PIDA-Net: A prior image-guided deformable attention network for high-quality 4D-CBCT reconstruction.
Authors
Affiliations (2)
Affiliations (2)
- School of Biomedical Engineering, Southern Medical University, China; Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, China.
- School of Biomedical Engineering, Southern Medical University, China; Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, China; Guangdong Province Engineering Laboratory for Medical Imaging and Diagnostic Technology, Southern Medical University, China. Electronic address: [email protected].
Abstract
In thoracic and abdominal radiotherapy, acquiring high-quality four-dimensional cone beam computed tomography (4D-CBCT) images is essential for accurate tumor localization and motion management. However, conventional 4D-CBCT reconstructions often suffer from streak artifacts and noise due to undersampled projection data at each respiratory phase. To address this challenge, this study presents a prior image-guided deformable attention network (PIDA-Net) for generating high-fidelity 4D-CBCT reconstructions with enhanced anatomical accuracy. In the proposed framework, a prior reference image is first constructed-either the averaged image over the entire respiratory cycle or the planning CT image selected for its structural similarity to the target phase. The PIDA-Net then employs a dual-branch convolutional network to hierarchically extract multi-scale features from both the original 4D-CBCT data and the selected reference image. With the deformable attention module, the PIDA-Net dynamically identifies and weights the most relevant anatomical features from the high-quality prior image, thereby enabling context-aware fusion of complementary anatomical information from the reference image to effectively restore artifact-degraded structures in the target image. Experimental results on both simulated and clinical patient datasets reveal that the PIDA-Net significantly improves 4D-CBCT image quality by suppressing streak artifacts while preserving fine anatomical details.