ROI Extraction Model for Multi-Format Medical Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for extracting regions of interest from medical images with different representation formats require significant cost and effort due to the need for ground-truth masks for each format, and existing generative adversarial networks are limited to one-to-one domain conversion or require domain labels.
Innovation Solution
A learning apparatus and method that utilizes an encoder, multiple decoders, and discriminators to convert and extract regions of interest across multiple representation formats without requiring ground-truth masks for each format, using generative adversarial networks to adapt feature spaces and perform domain conversion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ground-truth masks are prepared for each representation format (CT, MRI T1-weighted, MRI T2-weighted), then region extraction accuracy for each format is improved, but the cost and complexity of model training and data preparation increases significantly
Solution Approach 1:
The patent applies a single unified extraction model that can process multiple representation formats (CT, MRI T1-weighted, MRI T2-weighted) simultaneously, eliminating the need for separate models for each format. This multi-functional approach maintains high extraction accuracy across all formats while significantly reducing training complexity and data preparation costs.
Solution Approach 2:
The patent introduces a domain adaptation mechanism that acts as an intermediary between different image formats and the extraction model. This intermediary layer enables the model to adapt to various representation formats without requiring format-specific ground-truth masks, thereby maintaining accuracy while simplifying the training process.
2Measurement precision
If separate extraction models are created for each representation format, then extraction accuracy for each format is optimized, but the time and resources required for model development increases
Solution Approach 1:
By developing a single universal extraction model that handles multiple representation formats through domain adaptation, the patent eliminates the time-consuming process of creating and training separate models for each format. This unified model achieves optimized extraction accuracy across all formats while significantly reducing development time and resources.
3Measurement precision
If domain adaptation is performed between source and target domains, then learning accuracy for target domain with few or no labels is improved, but the complexity of feature space alignment increases
Solution Approach 1:
The patent employs a domain adaptation mechanism that serves as an intermediary between source and target domains. This intermediary layer aligns feature spaces while maintaining the ability to extract regions accurately in the target domain even with few or no labels, achieving high learning accuracy without excessive complexity.
Solution Approach 2:
The patent incorporates feedback mechanisms in the domain adaptation process, where the model continuously refines its feature space alignment based on available data. This feedback loop enables accurate learning in the target domain while managing the complexity of feature space alignment through iterative optimization.
Data Source
AI summary
A learning apparatus of an extraction model that extracts a region of interest from images having a plurality of different representation formats is provided. The learning apparatus comprising one processor is configured to perform learning of an encoder, a first decoder, a first discriminator, a second decoder, and a second discriminator which constitute the extraction model. The encoder extracts a feature amount of a first image of a first representation format. The first decoder derives a second virtual image of a second representation format. The first discriminator outputs a first discrimination result on a representation format of an input image and whether the input image is a real image or a virtual image. The second discriminator outputs a second discrimination result on whether an extraction result of a region of interest by the second decoder is from a first image with or without a ground-truth mask.