WMH Segmentation Using CNN Ensembles for MRI Differentiation Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for automatic segmentation of white matter hyperintensities in magnetic resonance images lack the precision needed for accurate differentiation between pathological and physiological hyperintensities, posing a challenge for radiologists in diagnosing brain injuries.
Innovation Solution
A method and system using a combination of convolutional neural networks, including standard, residual, dilated, and dilated-residual architectures, to process magnetic resonance brain images, identify voxels, calculate probabilities, and generate image masks by averaging probabilities and applying a threshold, with pre-processing techniques to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If current automatic segmentation systems are used, then productivity is improved by reducing manual work, but measurement precision deteriorates due to inability to accurately differentiate pathological from physiological hyperintensities
Solution Approach 1:
The patent combines five different convolutional neural network architectures (standard CNN, residual CNN, dilated CNN, and their combinations) into an ensemble system. Each network processes the input image independently, and their probability outputs are averaged to produce the final segmentation mask. This merging of multiple segmentation approaches resolves the contradiction by achieving both automation and high precision through collective decision-making of diverse models.
Solution Approach 2:
The patent applies multiple parameter transformation techniques including intensity normalization (scaling to 0-1 range), standardization (z-score normalization using mean and standard deviation from training data), and probability thresholding (0.5 cutoff). These parameter changes enable the automated system to achieve measurement precision comparable to expert radiologists by optimizing the numerical characteristics of input data and decision boundaries.
2Measurement precision
If multiple convolutional neural networks are combined, then measurement precision is improved through ensemble averaging, but device complexity increases due to multiple network architectures
Solution Approach 1:
The patent segments the overall segmentation task into five independent computational paths, each handled by a different convolutional neural network architecture. Each network is trained independently on the same dataset and processes images through its own unique architectural characteristics. This segmentation of the computational task allows the system to leverage diverse strengths of different architectures while maintaining modular independence, resolving the contradiction between precision and complexity.
Solution Approach 2:
The patent creates five copies of the segmentation system using different network architectures (standard CNN, residual CNN, dilated CNN, and their combinations). Each copy is a complete, independently trained model that produces a probability map. The final segmentation is obtained by averaging these five copies' outputs. This copying approach enables precision improvement through ensemble methods while keeping each individual network component relatively simple and manageable.
3Measurement precision
If pre-processing techniques are applied, then measurement precision is improved through better feature normalization, but loss of time occurs during the additional processing steps
Solution Approach 1:
The patent performs pre-processing operations (intensity scaling to 0-1 range and z-score standardization using mean and standard deviation) before feeding images to the neural networks. These preliminary actions normalize the input data distribution, which significantly improves the convergence and accuracy of the training process. By performing these transformations in advance during the training phase, the system achieves higher measurement precision without requiring complex real-time processing during inference.
Data Source
AI summary
The present invention relates to a method and a system for the segmentation of white matter hyperintensities (WMHs) present in magnetic resonance brain images, comprising: providing an array of trained convolutional neural networks (CNNs) with a magnetic resonance brain image; determining, for each of the CNNs and for each voxel, the probability that the given voxel corresponds to a pathological hyperintensity; calculating the average of all the probabilities determined for each voxel; comparing the averaged probabilities for each voxel with a threshold; generating an image mask with the voxels that exceed the threshold.