Real Time 4D-Cone Beam CT via Single-Angle Projection with Dual Attention Mechanism

Abstract

The hypothesis is that dual attention integrated Residual Network (DA-ResNet) together with a Principal Component Analysis (PCA) model accurately maps the measured on-treatment CBCT projections from different phases to each phase's 3D-CBCT real timely. Finally, it achieves real time 4D-CBCT. The DA-ResNet contains a CBAM (Convolutional Block Attention Module) and a SPA (Spatial Pyramid Structure). Hundreds of forward projections from one view of patient 4D-CT were simulated to train DA-ResNet. The extracted motion features were expressed as PCA coefficients. To test DA-ResNet performance, several single measured on-treatment projection is sent into it and the network output several groups of PCA coefficients real timely. Each group of PCA coefficient stands for the corresponding deformation field between the 0% phase and the predicted phase that the PCA label stands for. In labels such as Mean Absolute Error (MAE), R square, Normalized Cross Correlation (NCC), Structural Similarity Index Measure (SSIM) were used to evaluate DA-ResNet performance with other state-of-the-art networks such as CNN/Unet/ResNet. DA-ResNet outperforms other popular networks on all of the quantification labels, with MAE (14.41 ± 8.85), R square (0.95 ± 0.07) in PCA coefficient prediction, and NCC (0.9981 ± 0.0006), SSIM (0.98 ± 0.04) in image reconstruction. Compared with conventional ResNet, DA-ResNet reduced MAE by 79%. Meanwhile R square rises from 0.88 to 0.95, NCC rises from 0.996 to 0.998, SSIM rises from 0.94 to 0.98. DA-ResNet offers a solution for real time on-treatment 4D-CBCT image guidance by using single 2D on-line projections. It offers better motion prediction and image reconstruction accuracy compared with traditional deep learning networks (e.g., CNN/Unet/ResNet). The initial results are promising. The prospect of clinical application needs further clinical verification.