Virtual Microwave Phantom Generation for AI Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MW breast imaging techniques suffer from low spatial resolution, limiting their effectiveness in breast cancer detection compared to modalities like PET and MRI, and existing AI medical imaging techniques often rely on conventional image reconstruction methods due to the lack of accurate phantom data.
Innovation Solution
The use of a generative adversarial network (GAN) to generate artificial breast phantoms that mimic real human breast tissue, which can be used as high-quality image labels for training AI systems in MW breast imaging, thereby improving image resolution and reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional image reconstruction methods are used in MW breast imaging, then the imaging process is simple and fast, but the spatial resolution remains at centimeter level which is insufficient for accurate cancer detection
Solution Approach 1:
The patent creates virtual copies of real breast tissue through GAN-generated phantoms. These synthetic phantoms replicate the complex dielectric properties and anatomical structures of real breast tissue, enabling AI training without requiring extensive real patient data. The generator network learns to copy realistic tissue characteristics from training images, producing high-resolution phantom data that preserves the statistical and physical properties of actual breast tissue.
Solution Approach 2:
The patent performs preliminary action by generating high-quality phantom data before actual AI reconstruction training. The GAN is trained beforehand to create realistic breast tissue phantoms with accurate dielectric properties, which then serve as pre-prepared training datasets for subsequent AI reconstruction algorithms. This preliminary data preparation enables the main AI system to achieve high resolution without complex hardware modifications.
2Manufacturing precision
If AI technologies are used to improve spatial resolution, then image quality can be significantly enhanced, but accurate training data (ground truth) is unavailable in practice
Solution Approach 1:
The GAN creates synthetic copies of ground truth data by generating realistic phantom images that mimic actual breast tissue. The generator produces high-resolution phantom images with known ground truth dielectric properties, while the discriminator ensures these synthetic images indistinguishably replicate real tissue characteristics. This copying approach solves the ground truth availability problem by creating artificial but realistic training data.
Solution Approach 2:
The patent transforms the availability problem by changing the parameter of data origin from real patient measurements to synthetic generation. The GAN modifies the data generation process to produce phantom images with controllable and known ground truth parameters, enabling supervised AI training without relying on unavailable real-world ground truth labels.
3Reliability
If a small set of training data is used to train AI systems, then the training process is fast and resource-efficient, but the systems become overfitted and lose reliability
Solution Approach 1:
The patent implements dynamics by creating a dynamic and diverse training dataset through the GAN. Instead of using a static small dataset, the generator continuously produces varied phantom images with different anatomical structures, tissue compositions, and dielectric properties. This dynamic data generation approach provides the AI system with extensive diverse training examples, preventing overfitting while maintaining training efficiency through on-demand generation.
Solution Approach 2:
The GAN-based phantom generation system serves multiple functions: it generates diverse training data for AI training, provides ground truth labels for supervision, creates realistic test cases for validation, and enables various imaging scenarios without requiring additional real patient data. This multi-functional approach enhances reliability through comprehensive training while maintaining productivity through a single versatile data generation system.
Data Source
AI summary
A method for generating virtual microwave (MW) phantoms at a specific MW frequency or multiple MW frequencies, comprising generating virtual MW phantoms using a generative neural network (generator) and identifying the authenticity of the virtual MW phantoms using a discriminative neural network (discriminator).