Projection‐domain scatter correction for cone beam computed tomography using a residual convolutional neural network

• DOI: 10.1002/mp.13583
• Publication Date: 2019-05-11T15:59:16Z
• Authors: Yusuke Nomura; Qiong Xu; Hiroki Shirato; Shinichi Shimizu; Lei Xing
• Keywords: Kernel (algebra)

Scatter is a major factor degrading the image quality of cone beam computed tomography (CBCT). Conventional scatter correction strategies require handcrafted analytical models with ad hoc assumptions, which often leads to less accurate removal. This study aims develop an effective method using residual convolutional neural network (CNN).A U-net based 25-layer CNN was constructed for CBCT correction. The establishment model consists three steps: training, validation, and testing. For total 1800 pairs x-ray projection corresponding scatter-only distribution in nonanthropomorphic phantoms taken full-fan scan were generated Monte Carlo simulation scanner installed proton therapy system. An end-to-end training implemented two loss functions 100 epochs mini-batch size 10. Image rotations flips randomly applied augment datasets during training. 200 projections digital head phantom collected. proposed CNN-based compared conventional projection-domain named fast adaptive kernel superposition (fASKS) 360 anthropomorphic phantom. Two different same evaluate impact on final results. Furthermore, trained fine-tuned half-fan by transfer learning additional phantoms. tuned-CNN fASKS as well without fine-tuning lung projections.The provides significantly reduced images more Hounsfield Units (HUs) than that fASKS-based method. Root mean squared error CNN-corrected improved 0.0862 0.278 uncorrected or 0.117 fASKS-corrected projections. reconstruction provided better HU quantification, especially regions near air bone interfaces. All four measures, include absolute (MAE), (MSE), peak signal-to-noise ratio (PSNR), structural similarity (SSIM), indicated images. Moreover, technique made it possible be applicable remove scatters after only small number datasets. SSIM value tuned-CNN-corrected 0.9993 0.9984 non-tuned-CNN-corrected 0.9990 Finally, computationally efficient - time took 5 s reported experiments PC (4.20 GHz Intel Core-i7 CPU) single NVIDIA GTX 1070 GPU.The deep learning-based tool holds significant quantitative imaging image-guided radiation therapy.