Author: Ziqi Gao, Lei Xing, Siqi Ye, S. Kevin Zhou 👨🔬
Affiliation: Department of Radiation Oncology, Stanford University, University of Science and Technology of China (USTC) 🌍
Purpose: To address the challenge of high memory usage in volumetric cone-beam CT (CBCT) imaging, we propose a method that combines joint reconstruction and super-resolution for sparsely sampled CBCT using implicit neural representation (INR). This approach enables, for the first time, high-quality CBCT imaging on a consumer-level GPU with just 24 GB of memory.
Methods: We introduce a joint optimization framework for sparsely sampled reconstruction and super-resolution enabled by a memory-efficient, self-supervised INR approach. Specifically, we propose a multi-view physics consistency optimization technique, which enables memory-efficient optimization of INR and enforces the physical constraints on a reconstructed volume of a high resolution. We demonstrate the effectiveness of our method on 10 patient cases from the public Lung Image Database Consortium (LIDC) image collection dataset [2] and a pancreas 4D CT dataset. The low resolution (LR) - high resolution (HR) volume pair is constructed by resizing the volumes into (128, 128, 40) and (512, 512, 40). Sinogram projection is simulated using the LR volume. PSNR and SSIM are used for quantitative evaluation.
Results: The results on 512×512×40 volumes from LIDC and the pancreas CT data demonstrate that our multi-view physics consistency optimization successfully improves the visual quality of NeRP at the 512×512 resolution, at the cost of a longer training time. The improved visual quality on a higher spatial resolution indicates INRs’ potential for more accurate disease diagnosis under the current hardware constraints. The reconstruction on a consumer-level GPU showcases the potential of INRs for real-world deployment.
Conclusion: This work provides a novel solution for high-resolution INR reconstruction of CBCT with a consumer-level GPU. Quantitative and qualitative results of patient case studies showcase its strong potential for clinical deployment. This methodology is extendable to many other image reconstruction tasks (e.g., accelerated and memory-efficient MR imaging).