Automated patient-specific and organ-based image quality metrics on dual-energy CT datasets for large scale studies

The purpose of this study was to develop an automated patient-specific and organ-based image quality (IQ) assessment tool for dual energy (DE) computed tomography (CT) images for large scale clinical analysis. To demonstrate its utility, this tool was used to compare the image quality of virtual monoenergetic images (VMI) with mixed images. The tool combines an automated organ segmentation model developed to segment key organs of interest and a patient-based IQ assessment model. The organ segmentation model was reported in our previous study and used to segment liver in this study; specifically, the model used 3D Unet architecture, developed by training on 200 manually labeled CT cases. We used task-based image quality assessment to define a spectral detectability index (ds'), which enables the task definition to be lesion with specific contrast properties depending on DE reconstruction chosen. For actual testing of the tool, this study included 322 abdominopelvic DECT examinations acquired with dual-source CT. Within regions of segmented organ volumes, the IQ assessment tool automatically measures noise and calculates the spectral dependent detectability index (ds') for a detection task (i.e., liver lesion). This organ-based IQ tool was used to compare the image quality of DE images including VMIs at 50 keV, 70 keV and mixed images. Compared to mixed images, the results showed that VMI at 70 keV had better or equivalent spectral detectability index (difference 12.62±2.95%), while 50 keV images showed improved detectability index (61.62±10.23%). The ability to automatically assess image quality on a patient-specific and organ-based level may facilitate large scale clinical analysis, standardization, and optimization.