Deep Learning Super-Resolution for Anisotropic MRI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for enhancing magnetic resonance image resolution face challenges such as trade-offs between resolution, acquisition time, and noise, with limitations in achieving high through-plane resolution due to anisotropic voxels and the need for multiple low-resolution image acquisitions, which are time-consuming and prone to artifacts.
Innovation Solution
The development of anti-aliasing and super-resolution techniques using deep learning networks that operate on single anisotropic data sets, eliminating the need for external training data and multiple image acquisitions, by applying these networks along different orientations and recombining outputs through Fourier burst accumulation, thereby enhancing through-plane resolution without deconvolution steps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high resolution magnetic resonance images are acquired with adequate signal-to-noise ratio, then image resolution is improved, but acquisition time increases and cost increases
Solution Approach 1:
The patent creates synthetic high-resolution training data by generating artificial image pairs through downsampling and blurring operations. Instead of requiring actual paired high-resolution and low-resolution images from multiple acquisitions, the system copies and transforms existing low-resolution images to create training datasets, eliminating the need for time-consuming multiple acquisitions.
Solution Approach 2:
The patent performs preliminary training of deep learning networks using synthetically generated data before applying them to actual medical images. This preliminary preparation allows the network to learn super-resolution mappings in advance, so that when processing real patient data, high-resolution images can be generated quickly from single low-resolution acquisitions without requiring multiple scans.
2Loss of time
If images are acquired with good in-plane resolution and poor through-plane resolution to reduce acquisition time, then acquisition time is reduced, but through-plane resolution deteriorates and aliasing artifacts are introduced
Solution Approach 1:
The patent changes the resolution parameters anisotropically, accepting poor through-plane resolution in the acquired images while using deep learning to restore and enhance through-plane resolution in the reconstructed images. The system processes images with voxel dimensions such as 0.5mm×0.5mm×2mm, where the through-plane dimension is intentionally undersampled, then uses trained networks to recover fine through-plane details that would otherwise be lost.
3Measurement precision
If multiple low-resolution image acquisitions are performed to achieve high resolution, then image resolution can be improved, but acquisition time increases and motion artifacts increase
Solution Approach 1:
The patent generates synthetic training pairs by copying and transforming single low-resolution images through mathematical operations (downsampling, blurring) rather than requiring multiple physical acquisitions. This creates artificial high-resolution references from single scans, allowing the network to learn from synthesized data without exposing patients to motion artifacts from repeated scanning.
4Measurement precision
If state-of-the-art machine-learning based super-resolution methods are used, then through-plane resolution is improved, but external training data consisting of paired low resolution and high resolution images are required which are often unavailable due to scanner limitations
Solution Approach 1:
The patent creates training data by copying and mathematically transforming existing low-resolution images into synthetic high-resolution versions. Through downsampling and controlled blurring operations, the system generates artificial image pairs without requiring actual high-resolution scans, thus eliminating the need for specialized scanner capabilities or external datasets.
Solution Approach 2:
The system is self-sufficient in generating its own training data from any available low-resolution medical images. The deep learning network is trained using synthetic pairs derived from the same imaging modality and scanner, making the system adaptable to different scanners and protocols without requiring external training datasets or specialized hardware.
Data Source
AI summary
Disclosed techniques for image processing three-dimensional image data include: obtaining three-dimensional image data representing contiguous slices parallel to a plane, constructing training data from the image data by, for each of a plurality of angles: rotating the image data in the plane to produce rotated image data, blurring the rotated image data in a dimension parallel to the plane to produce low resolution rotated image data, and introducing aliasing into the low resolution rotated image data in the dimension parallel to the plane to produce aliased low resolution rotated image data, training an anti-aliasing neural network with the aliased low resolution image data and the low resolution image data, training a super-resolution neural network with the aliased low resolution image data and the rotated image data, and processing the image data using the trained anti-aliasing neural network and the trained super-resolution neural network to produce processed image data.