Medical Image Learning Device Using Uncertainty-Based Data Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for generating learning models for disease detection from medical images require high annotation costs and do not effectively utilize uncertainty values from different networks, leading to inefficiencies in data preparation and model reliability.
Innovation Solution
A learning device and method that uses a first processor to extract organ regions and estimate uncertainty from medical images, normalizing input data for a second learning model, which is trained using the extracted information and uncertainty values, reducing the need for manual annotation and improving detection reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If organ region cropping is performed to improve detection accuracy, then detection accuracy is improved, but annotation cost is increased
Solution Approach 1:
The system uses the deep learning network itself to automatically generate organ region extraction results, which are then used as annotations for training. This self-generated annotation approach eliminates the need for manual expert annotation while maintaining high detection accuracy, as the network learns to focus on relevant organ regions through self-supervised learning.
Solution Approach 2:
The system performs preliminary organ region extraction using the deep learning network before the actual disease detection training. This preliminary extraction of organ regions creates pre-processed training data that guides the network to focus on clinically relevant areas, improving detection accuracy without requiring manual annotation of each organ region.
2Reliability
If uncertainty values from multiple networks are used for classification, then classification reliability is improved, but device complexity is increased
Solution Approach 1:
The system merges the uncertainty values from multiple deep learning networks with different architectures into a unified classification framework. By combining the uncertainty assessments from diverse networks, the system achieves more robust and reliable classification, as the multiple perspectives reduce the impact of individual network biases or errors.
Solution Approach 2:
The system creates a universal classification framework that can accommodate uncertainty values from various types of deep learning networks. This multi-functional approach allows different network architectures to contribute their uncertainty assessments to a common decision-making process, improving reliability without requiring network-specific customization.
Data Source
AI summary
A first processor of a learning device reads out a first medical use image having a disease label from a data set stored in a memory and inputs the read out first medical use image to a first learning model. The first medical use image is normalized based on a lung field region extracted by the first learning model, and a second learning model that has not been trained and detects a disease is trained by using the normalized first medical use image and the disease label. In a case in which the second learning model is trained, a value of the uncertainty of the first medical use image is calculated based on the uncertainty simultaneously estimated by the first learning model, and the first medical use image having a large value of uncertainty is excluded from learning data.