Daily edge deformation prediction using an unsupervised convolutional neural network model for low dose prior contour based total variation CBCT reconstruction (PCTV-CNN). (7th October 2019)

• Record Type: Journal Article
• Authors: Chen, Yingxuan; Yin, Fang-Fang; Jiang, Zhuoran; Ren, Lei
• Is Part Of: Biomedical physics & engineering express. Volume 5:Number 6(2019)
• Publication Date: 2019-10-07
• Subjects: low dose CBCT reconstruction -- prior contour based total variation (PCTV) -- unsupervised convolutional neural networks (CNN); Medical physics -- Periodicals; Biophysics -- Periodicals; Biomedical engineering -- Periodicals; Medical sciences -- Periodicals; 610.153
• DOI: 10.1088/2057-1976/ab446b
• Languages: English
• ISSNs: 2057-1976

Abstract

Abstract: Purpose : Previously we developed a PCTV method to enhance the edge sharpness for low-dose CBCT reconstruction. However, the iterative deformable registration method used for deforming edges from planning-CT to on-board CBCT is time-consuming and user-dependent. This study aims to automate and accelerate PCTV reconstruction by developing an unsupervised CNN model to bypass the conventional deformable registration. Methods : The new method uses unsupervised CNN model for deformation prediction and PCTV reconstruction. An unsupervised CNN model with a u-net structure was used to predict deformation vector fields (DVF) to generate on-board contours for PCTV reconstruction. Paired 3D image volumes of prior CT and on-board CBCT are inputs and DVF are predicted without the need of ground truths. The model was initially trained on brain MRI images, and then fine-tuned using our lung SBRT data. This method was evaluated using lung SBRT patient data. In the intra-patient study, the first n −1 day's CBCTs are used for CNN training to predict nth day edge information ( n = 2, 3, 4, 5). 45 half-fan projections covering 360̊ from nth day CBCT is used for reconstruction. In the inter-patient study, the 10 patient images including CT and first day's CBCT are used for training. Results from Edge-preserving (EPTV), PCTV and PCTV-CNN are compared. Results : The cross-correlations of the predicted edge map and the ground truth were on average 0.88 for both intra-patient and inter-patient studies. PCTV-CNN achieved comparable image quality as PCTV while automating the registration process and reducing the registration time from 1–2 min to 1.4 s. Conclusion : It is feasible to use an unsupervised CNN to predict daily deformation of on-board edge information for PCTV based low-dose CBCT reconstruction. PCTV-CNN has a great potential for enhancing the edge sharpness with high efficiency for low-dose CBCT to improve the precision of on-board target localization and adaptive radiotherapy.