Synthetic Microwave 3D Breast Models: A Step Forward with Denoising Diffusion Models

With the increased interest in utilizing machine learning techniques for microwave breast imaging there is an associated increase in the demand for anthropomorphically accurate numerical breast phantoms. Here we investigate the use of generative deep-learning techniques to create 3D complex-valued permittivity numerical breast phantoms. A Denoising Diffusion Probabilistic Model is trained using a publicly available MRI breast dataset that has been converted to three relevant 3D permittivity images: real and imaginary parts of the complex-valued permittivity, as well as a tissue segmented image. Different diffusion models are investigated with the ultimate goal of generating new 3D breast phantoms. The value of having a large database of breast phantoms is that microwave data can then be obtained using traditional electromagnetic forward solvers that model the imaging system being used. This creates labelled training data consisting of microwave data corresponding to specific permittivity breast phantoms.