-resolution images.
Innovation Solution
The proposed solution involves training a CNN to perform ringing correction for a lower-dimensional direction than the image being corrected, and applying this CNN in multiple stages to achieve effective ringing correction. This approach reduces the burden of training the CNN and allows for efficient correction of ringing artifacts in both two-dimensional and three-dimensional images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a CNN is trained using high-resolution images with large measurement matrix sizes to achieve accurate ringing correction, then the ringing correction accuracy is improved, but the training time and computational load increase significantly
Solution Approach 1:
The patent transforms the training problem from three-dimensional image data to two-dimensional image data by applying the CNN in multiple stages. Specifically, a CNN trained on 2D images is applied sequentially to correct different dimensions of 3D images, reducing the training dimensionality while maintaining correction effectiveness. This dimensional transformation allows training on lower-dimensional data that is faster and less computationally intensive.
Solution Approach 2:
The patent segments the ringing correction process into multiple stages, where each stage corrects ringing in a specific dimension. Instead of training a single CNN on complete 3D high-resolution images, the method divides the correction task across multiple lower-dimensional stages, making the training process more manageable and time-efficient while achieving comprehensive ringing correction.
2Measurement precision
If the measurement matrix size is increased to obtain high-resolution images for training, then the image resolution and training data quality are improved, but the imaging time increases
Solution Approach 1:
The patent applies partial action by training the CNN on 2D image data rather than complete 3D high-resolution images. This partial training approach uses a subset of the information (2D slices) that is sufficient to learn ringing correction patterns, avoiding the need to acquire and process complete high-resolution 3D images for training, thus reducing imaging time requirements.
3Object-generated harmful factors
If a k-space filter is applied to suppress ringing artifacts, then the ringing is reduced, but the image sharpness decreases and the image becomes blurred
Solution Approach 1:
The patent replaces the traditional k-space filter approach (which operates in frequency space and blurs images) with a CNN-based method that operates in real space. The CNN learns to identify and correct ringing artifacts directly in the image domain, preserving image sharpness while effectively suppressing ringing artifacts through learned transformations rather than simple filtering.
Data Source
AI summary
Provided is a technology capable of effectively obtaining a ringing correction effect even for a two-dimensional image or a three-dimensional image with a simple CNN configuration.A CNN that has been trained to perform ringing correction for a direction of a dimension lower than a dimension of an image that is a correction target is prepared, and the CNN is applied in multiple stages to perform the ringing correction. For training the CNN, an image captured by increasing a measurement matrix size in one or two directions need only be used, thereby reducing an imaging time for acquiring training data and a burden of data processing, and enabling handling of images of various dimensions.