MRI Joint Segmentation via Two-Stage ML Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for quantifying joint inflammation in MRI images rely on manual visual inspection, which is subjective and unreliable, particularly in distinguishing bone regions from soft tissues, and automated classifiers have not been effective due to the conflicted voting problem in distinguishing inflammatory bone regions from soft tissues in T2-weighted images.
Innovation Solution
A two-stage automated segmentation process using machine-learning algorithms, where image slices are categorized based on the presence of bony structures and tendons, and then segmented into bone and non-bone regions using separate classifiers, such as the Inception v3 and U-Net architectures, to generate final segmented image data for quantifying bone marrow edema.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated classifiers are used to segment bone regions from soft tissues in MRI images, then productivity is improved, but measurement precision deteriorates due to the conflicted voting problem in distinguishing inflammatory bone regions from soft tissues
Solution Approach 1:
The patent applies segmentation by dividing the image processing task into two distinct stages: (1) image-slice categorization to identify slices containing bony structures or tendons, and (2) bone segmentation to distinguish bone from non-bone regions. This multi-stage segmentation approach resolves the conflicted voting problem by separating different classification functions, thereby maintaining both automation and measurement precision.
2Measurement precision
If manual visual inspection is used to assess inflammation, then measurement precision is maintained, but productivity deteriorates due to subjectivity and time consumption
Solution Approach 1:
The patent implements self-service by developing an automated system that performs inflammation quantification without requiring manual visual inspection. The machine learning models automatically segment bone regions, quantify bone marrow edema, and generate assessment results, eliminating the need for practitioner time and subjective judgment while maintaining objective measurement precision.
3Device complexity
If a single automated classifier is used for segmentation, then device complexity is reduced, but reliability deteriorates due to inability to handle different image slice categories
Solution Approach 1:
The patent applies segmentation by dividing the image processing task into two distinct stages: (1) image-slice categorization to identify slices containing bony structures or tendons, and (2) bone segmentation to distinguish bone from non-bone regions. This multi-stage segmentation approach resolves the conflicted voting problem by separating different classification functions, thereby maintaining both automation and measurement precision.
Solution Approach 2:
The patent applies local quality by training different segmentation classifiers on different types of image slices based on their specific characteristics. Slices containing both bone and tendon are processed by one classifier, while slices containing only bone are processed by another classifier. This localized approach ensures each classifier is optimized for its specific input type, improving reliability without excessive complexity.
Data Source
AI summary
Segmentation of bony regions in MRI images of joints is automated using a two-stage process. In a first stage, a machine-learning image-slice categorizer is used to categorize image slices of the MRI image data as belonging to one of a set of image-slice categories, depending on presence or absence of bone and/or tendon in the image slice. In a second stage, a first instance of a machine-learning segmentation classifier is used to segment image slices that contain both bone and tendon into bone and non-bone regions, and a second instance of a machine-learning segmentation classifier is used to segment image slices that contain bone but not tendon into bone regions and non-bone regions. Results from the two segmentation classifiers can be combined across image slices to provide a final segmentation of the bony structures, including inflammatory regions, in the image data.