Medical Image Segmentation With Dropout for Variable MRI Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for medical image segmentation fail when applied to data not well represented in the training process, particularly due to variability from different scanners and incomplete data sets, necessitating models that can function with variable inputs and scanner variations without retraining.
Innovation Solution
A deep convolutional neural network (DCNN) is trained using random dropout and standardized inputs to segment medical images, incorporating dropout layers and iterative weight updates to enhance robustness and adaptability across different imaging systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are trained with specific data combinations (e.g., MRI and PET scans), then segmentation accuracy is improved for that specific data type, but the model fails when applied to variable subsets of data (e.g., only one modality)
Solution Approach 1:
The model employs dynamic dropout layers that can adaptively deactivate different input channels based on availability. During training, dropout layers randomly deactivate portions of the input data to teach the model to function with incomplete information. During inference, the same dropout layers can be configured to handle variable combinations of available modalities, enabling the model to maintain segmentation accuracy across different data scenarios without requiring separate models for each configuration.
2Measurement precision
If models are trained with data from specific scanners, then segmentation performance is optimized for that scanner, but the model cannot reliably process data from other scanners
Solution Approach 1:
The model incorporates preliminary standardization processing that normalizes image intensities and characteristics before they enter the neural network. This pre-processing step transforms images from different scanners into a common representation space, allowing the model to learn scanner-independent features. The standardization is applied during training to create a robust model that generalizes across scanners, eliminating the need for scanner-specific training for each new imaging system.
3Measurement precision
If a new model is trained for each scanner type, then optimal segmentation performance is achieved for that scanner, but the complexity and time required for deployment increases
Solution Approach 1:
The model is designed as a universal segmentation system that can process images from multiple scanner types using a single trained model. The standardization preprocessing layer and adaptive dropout mechanisms enable the same model architecture to handle variable input formats from different scanners. This eliminates the need to train and deploy separate models for each scanner type, significantly reducing deployment complexity while maintaining high segmentation performance across all supported scanners.
4Stability of the object's composition
If image standardization methods are applied to all input images, then consistency across scanners is improved, but the computational processing time increases
Solution Approach 1:
The standardization processing is performed as a preliminary step before image input to the neural network, transforming images into a standardized representation space that ensures consistency across scanners. By completing this computationally intensive standardization beforehand, the subsequent segmentation inference can proceed more efficiently with consistent input characteristics, balancing the trade-off between image consistency and processing time.
Data Source
AI summary
Systems and methods are described for segmenting medical images, such as magnetic resonance images, using a deep learning model that has been trained using random dropped inputs, standardized inputs, or both. Medical images can be segmented based on anatomy, physiology, pathology, other properties or characteristics represented in the medical images, or combinations thereof. As one example, multi-contrast magnetic resonance images are input to the trained deep learning model in order to generate multiple segmented medical images, each representing a different segmentation class.