Dielectric Breast Phantoms by Generative Adversarial Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Machine learning (ML)-based microwave breast imaging (MBI) requires extensive digital dielectric breast phantoms for training.
• Existing phantoms are limited in number and diversity, hindering robust ML algorithm development for MBI.

Purpose of the Study:

• To present a novel neural network method for generating diverse 2D virtual breast phantoms.
• To provide a solution for the scarcity of training data in ML-based MBI research.

Main Methods:

• A neural network was employed to generate 2D virtual breast phantoms.
• Each phantom comprises multiple images representing dielectric parameter distributions.
• Statistical analysis was conducted on 10,000 generated phantoms to evaluate the generative network's performance.

Main Results:

• The neural network successfully generated virtual breast phantoms similar to real ones but with crucial variations.
• The generated phantoms differ from the training data, enhancing model generalizability.
• The generative network demonstrated the capability to produce a virtually unlimited supply of varied breast images.

Conclusions:

• The developed generative network can produce unlimited, diverse breast images for ML-based MBI.
• This method addresses the critical need for large, varied datasets in developing robust ML algorithms for MBI.
• The generated phantoms will accelerate the deployment readiness of ML-based MBI systems.