Feature-Aware Deep-Learning Reconstruction for Context-Sensitive X-ray Computed Tomography

It is common for CT images to be reconstructed differently for different clinical examination purposes. It is difficult for conventional filtered backprojection (FBP) methods and standard model-based iterative reconstruction (MBIR) methods to produce a single context-sensitive image without multiple reconstructions. In this article, we address this challenge by leveraging the power of deep learning. We propose to train a deep convolution neural network to reconstruct a universal image from one FBP image. We present a new data argumentation method that generates the specific training target image, specifically the feature-aware target. The resulting method, called feature-aware deep-learning reconstruction (DLR), requires only one FBP image as input and is much faster than MBIR. In experiments, we investigate one application of low-dose CT feature-aware DLR which aims to achieve noise and resolution consistency across different body parts. We evaluate the performance of the proposed method using both simulated and real clinical low-dose CT scans. The results show that our feature-aware DLR outperforms both FBP and standard MBIR by producing improved CT image quality potentially suitable for a broad range of clinical diagnoses.