Multi-stream Progressive Up-Sampling Network for Dense CT Image Reconstruction.

Qiuyue Liu,Zhen Zhou,Feng Liu,Xiangming Fang,Yizhou Yu,Yizhou Wang
DOI: https://doi.org/10.1007/978-3-030-59725-2_50
2020-01-01
Abstract:Pulmonary computerized tomography (CT) images with small slice thickness (thin) is very helpful in clinical practice due to its high resolution for precise diagnosis. However, there are still a lot of CT images with large slice thickness (thick) because of the benefits of storage-saving and short taking time. Therefore, it is necessary to build a pipeline to leverage advantages from both thin and thick slices. In this paper, we try to generate thin slices from the thick ones, in order to obtain high quality images with a low storage requirement. Our method is implemented in an encoder-decoder manner with a proposed progressive up-sampling module to exploit enough information for reconstruction. To further lower the difficulty of the task, a multi-stream architecture is established to separately learn the inner- and outer-lung regions. During training, a contrast-aware loss and feature matching loss are designed to capture the appearance of lung markings and reduce the influence of noise. To verify the performance of the proposed method, a total of 880 pairs of CT images with both thin and thick slices are collected. Ablation study demonstrates the effectiveness of each component of our method and higher performance is obtained compared with previous work. Furthermore, three radiologists are required to detect pulmonary nodules in raw thick slices and the generated thin slices independently, the improvement in both sensitivity and precision shows the potential value of the proposed method in clinical applications.
What problem does this paper attempt to address?