Author: Penghao Gao, Zejun Jiang, Huazhong Shu, Linlin Wang, Gongsen Zhang, Jian Zhu 👨🔬
Affiliation: Laboratory of Image Science and Technology, Key Laboratory of Computer Network and Information Integration, Southeast University, Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Artificial Intelligence Laboratory, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences 🌍
Purpose: We propose a cascading framework for time-varying anatomical volumetric reconstruction and tumor-tracking, guided by onboard orthogonal-view X-ray projections.
Methods: We employe multiple deep learning components to decompose the volumetric tumor mask inference into two subtasks: i) volumetric reconstruction from planar projections, and ii) tumor localization on the reconstructed CT, to facilitate anatomical dimensionality elevation and phased regulation of component performance. Embedded with convolutional block attention modules (CBAM), a conditional generative-adversarial network (cGAN) is employed as backbone of the reconstruction network, for which filtered back projection (FBP) is improved to compensate for highly limited anatomical information from extremely sparse-view projections. For tumor localization, we use a hybrid network integrating Swin-Transformer and CNN blocks to perform segmentation on the reconstructed CT. This component is enhanced by full-scale fused features from the reconstruction component. Additionally, the framework incorporates planning images and clinical delineation-based priori-knowledge. For deep learning development, validation, and evaluation pipeline, we enrolled 312 real patients with non-small cell lung cancer (NSCLC).
Results: The peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) of the reconstructed CTs were 31.02 dB and 0.95, respectively, while tumor localization achieved a centroid deviation amplitude of 0.47 mm and a Dice similarity coefficient (DSC) of 0.96. Experiments indicated that the number of sparse-view projections, rather than angles, affects reconstruction performance. Effectiveness of improvements in network settings and framework workflow optimization are validated by ablation experiments. The FBP enhancement and the embedded CBAM modules improved the PSNR of reconstruction by 2.94 dB and 0.14 dB, respectively. The cascading localization and full-scale feature fusion strategies improved the DSC of tumor-tracking by 0.06 and 0.05, respectively. The total time taken for time-varying reconstruction and tumor-tracking was 197.35±7.12 ms.
Conclusion: With a cascading volumetric reconstruction and tumor-tracking pipeline, our proposed deep learning framework represents a promising step toward image-guided adaptive intra-fractional radiotherapy.