Anatomy-Guided Synthesis of Novel CT Images at Full Hounsfield Range
Arjun Krishna, Ge Wang, Klaus D. Mueller · 2022
In this work, we present an approach that can synthesize novel CT images across the full Hounsfield range using a very small annotated dataset of around thirty patients and a large non-annotated dataset with high resolution medical images. Our method uses these two datasets in a sequence of steps involving texture learning via StyleGAN and semi-supervised learning via CycleGAN to generate a large annotated medical dataset suitable for use in deep learning algorithms for medical applications. Using an anatomy exploration interface we can then generate CT images with anatomies that were non-existent within either of the datasets, without compromising accuracy and quality. We show that our approach works for all Hounsfield windows with minimal depreciation in anatomical plausibility.