Isotropic 3D Image Reconstruction via Patch-Based Self-Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Magnetic Resonance Imaging (MRI) techniques face challenges in producing high-resolution, isotropic 3D cardiac images due to hardware and time limitations, resulting in poor image quality, especially in the through-plane direction, and existing super-resolution methods are computationally intensive and prone to noise amplification.
Innovation Solution
A method combining denoising and super-resolution reconstruction using an Augmented Lagrangian algorithm, which integrates self-similarity learning into a geometrical framework, allowing for the merging of information from multiple anisotropic volumes and recovery of sharp edges and thin anatomical structures, while reducing computational time through multithreading and GPU acceleration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple anisotropic acquisitions are combined using super-resolution techniques, then isotropic 3D image quality is improved, but computational complexity and noise amplification increase
Solution Approach 1:
The patent segments the image reconstruction problem into local patch processing units. Instead of processing the entire 3D volume at once, the method divides it into overlapping patches that can be independently processed and then combined. This segmentation reduces computational complexity while maintaining image quality by enabling parallel processing and reducing the memory burden of handling large volumetric data.
Solution Approach 2:
The patent uses self-similarity by copying and comparing patches from different locations and orientations within the multi-acquisition data. By identifying and copying similar patches from various anisotropic acquisitions, the method reconstructs high-resolution isotropic volumes without requiring equally complex computational resources, as it leverages redundant information already present in the acquired data.
2Reliability
If local regularization techniques such as Total Variation are applied, then noise amplification is limited and sharp edges are preserved, but image smoothness and detail recovery are compromised
Solution Approach 1:
The patent merges local regularization (Total Variation) with non-local self-similarity methods. By combining these two approaches, the method maintains the noise control and edge preservation benefits of local regularization while adding the detail recovery capabilities of non-local methods that exploit similarities across different regions of the image. This hybrid approach resolves the contradiction by achieving both noise control and detailed structure recovery.
Solution Approach 2:
The patent creates a composite regularization approach by integrating TV regularization with self-similarity constraints. This composite method combines the strengths of different regularization techniques, using TV for local noise control and edge preservation, while using self-similarity for global structure recovery and detail enhancement, thereby achieving superior overall performance compared to either method alone.
3Measurement precision
If high spatial resolution 3D isotropic cardiac images are acquired, then diagnostic quality is improved, but scan time exceeds patient breath-holding capacity
Solution Approach 1:
The patent applies preliminary action by acquiring multiple lower-resolution anisotropic volumes during the patient's breath-hold period, rather than attempting to acquire a single high-resolution isotropic volume. The super-resolution reconstruction is then performed as a post-processing step, effectively performing the high-resolution acquisition in advance through computational methods, thereby avoiding the need to extend the scan time beyond breath-holding capacity.
Solution Approach 2:
The patent replaces the mechanical/physical limitation of scan time with a computational solution. Instead of relying on longer acquisition times to achieve high resolution, the method uses computational super-resolution algorithms to reconstruct high-resolution images from faster-acquired low-resolution data, substituting computational processing for extended mechanical acquisition time.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The invention concerns a method for producing an isotropic 3D image, said method comprising steps of : a) acquisition of at least two anisotropic 3D images or anisotropic 2D multislice images of an object, b) determining an isotropic 3D image estimate from said two anisotropic 3D images, c) applying a denoising technique and a 3D super-resolution reconstruction algorithm, both combined in a constrained optimization problem which can be solved by an Augmented Lagrangian algorithm, by realizing following steps: d) defining 3D patches in the isotropic 3D image estimate, at iteration i=0, this estimate being obtained from step b), at iteration i+1 this estimate being obtained from step f), e) applying a denoising technique on said 3D patches in order to obtain denoised 3D patches, f) applying a geometrical 3D super-resolution reconstruction based on the denoised 3D patches used as a prior to produce the 3D isotropic image; the geometrical 3D super-resolution reconstruction including estimates of an Augmented Lagrangian algorithm, g) updating the Augmented Lagrangian estimates, h) iterating steps d-h until estimates meet a stopping criteria.