1152 - The Effect of a Novel Quantitative CBCT Imaging Method on Deep Learning Based Image Segmentation Performance

Purpose/Objective(s): Deep Learning (DL) methods play an essential role in automatic segmentation of planning CT images, improving the efficacy of treatment planning processes. However, adaptation of DL methods to CBCT image segmentation for dose delivery monitoring is challenging due to poor CBCT image quality. To address this challenge, a novel quantitative CBCT (qCBCT) method has been developed and its effect on DL-based segmentation performance was evaluated in a prospective trial.

Materials/Methods: Twenty-six patients with head & neck, pelvic, and abdomen cancers were enrolled into an IRB-approved study. Each participant was scanned with qCBCT and standard-of-care CBCT using a Varian medical linear accelerator and using identical acquisition protocols. Standard-of-care scans were reconstructed using standard (FDK) and advanced (iCBCT) methods. qCBCT pipeline employs a 2D antiscatter grid prototype and data processing algorithms, scans were reconstructed using the FDK method. A DL segmentation model with Transformer architecture was utilized to segment up to 5 structures in each CBCT and planning CT image set. More than 100 structures were segmented per imaging modality. DL segmentation accuracy was compared to reference segmentations by a human observer using the Dice similarity coefficient (DSC) and Hausdorff distance (HD).

Results: For five of the sixteen unique anatomical structures, qCBCT provided higher segmentation accuracy than standard-of-care CBCT images. These structures were liver, kidney, prostate, bladder, and rectum. Their average DSC values were 0.73 ± 0.18, 0.74 ± 0.17, and 0.81 ± 0.13 for standard CBCT, iCBCT, and qCBCT, respectively (p < 0.001). Likewise, average HD values were 3.26 ± 0.75 mm, 3.28 ± 0.87 mm, and 2.96 ± 0.85 mm for standard CBCT, iCBCT, and qCBCT, respectively (p < 0.001). Planning CT had higher segmentation accuracy than both CBCT modalities, with an average DSC of 0.83 ± 0.14 and HD of 2.81 ± 0.85 mm. For other structures, DSC and HD values of qCBCT and stand-of-care CBCTs were not significantly different (p>0.05). Overall, H&N structures had a mean DSC of 0.48 ± 0.22, while pelvic/abdominal structures had 0.70 ± 0.23. In a subset analysis, structures with larger cross-sectional area, such as bladder and liver had significantly higher segmentation accuracy in all imaging modalities combined (DSC: 0.78 ± 0.16) compared to smaller structures (DSC: 0.52 ± 0.25), such as lymph nodes, submandibular glands, and common vessels. Lower DSC values of H&N structures were attributed to their small cross-sectional area.

Conclusion: Proposed qCBCT significantly improved Deep Learning based segmentation accuracy for liver, kidney, prostate, bladder, and rectum, when compared to standard and advanced standard-of-care CBCTs. These results indicate that quantitatively accurate CBCT images can potentially improve the efficacy of automated segmentation workflows in CBCT-based treatment dose delivery monitoring.