2306 - Development and Evaluation of Deep Learning-Based CBCT Enhancement in Automatic OAR Segmentation

Q. Hu¹, X. Feng², Y. Chen¹, Y. Zhu¹, J. Li³, H. Lin¹, S. Song³, L. Zhao⁴, and X. Zhang¹; ¹Department of Radiation Oncology, Tong Ren Hospital Affiliated to Medical College of Shanghai Jiao Tong University, Shanghai, China, ²Carina Medical LLC, Lexington, KY, ³Hangzhou Fortunelight Fuzhao Technology Co., Ltd., Hangzhou, China, ⁴Zhejiang University, Hangzhou, China

Purpose/Objective(s): Cone-beam computed tomography (CBCT) imaging plays an important role in the treatment delivery. However, due to the limited image quality, it cannot provide accurate guidance in online plan evaluation, which is a limiting factor for adaptive radiation therapy. Deep learning has shown the potential to improve CBCT image quality, but most studies have focused on the image-based metrics alone and did not evaluate the impact to key aspects in plan evaluation such as organ-at-risk (OAR) segmentation. This study developed a CBCT enhancement model and compared the OAR segmentations before and after the enhancement using an AI-based automatic segmentation product. Furthermore, the datasets used to train and test the enhancement network are from completely different institutions to avoid overfitting and gain a more objective evaluation. Materials/

Methods: The training data contains 40 paired abdominal cases from TCIA public dataset, 180 head and neck and 180 pelvic cases from SynthRad2023 challenge, and 28 thoracic cases from Learn2Reg2023 public dataset. A body-part-agnostic model was trained based on 2D U-Net with residual encoder. A masked l1-loss was leveraged during training to reduce errors from mis-alignment due to respiratory motion and anatomy changes. Specifically, the voxels with over 400 HU difference between CBCT and CT were ignored when calculating l1-loss. Testing dataset was collected from a completely independent data source containing 11 subjects from our institution of pelvic regions. A commercial auto-segmentation software was applied to original CBCT and enhanced CBCT separately. Dice similarity coefficient (DSC) was computed between ground truth labels and automatic segmented labels. Translation-based registration was applied to obtain aligned ground truth labels for CBCT from CT.
Results: Visually, CBCT enhancement showed improvement in soft tissue HU consistency with less artifacts, and more visible soft tissue on the same HU window level. There was a statistically significant improvement of 0.119 DSC for rectum with p-value=0.024.
Conclusion: The proposed CBCT enhancement model improved the image quality of CBCT evaluated from a separate dataset. For small-sized OARs, like rectum, whose segmentation may depend on the image contrast, the proposed model significantly helped the auto-segmentation performance on CBCT. Abstract 2306 – Table 1

OAR	DSC_Original	DSC_Enhanced	DSC_Improvement	p-value (one-tail to test if there is substantial improvement)
Bladder	0.722	0.742	0.020	0.208
Rectum	0.453	0.572	0.119	0.024
FemurHead	0.604	0.639	0.035	0.309
BoneMarrow	0.669	0.675	0.006	0.371