Medical Image Segmentation with Quality-Based Atlas Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical image segmentation methods, whether multi-atlas based or neural network based, struggle with accuracy and generalization to new datasets with slightly different properties, leading to potential failure-mode segmentations.
Innovation Solution
A method combining a segmentation function with multiple atlas images, using machine learning algorithms like neural networks or decision forests, and a voting scheme to enhance segmentation accuracy and robustness, allowing for accurate and robust segmentations across varied datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network based segmentation approach is used, then segmentation accuracy is improved, but generalization to new datasets with slightly different properties deteriorates
Solution Approach 1:
The patent segments the training process into two distinct phases: (1) training the segmentation function on a source dataset, and (2) selecting atlas images from that trained function's predictions on a target dataset. This separation allows the segmentation function to achieve high accuracy on training data while the atlas selection process handles adaptation to new datasets, resolving the contradiction between accuracy and generalization.
Solution Approach 2:
The patent introduces atlas images as an intermediary between the trained segmentation function and the target dataset. Instead of directly applying the segmentation function to new data, the system first generates atlas images from the trained function, then uses these atlases to guide segmentation on the target dataset. This intermediary layer enables better generalization while maintaining accuracy.
2Reliability
If multiple atlas images are used for segmentation, then robustness and generalization are improved, but computational complexity increases
Solution Approach 1:
The patent performs atlas image selection in advance, before actual segmentation is needed. By pre-selecting a limited set of atlas images from the target dataset based on the trained segmentation function's predictions, the system reduces the computational burden during actual segmentation operations. This preliminary action maintains robustness while reducing real-time complexity.
Solution Approach 2:
The patent applies different processing strategies to different parts of the image set: the trained segmentation function processes the source dataset to generate predictions, while a separate atlas selection mechanism processes the target dataset to identify suitable atlas images. This localized approach optimizes each stage for its specific purpose, balancing robustness and complexity.
Data Source
AI summary
Some embodiments are directed to a segmentation of medical images. For example, a medical image may be registering to multiple atlas images after which a segmentation function may be applied. Multiple segmentation may be fused into a final overall segmentation. The atlas images may be selected on the basis of high segmentation quality or low registration quality.