Medical Image Reliability via Neural Network Variance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In medical image diagnoses, there is a lack of established methods to measure the reliability of images reconstructed using deep neural networks (DNNs), making it unclear how reliable these images are.
Innovation Solution
A medical data processing apparatus that generates multiple medical images by applying different machine learning models or varying parameters to raw data, and outputs a reliability measure based on these images, allowing for the evaluation of image reliability through display and processing of DNN reconstruction images and reliability maps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If DNN reconstruction is used to generate medical images from raw data, then image quality and diagnostic capability are improved, but reliability measurement capability deteriorates (no established method to measure reliability)
Solution Approach 1:
The patent segments the reliability measurement process into distinct components: generating multiple provisional reconstruction images through different DNN models or parameter variations, calculating variance metrics for each pixel position, and synthesizing these into a comprehensive reliability map. This segmentation enables systematic reliability assessment that was previously unavailable.
Solution Approach 2:
The patent creates multiple copies of the reconstruction process by generating several provisional reconstruction images using different DNN models or varying parameters. These copied reconstruction results are then compared to determine reliability, allowing the system to assess confidence without requiring a single ground-truth reference.
2Reliability
If multiple machine learning models are applied to generate provisional reconstruction images, then reliability evaluation capability is improved, but device complexity increases
Solution Approach 1:
The patent implements a universal reliability assessment framework that can work with multiple DNN models or parameter variations without requiring model-specific customization. The same variance calculation and reliability map generation process applies regardless of how many provisional images are generated or what specific models are used, providing multi-functional reliability evaluation.
Solution Approach 2:
The patent enables reliability assessment by changing parameters in the reconstruction process - either by varying model parameters across multiple models or by adjusting parameters within a single model across multiple runs. This parameter variation approach provides the necessary diversity for reliability measurement without fundamentally changing the core reconstruction architecture.
3Measurement precision
If reliability is defined based on variations among provisional reconstruction images, then measurement precision of reliability is improved, but loss of time increases due to multiple reconstructions
Solution Approach 1:
The patent calculates reliability metrics at partial levels - computing variance for individual pixel positions or regions of interest rather than requiring complete processing of all image data. This partial action approach provides sufficient reliability measurement precision for clinical decision-making while reducing the computational burden and time required compared to exhaustive analysis.
Data Source
AI summary
A medical image processing apparatus includes processing circuitry. The processing circuitry generates a plurality of first medical images by applying a plurality of first machine learning models having different elements to a set of raw data, or applying a first machine learning model to a set of raw data a plurality of times while changing elements. The processing circuitry outputs a second medical image and a first reliability relating to the second medical image based on the first medical images.