3D Myocardium Segmentation Using Vision Transformer and Mesh Deformer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to create consistent 3D models of myocardium that are accurate for both cardiac segmentation and CFD simulations, often leading to misclassification and artifacts due to similarities between ischemic myocardium and surrounding tissue, and shape changes during the cardiac cycle.
Innovation Solution
A computer-implemented method using ensembled models, combining vision transformers for precise segmentation and mesh deformers to correct local errors, ensuring accurate and reliable 3D myocardium models suitable for both cardiac scale visualization and CFD simulations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single segmentation model is used for both cardiac segmentation and CFD simulations, then the process is simplified, but the accuracy and reliability deteriorate due to conflicting requirements of the two applications
Solution Approach 1:
The patent divides the segmentation process into two distinct models: a first segmentation model optimized for cardiac segmentation accuracy and a second segmentation model optimized for CFD simulation requirements. This segmentation of the segmentation process resolves the contradiction by allowing each model to be specialized for its specific application rather than forcing a single general-purpose model to satisfy both conflicting requirements.
Solution Approach 2:
The patent creates a unified segmentation system that serves multiple functions through the combination of two specialized models. The first model handles cardiac segmentation while the second handles CFD preparation, and together they provide a comprehensive solution that addresses both applications' needs, achieving multi-functionality without compromising the specialized requirements of either application.
2Measurement precision
If vision transformers are used for semantic segmentation, then segmentation precision is improved, but local errors and misclassification artifacts increase due to similarities between ischemic myocardium and surrounding tissue
Solution Approach 1:
The patent introduces a second segmentation model as an intermediary that processes the output of the first vision transformer model. This intermediary model specifically targets and corrects the local errors and misclassification artifacts produced by the first model, particularly in challenging regions where ischemic myocardium resembles surrounding tissue. The second model acts as a mediator that refines the initial segmentation without requiring complete redesign of the primary segmentation approach.
3Reliability
If mesh deformers are applied to correct local errors, then model reliability for CFD simulations is improved, but the overall segmentation accuracy may deteriorate due to over-correction of valid features
Solution Approach 1:
The patent applies local quality by using the second segmentation model specifically targeted at correcting local errors and artifacts in the first model's output, rather than re-processing the entire segmentation. This localized correction approach ensures that valid features captured by the first model are preserved while only the problematic regions are refined, thus maintaining overall segmentation accuracy while improving CFD simulation reliability.
Data Source
Figure 1~2a
Figure 2b~2d
Figure 3
AI summary
Computer implemented method for enhancing diagnosis by processing medical images, comprising step (130) of semantic segmentation of the medical image to obtain a model of organ, step (140) of orientation of the model, step (150) of medical segmentation of the model of an organ and step (170) of preparing visualization of the model of the organ and/or the medical image together with the result of medical segmentation, according to the invention comprises using the model of organu ensembled from a first and a second model. The first model is obtained with vision transformer, and the second with mesh deformer. The invention further concerns computer program product and computing system.