P. Pathak1, Z. A. Siddiqui2, C. Claunch2, C. Stewart2, B. Sun2, S. Sharma2, D. A. Hamstra2, P. M. Jhaveri2, and A. S. Mohamed2; 1Department of Radiation Oncology, Baylor College of Medicine, Houston, TX, 2Department of Radiation Oncology, Dan L. Duncan Comprehensive Cancer Center, Baylor College of Medicine, Houston, TX
Purpose/Objective(s):Enhanced cone beam computed tomography (eCBCT) is a novel imaging modality offering improved image quality and monitoring of tumor and normal tissue changes during radiation therapy (RT), holding the potential to revolutionize adaptive RT approaches. A critical step toward application in these domains is the assessment of eCBCT performance in accurately segmenting normal tissues and tumor volumes. Consequently, our study is designed to assess eCBCTs efficacy in precise segmentation of normal tissues and tumor volumes by comparing observer agreement and accuracy against traditional CT simulations (CTsims) in a head and neck cancer (HNC) patient cohort. Materials/
Methods: We acquired paired CTsims and first fraction eCBCT images for five patients undergoing definitive RT for HNC. Regions of interest (ROIs) for primary tumors and organs at risk (OARs) were manually segmented on both modalities by a primary observer, without additional imaging information, to prevent bias. Seventeen ROIs were delineated per image. Four independent observers (two experts and two residents) also segmented eCBCTs. Paired CTsim/eCBCT were co-registered for each patient and ROIs were propagated to the CTsim for comparison. Agreement of eCBCT segmentation volumes to CTsim volumes were calculated using volume overlap and surface distance metrics. Mann-Whitney U test was used to compare the performance between different ROI types and the Intraclass Correlation Coefficient (ICC) was used to calculate the inter-observer variability. Results: Across all patients, 425 ROIs were segmented. Dice similarity coefficients (DSC) had a median and interquartile range of 0.8 (0.7-0.83), mean surface distance (MSD) was 1.5 mm (1.1-2.3), and the maximum Hausdorff distance (Max HD) was 9.3 mm (6.4-13.7). Mandibular bone ROIs achieved the best volume overlap performance (average DSC 0.88) and minimal MSD (average 0.89 mm), with muscles of mastication also performing well (average DSC 0.82 and MSD 1.18 mm). Salivary glands displayed commendable results (average DSC 0.77 and MSD 2 mm). However, GTV ROIs had the worst performance with an average DSC of 0.5 and MSD of 4 mm. GTV ROIs had statistically significant lower DSC, and higher MSC and Max HD compared to all other OARs (P <0.001). The ICC showed excellent inter-observer agreement for both DSC and Max HD metrics of 0.93 (95%CI 0.9-0.95) and 0.97 (95% CI 0.96-0.98), respectively, for normal tissue ROIs as well as a good agreement of 0.87 (95% CI 0.74-0.94) for MSD. On the contrary, ICC showed poor inter-observer agreement for GTV ROIs using DSC 0.5 (95% CI -0.8-0.9) and moderate agreement 0f 0.71 (95% CI -0.07-0.97) using MSD. Conclusion: This pilot study reveals that manual segmentation of eCBCT in HNC patients undergoing RT aligns closely with CTsim for most OARs, maintaining high observer agreement. However, GTV segmentation using eCBCT alone, without contrast or diagnostic imaging integration, is unreliable.
