Cancer Hospital Chinese Academy of Medical Science, Shenzhen Center Shenzhen, Guangdong
J. Dang; National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital & Shenzhen Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Shenzhen, China
Purpose/Objective(s): To incorporate the temporal-spatial correlation into the prediction of real time on-line 4D-CBCT prediction process for lung cancer on-treatment imaging. Materials/
Methods: We develop a cascade network that contains: 1) a Convolutional LSTM (ConvLSTM); and 2) a Principal Component Analysis (PCA) model with a prior deformation vector fields (DVFs) from patient 4D-CT. The 4D-DVFs are further transferred into PCA label groups and expanded with 15% motion amplitudes. The expanded PCA labels are further random sampled into ~1000 label groups. Then hundreds of deformed CBCTs with continuous motion are obtained via DVFs mapped from PCA label groups. Finally the forward projection under a fixed angle will generate hundreds of Digital Reconstructed Radiography (DRRs) form CBCTs for network training. For testing, a on-line measured 2D projection is sent into trained network to predict 3 PCA labels. Then a deformed 3D-CBCT will be obtained via deformable image registration with DVF mapped from predicted PCA labels. After measuring several 2D projections that covers the whole respiration cycle, a full 4D-CBCT can be obtained. We performed a digital XCAT phantom experiment followed with an initial 5 patient data validation. Quantification labels of FSIM (Feature Similarity Index Measure), PSNR (Peak Signal-to-Noise Ratio), MSSIM (Multi-scale Structural Similarity Index Measure) are used for parallel network (CNN/Unet/ResNet/ConvLSTM) performance evaluations. Results: ConsLSTM outperforms all of the parallel networks for both of the XCAT phantom and clinical experiments. For phantom experiment, ConvLSTM achieves the highest quantification accuracy with FSIM, PSNR, MSSIM of 0.9998, 64.6742, 0.9998, respectively. For patient clinical evaluation, ConvLSTM also achieves the best quantification accuracy with that average value of 0.9999, 63.7294, 0.9999, respectively. Conclusion: ConvLSTM supplied a promising solution for accurate real time 4D-CBCT prediction.
