A novel convolutional neural network with high convergence rate: Application to CT synthesis from MR images

Abstract

Synthetic CT (sCT) generation from MR images is yet one of the major challenges in the context of MR-guided radiation planning as well as quantitative PET/MR imaging. Deep convolutional neural networks have recently gained special interest in large range of medical imaging applications including segmentation and image synthesis. In this study, a novel deep convolutional neural network (DCNN) model is presented for synthetic CT generation from single T1-weighted MR image. The proposed method has the merit of highly accelerated convergence rate suitable for applications where the number of training da-taset is limited while highly robust model is required. This algorithm exploits a Visual Geometry Group (VGG16) model without fully connected layer coupled to a residual network in the form of encoder-decoder structure. The training of the proposed algorithm was performed using pelvis image of only 15 patients in a five-fold cross-validation scheme. No network pre-training and data augmentation was used. The outcome of the proposed algorithm was compared to an atlas-based approach in terms of accuracy of CT intensity estimation within different body tissues. After only 100 epochs, the proposed algorithm resulted in mean absolute error (MAE) and mean error (ME) of 40.64 ± 12.66 and -2.80 ± 10.98 (HU) for the entire pelvis region, respectively. While atlas-based method led to MAE of 82.06 ± 52.59 and ME of -13.00 ± 60.19 (HU). Within the soft-tissue the atlas-based method and the proposed algorithm achieved MAEs of 51.01±14.3 and 25.51±7.72 (HU), respectively. Likewise in bony tissue, MAEs of 212.65±78.45 and 221.99±76.28 (HU) were obtained when using DCNN and atlas-based methods, respectively. The proposed algorithm showed superior performance to the atlas-based method with only relying on a limited number of training subjects. The proposed algorithm is suitable for the clinical applications where accurate models are required while accessing a large number of training cases is limited.