ESTRO 2025 - Abstract Book

S2994

Physics - Image acquisition and processing

ESTRO 2025

1929

Digital Poster Prior knowledge-induced CBCT reconstruction from single-view projection for adaptive radiotherapy Danyang Li, Li Lin, Xiaoyan Huang, Ying Sun Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, China Purpose/Objective: Cone-beam computed tomography (CBCT) is a commonly used imaging modality in adaptive radiotherapy (ART) to monitor variations arising from inter- and intrafraction, and improve the accuracy and efficiency of the radiation therapy treatment [1]. Online ART requires fast imaging of the anatomies of patients. However, the current on board CBCT scanner usually takes about one minute to finish a full scan, which is far from real-time imaging. To address this challenge, we presented a Prior Knowledge-Induced deep-learning model for CBCT reconstruction, shortened as PKI-CBCT, with single-view projection data. The proposed model achieved fast CBCT scanning and reconstructing, which enhanced the ART process and reduced the radiation exposure to the patients. Material/Methods: The presented PKI-CBCT model contains four components, i.e., the encoder, transfer, generator, and prior modules, as shown in Figure 1. 1610 paired data (planning CT images, CBCT images, and projections) from patients with nasopharyngeal carcinoma (NPC) were retrospectively collected to establish the whole dataset, where 1120 paired data for training and the remaining 490 paired data for testing. A specific model without inducing planning CT, shortened as W/O-PKI, was evaluated to gain more insights into the importance of prior knowledge. Mean absolute error (MAE), structure similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR) were utilized to evaluate the performances of the methods.

Results: Figure 2 shows the results of the W/O-PKI and proposed PKI-CBCT. It can be seen that W/O-PKI produces oversmoothed results. In contrast, the PKI-CBCT reconstructs high-quality CBCT images preserving spatial resolution. In addition, the PKI-CBCT model achieves an average MAE of 37±11.6 HU (mean±std), SSIM of 0.9508±0.0184, and PSNR of 38.40±1.41 dB on the whole testing dataset, respectively. The W/O-PKI model achieves an average MAE of 55±20.4 HU, SSIM of 0.8791±0.0251, and PSNR of 27.52±2.76 dB, respectively. These quantitative results also illustrate that the proposed PKI-CBCT model is better than the W/O-PKI model indicating that the integration of planning CT images is significant for reconstructing desired CBCT images.