Cone-beam CT to CT harmonization by learning disentangled representations

Abstract

Chest computed tomography (CT) is a widely used imaging modality for evaluating thoracic pathology, with fan-beam CT (FBCT) and cone-beam CT (CBCT) being the primary types. While CBCT reduces radiation exposure, it often results in lower image quality, limited field of view, and increased artifacts, restricting its clinical applications. Given these challenges and the predominance of FBCT data and automated models designed for FBCT, developing high-quality CBCT-to-CT image synthesis is essential for improving CBCT image quality and expanding its applications. One promising approach is image harmonization, which mitigates domain shifts in medical images by translating from a source image to a target image acquisition setting while preserving the underlying anatomy. Recent methods, such as HACA3 for magnetic resonance imaging, have used disentangled representations of anatomy, contrast, and artifacts to respect anatomical differences between contrasts. To this end, this study aimed to perform image synthesis harmonization between CBCT and FBCT to capture CBCT anatomy while preserving FBCT quality and resolution. The harmonized images were evaluated through downstream tasks requiring various detail levels: lobes, nodules, and airways segmentation. Results showed that the model effectively reduced artifacts and smoothed the HU distribution, leading to improved segmentation performance, particularly in lobes and airways, although some detail was lacking in segmented airways. External evaluations suggested the model’s potential generalizability to other distributions. Overall, this harmonization approach enhances CBCT image quality, expanding its applicability across various imaging tasks.