Comparison of Auto-Generated Contours on Virtual Monoenergetic Images across Multiple Energy Levels
Abstract
Purpose
To assess the effect of virtual monoenergetic image (VMI) energy variation (50-160 keV) on the accuracy of auto-generated contours using commercial autocontouring software.
Methods
An anthropomorphic thorax phantom (CIRS) was scanned using both conventional single-energy CT (SECT) at 120 kVp and dual-energy CT (DECT) on a SOMATOM go.Open Pro® scanner (Siemens Healthineers, Erlangen, Germany). DECT acquisitions were performed sequentially at 80 and 140 kVp and reconstructed into corresponding VMI series with 10 keV intervals ranging from 50 to 160 keV. Air calibration was performed on the same day prior to image acquisition. SECT and DECT scans were performed under identical acquisition conditions, with the phantom maintained in the same position for both scans. All images were reconstructed using a medium kernel (Q40), a 2 mm slice thickness, and sinogram-affirmed iterative reconstruction (SAFIRE) at strength level 3. Auto-contouring was performed using AutoContour® (Radformation, New York, NY, USA), generating contours for the left lung, right lung, heart, and spinal canal on all reconstructed VMI series. Manual contours for the same structures were delineated by a medical physicist for SECT series to serve as the reference standard (ground truth). Contour agreement was quantified using the Dice similarity coefficient (DSC).
Results
Across the evaluated energy range, DSC variation, expressed as ΔDSC (max–min), was organ dependent but generally demonstrated little sensitivity to energy variation (maximum ΔDSC < 4%). The ΔDSC for the left lung, right lung, heart, and spinal canal was approximately 2.4%, 3.9%, 1.9%, and 2.5%, respectively.
Conclusion
The limited ΔDSC observed across all evaluated organs indicates that auto-generated contours are minimally affected by VMI energy, supporting the use of a wide range of VMI energies without compromising contour accuracy.