Synthetic CT Generation for Cardiothoracic Injuries with Intracardiac Foreign Bodies

Abstract

This paper proposes novel methods for the generation of synthetic CT images of cardiothoracic injuries with foreign bodies (FB) localized in the cardiac region. These methods are intended to improve the FB segmentation performance by machine learning models through the augmentation of training datasets with synthetic images. A small real-world dataset was collected, consisting of 8 CT scans of combat cardiothoracic injuries with foreign bodies (FB amount ranging from 1 to 3) localized in the heart area, and an additional 4 "clean" CT scans of blast-induced cardiothoracic injuries where foreign bodies are absent in the cardiac region. Through the analysis of the collected dataset and findings from similar studies, the morphology of foreign bodies and the statistics of their localization in the heart during combat cardiothoracic trauma were examined. Two methods for synthesizing artificial CT images of cardiothoracic injuries with heart-localized FBs based on clean CT scans were developed: RealInsFB-CT, based on the insertion of regions of interest extracted from real injury CT scans-containing FB, artifacts, and surrounding damaged tissues – into clean CT scans and MorphGenFB-CT, based on the morphologically and statistically grounded generation of a 3D FB model, filling the model with Hounsfield Unit values corresponding to the selected FB material, inserting the model into one of the heart chambers on a clean CT scan, and the artificial generation of artifacts. Using both methods, CT images were synthesized and compiled into training datasets (19 CT scans + 19 FB masks each). The visual, structural, and statistical similarity between real and synthesized CT scans was proved.