Author: Ziheng Deng, Yao Hao, Runping Hou, Deshan Yang, Jun Zhao, Yufu Zhou 👨🔬
Affiliation: Department of Radiation Oncology, Duke University, School of Biomedical Engineering, Shanghai Jiao Tong University, Washington University School of Medicine, Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine 🌍
Purpose: 4D CBCT has been developed to provide dynamic images for image-guided radiation therapy. However, as projection data are sorted into sparse and clustered phase-specific bins, 4D CBCT images are often degraded by streak artifacts. This study intends to generate high-quality 4D CBCT images for patients with upper abdominal cancer and improve image-guided radiation therapy.
Methods: Breathing signals were extracted from CBCT projection data, followed by gated-FDK reconstruction to produce phase-resolved CBCT images. A 4D convolutional neural network model named RSTAR4D-Net was developed to reduce the streak artifacts induced by sparse and clustered phase binning. The RSTAR4D-Net took 4D gated-FDK images as input and produced the artifact-removed images. A lightweight separable 4D convolution module was introduced, which enabled the model to process all phase-gated 4D CBCT images as a whole 4D volume and fully utilize the correlated image information among all 4D phases. The model was trained on simulated 4D thoracic-abdominal CBCT datasets in a supervised manner.
Results: Qualitative and quantitative evaluations were conducted on a simulated free-breathing abdominal dataset and a CIRS 4D thoracic motion phantom. The proposed method removed most streak artifacts and improved overall image quality. The enhanced 4D CBCT images displayed clearer structural details and more natural motion patterns. Quantitative metrics demonstrated that RSTAR4D-Net outperformed comparison methods.
Conclusion: The proposed RSTAR4D-Net algorithm showed the capability to generate high-quality 4D CBCT images without altering the clinical standard one-minute scan protocol. It provided an effective and efficient solution for 4D CBCT dynamic images recovery. In future work, we will evaluate the generalization of this method on additional clinical datasets and investigate its applications in downstream tasks such as upper abdominal tumor motion tracking for guiding online adaptive radiation therapy.