Click here to


Are you sure ?

Yes, do it No, cancel

Can Deep Learning Raise the Quality of Low-Dose CT Images Above That of Normal-Dose CT Images?

T Bai1*, D Nguyen2, G Wang3, S Jiang4, (1,2,4) The University of Texas Southwestern Medical Ctr, Dallas, TX, (3) Rensselaer Polytechnic Institute, Troy, NY


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: work is to investigate the possibility whether deep learning can raise the low-dose CT image quality above that of normal-dose CT.

Methods: maintain the inherent high-resolution property despite of strong noise in the origin low-dose CT image, a resolution-maintenance path is required by using non-downsample convolution operators. To remove the noise artifact effectively, strong semantic features are required to describe the intrinsic pixel correlations, which can be achieved by using repeated downsample convolution operators for larger receptive field. Furthermore, multiscale features should also be extracted to further enhance the semantic features by fusing the low-level and high-level features together. Combining the above three essential factors together, a high-resolution network (HRNet) was introduced, which by design can deliver an effectively noise-suppressed image without compromise the resolution. The group normalization (GN) and weight standardization (WS) techniques were also used to stabilize the training phase. The Low-dose CT Challenge 2016 dataset was employed for validation. For comparison, the UNet equipped with instance normalization (UNet + IN) served as the baseline. The root-mean-square-error (RMSE) and the structural similarity index (SSIM) were used as the quantitative metrics.

Results: inspection indicated that HRNet could produce much lower noise levels than both the quarter/normal dose CT images, but exhibited comparable resolution property as the normal-dose CT image. Moreover, the HRNet also outperformed the competing UNet based algorithm, regarding the better-preserved fine structures as well as the removed “bone-cliff” artifacts. In quantitative comparison, the UNet + IN baseline model had a SSIM/RMSE of 0.7079/62.78, which was boosted to 0.7380/55.55 by replacing the UNet with the HRNet, and was further improved to 0.7460/54.51 by using the HRNet + WS + GN model.

Conclusion: learning is able to raise the quality of low-dose CT images above that of normal-dose CT images by using the HRNet.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email