Author: Hoyeon Lee 👨🔬
Affiliation: University of Hong Kong 🌍
Purpose: Deep-learning approaches are widely investigated for Cone-Beam CT (CBCT) scatter correction to improve the quality of the linear-accelerator mounted CBCT. This study aims to optimize the deep-neural-network (DNN) structure utilizing the network architecture search (NAS) approach to further improve deep-learning-based CBCT scatter correction performance.
Materials & Methods: We collected diagnostic CT images of 38 H&N patients from The Cancer Imaging Archive. Elekta’s XVI CBCT system was modeled using the MC-GPU, a GPU-accelerated Monte Carlo code for CBCT simulation, to obtain projection data and corresponding scatter-free projection data. Out of 38 patients, 31 were utilized for training and internal validation, while 7 were reserved for external testing. We employed U-Net, a network structure commonly adopted for image processing, as a baseline network architecture and trained networks to estimate the scatter signals from the projection data. Kernel sizes of convolution layers as well as types of activation, normalization, down-sampling, and up-sampling layers were tuned via NAS to minimize the mean-absolute-percentage-error (MAPE). The optimization employed a grid search algorithm over 100 iterations. The model architecture yielding the lowest MAPE on the internal validation dataset was selected as the optimal architecture. Scatter correction in the external testing data was performed using both the baseline and optimal architectures for comparison.
Results: The scatter-corrected and scatter-free projection data were reconstructed employing the FDK algorithm and the MAPE was calculated. The mean and standard deviation of the MAPE for the baseline and optimized architectures were 6.42 HU ± 2.12 and 4.34 HU ± 1.38, respectively. The optimized architecture achieved 32% improvement in scatter correction performance compared to the baseline network on the same dataset.
Conclusion: We employed NAS to optimize the DNN structure for CBCT scatter correction. This study demonstrated the potential of task-specific DNN architecture optimization to improve the CBCT scatter correction performance.