Author: Cem Altunbas, Farhang Bayat, Roy Bliley, Rupesh Dotel, Brian Kavanagh, Uttam Pyakurel, Tyler Robin, Ryan Sabounchi 👨🔬
Affiliation: Department of Radiation Oncology, University of Colorado School of Medicine, Taussig Cancer Center, Cleveland Clinic, University of Colorado Anschutz Medical Campus 🌍
Purpose: Automatic segmentation of anatomical structures in CBCT images is key to enabling dose delivery monitoring and online plan modifications in radiotherapy. However, poor image quality can degrade segmentation performance. In this work, a quantitative CBCT (qCBCT) method with high image fidelity was developed, and its impact on Deep Learning (DL)-based image segmentation was evaluated in a pilot clinical trial.
Methods: Seventeen patients with cancers in the pelvis and abdomen were scanned with qCBCT and standard-of-care (SOC) TrueBeam CBCT in an IRB-approved study. Each SOC scan was reconstructed using standard (FDK) and advanced (iCBCT) reconstruction options. The qCBCT scans, employing a novel antiscatter grid and data processing algorithms, were reconstructed using the FDK method. Five anatomical structures were segmented in each CBCT and planning CT (planCT) by a DL model with Transformer architecture and compared to reference segmentations by a human observer. Segmentation accuracy was quantified using the Dice similarity coefficient (DSC) and Hausdorff distance (HD). Statistical significance was tested using the Wilcoxon signed-rank test.
Results: PlanCT yielded the highest DL segmentation accuracy, with an average DSC of 0.83 ± 0.14 and HD of 2.81 ± 0.85 mm. DL segmentations in standard CBCT, advanced CBCT, and qCBCT had average DSC values of 0.73 ± 0.18, 0.74 ± 0.17, and 0.81 ± 0.13, respectively (p < 0.001). Similarly, average HD values for DL segmentations in standard, advanced, and qCBCT were 3.26 ± 0.75 mm, 3.28 ± 0.87 mm, and 2.96 ± 0.85 mm, respectively (p < 0.001). Higher DSC and lower HD values indicate that DL segmentations in qCBCT are more similar to human-generated segmentations.
Conclusion: The qCBCT approach can potentially increase DL-based segmentation accuracy in the pelvis and abdomen compared to SOC CBCT. These findings indicate that robust scatter suppression with qCBCT can improve DL-based segmentation performance for radiotherapy.