MRI Harmonization Using Flow-Based Model Invertibility
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing MRI harmonization methods require multiple datasets from different domains for training and struggle with generalizability to unseen domains, necessitating 'traveling subjects' or large datasets, and are challenging to apply in novel domains due to the domain gap in MRI data.
Innovation Solution
A Blind Harmonization method using a flow-based model trained solely on target domain data, which enables harmonization of source domain images into the target domain by iteratively optimizing the harmonized image to maintain anatomical structure and contrast, leveraging the invertibility of normalizing flows.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple datasets from different domains are used for training harmonization networks, then harmonization performance is improved, but data acquisition complexity and cost increase
Solution Approach 1:
The patent extracts only the target domain data from the training process, eliminating the need for source domain data. The harmonization network is trained exclusively on target domain data, and source domain images are harmonized by optimizing their likelihood in the target domain distribution, thereby removing the complexity of acquiring and managing multiple datasets
Solution Approach 2:
The patent creates a probabilistic model (flow model) that copies the distribution characteristics of the target domain. This model is trained on target domain data and then used to evaluate and optimize source domain images, replacing the need for actual target domain images from multiple sources during inference
2Measurement precision
If harmonization networks are trained on specific source and target domain pairs, then harmonization accuracy is improved, but generalizability to novel domains deteriorates
Solution Approach 1:
The patent creates a universal harmonization framework where a single flow model trained on target domain data can harmonize images from any source domain. The method is not limited to specific source-target pairs but can generalize to unseen domains, as it optimizes source images to match the target domain distribution without requiring source domain training data
Solution Approach 2:
The patent changes the fundamental parameter of training data composition, shifting from requiring both source and target domain data to using only target domain data. This parameter change enables the model to adapt to any source domain while maintaining harmonization accuracy for the target domain
3Quantity of substance
If traveling subjects undergo multiple MRI scans with different scanners, then both source and target domain images are obtained, but subject availability and study complexity increase
Solution Approach 1:
The patent extracts the essential requirement to only target domain data, eliminating the need for traveling subjects. Source domain images can be harmonized using the flow model trained on target domain data, removing the operational burden of coordinating multiple scans of the same subject across different scanners
Data Source
AI summary
An MR image processing method that repeats, by a computing device, a unit transformation method N number of times is provided. The unit transformation method in nth iteration comprises, generating a corrected image corrected from a prepared nth image using a predetermined reversible generative model, calculating a differential value of a distance between the corrected image and a source image, and generating a harmonized image by subtracting the differential value from the corrected image. The generating the corrected image comprises, generating a predetermined array by inputting the nth image into the reversible generative model in a forward direction of the reversible generative model, and generating the corrected image by inputting a scaled array in a reverse direction of the reversible generative model, the scaled array being obtained by scaling a value of each element of the generated array by a predetermined scaling factor (1−α) (0<α<1).