Medical Imaging System Using Hierarchical Autoencoders for Disease Prognosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical imaging systems face challenges in processing high-dimensional data from tools like MRI and PET due to a mismatch between the number of data dimensions and the limited number of samples, leading to overfitting and ineffective disease prognosis.
Innovation Solution
A medical imaging system that divides high-dimensional data into smaller subsets for independent training of multiple machine learning modules, using denoising autoencoders with varying parameters to reduce overfitting and improve generalizability, allowing for effective disease prognosis with limited clinical data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning systems are trained using high-dimensional clinical data from medical imaging, then the ability to measure and predict disease progression is improved, but the system suffers from overfitting due to the mismatch between the number of data dimensions and the limited number of samples
Solution Approach 1:
The patent applies segmentation by dividing the high-dimensional data processing into multiple stages: first dividing data into super-voxels, then into super-patches, and finally processing through multiple levels of autoencoders. This hierarchical segmentation reduces the dimensionality at each stage, allowing the model to learn from limited samples without overfitting while maintaining the ability to capture complex disease progression patterns.
Solution Approach 2:
The patent transforms the data representation by converting 3D volumetric data into 2D super-patches through projection, and then further transforms these into compressed representations through autoencoders. This dimensional transformation reduces the data complexity from millions of voxels to manageable super-patches, enabling effective training with limited clinical samples while preserving disease-relevant information.
2Loss of information
If the number of data dimensions from imaging equipment is increased to provide more detailed measurements, then the information available for diagnosis is improved, but the complexity of the data processing system increases beyond the analysis abilities of individual diagnosticians
Solution Approach 1:
The patent merges adjacent voxels into super-voxels, and then merges super-voxels into super-patches, effectively combining redundant fine-grained information into coarser representations. This merging process reduces data complexity while preserving the essential clinical information needed for diagnosis, making the data manageable for analysis without losing critical disease indicators.
Solution Approach 2:
The patent creates compressed representations (copies) of the original high-dimensional data through autoencoders. These compressed copies retain the essential disease-related information in a lower-dimensional format that is easier to process and analyze, while the original detailed data remains available if needed for verification or further analysis.
Data Source
AI summary
A medical imaging system that processes input data (imaging and/or non imaging) having high dimensionality and few samples to learn from, by using multiple ranks of machine learning modules each dealing with a separate portion of the clinical data. The outputs of the individual machine learning modules are the combined to provide a result reflective of the complete image data set.