GAN Feature Visualization for FMRI Tiny Texture Preservation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning visualization methods for FMRI images often overlook tiny texture features of the active cerebral cortex, affecting the quality of visualization results.
Innovation Solution
A Generative Adversarial Network (GAN) is trained using real and fictitious training data with positive and negative samples to learn a key feature map, which is then used to improve feature extraction and visualization of medical images by distinguishing between real and fictitious data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning visualization methods are used to visualize FMRI images, then the visualization can be achieved, but tiny texture features of the active cerebral cortex are easily ignored, affecting the quality of the visualization result
Solution Approach 1:
The patent segments the feature extraction process into two distinct components: a generator network that creates synthetic training data and a discriminator network that performs the actual feature extraction and visualization. This segmentation allows the discriminator to be specifically optimized for preserving tiny texture features while the generator focuses on creating realistic synthetic data, thereby resolving the contradiction between comprehensive feature extraction and preservation of fine details
Solution Approach 2:
The patent applies preliminary action by using the generator network to pre-process and augment the training data with synthetic FMRI images before the discriminator performs feature extraction. This preliminary generation of diverse training samples enables the discriminator to learn more robust feature representations that capture tiny texture features, improving measurement precision without losing information
Data Source
AI summary
An image feature visualization method and apparatus, and an electronic device during model training, inputs the real training data with positive samples into a mapping generator to obtain fictitious training data with negative samples. The mapping generator includes a mapping module configured to learn a key feature map that distinguishes the real training data with positive samples/negative samples, and the fictitious training data with negative samples is generated based on the real training data with positive samples and the key feature map. The training data with negative samples is input into a discriminator to obtain a discrimination result. An optimizer optimizes the mapping generator and the discriminator until training is completed. During model application, a target image that is to be processed is input into the mapping generator, and the mapper in the mapping generator extracts features of the target image.