| Abstract: |
Cone beam computed tomography (CBCT) is valuable for assessing anatomical changes between treatment sessions throughout the entire course of image-guided radiotherapy (IGRT). However, artifacts from various sources can limit the full utilization of CBCT in adaptive radiotherapy (ART). Inter-fraction anatomical changes during ART, such as variations in tumor size and normal tissue anatomy, can impact the effectiveness of radiation therapy. Acquiring high-quality CBCT images that accurately reflect patient-specific (PS) anatomical changes is essential for successful IGRT. To improve CBCT image quality, we developed patient-specific lung diffusion models (PS-LDMs). These models build on a pre-trained general lung diffusion model (GLDM), trained using historical paired data of deformed CBCT (dCBCT) and planning CT (pCT). For each patient, a new PS model is a general diffusion model fine-tuned with low-rank adaptation (LoRA) strategy to obtain additional PS knowledge which is learned from CBCT and deformed pCT (dpCT) pairs of each newly collected treatment fraction. This knowledge enables the generation of personalized synthetic CT (sCT) images with quality comparable to pCT or dpCT. The PS knowledge captures specific mapping relationships, including personal inter-fraction anatomical changes reflected in the CBCT-dpCT pairs. © 2025 SPIE |
