Conditional Synthetic Signal Generation for Microwave Head Imaging Using Diffusion Models

Abstract

Deep learning models have the potential to improve the accuracy and speed of medical microwave imaging. However, their performance often suffers due to a lack of high-quality data. Generative models, especially Denoising Diffusion Probabilistic Models (DDPM), solve this problem by creating realistic data for training and validation. These models have been used in various fields like text-to-image generation, time series generation, and EEG signal synthesis. However, they are not yet used in microwave head imaging for signal generation. Generating meaningful signals for stroke detection in microwave head imaging is challenging because the signals must show both the type and location of strokes. In this paper, DDPM is introduced for conditional signal generation in microwave head imaging. Also, different ways to embed the relevant conditions are explored. The generated signals are evaluated using quantitative metrics and the distorted Born iterative method to check their physical plausibility. Our results show that DDPM, with specially designed condition embeddings and noise schedulers, generates realistic signals, offering a new approach to train and validate deep learning models for microwave head imaging.