MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Residual Learning Based Projection Domain Denoising for Low-Dose CT

Y Zhang1 , R MacDougall2*, H Yu1 , (1) UMass Lowell, Lowell, MA,(2) Boston Children's Hospital, Boston, MA

Presentations

(Sunday, 7/29/2018) 3:00 PM - 6:00 PM

Room: Exhibit Hall

Purpose: Deep learning has achieved promising results in CT denoising in image domain. In this work, we apply a residual learning of convolutional neural network to reduce noise in projection domain for low-dose CT.

Methods: We design a five-layer convolutional neural network for sinogram denoising. To train to neural network, both noisy and noise-free are simulated. To ensure the generated sinogram to be close to realistic ones, we apply high-dose clinical CT images as phantoms. In all simulations, a 2D fan-beam CT geometry is performed. Poisson noise is considered and the number of photons is set to 2×10�. We randomly extract 2×10� patches with the size of 64×64 from the noisy and noise-free sinograms. Noisy patches are assumed as the input of network and the differences between the noisy and noise-free pairs are the target of network. Hence, the well trained network is capable to estimate the noise, and a clean sinogram can be obtained by subtracting estimated noise from the original sinogram. The network is obtained after 400 training epochs. Finally, effectiveness of the proposed method is validated on both numerical simulated data and real scanned data.

Results: It took ~4 hours to train the network and the denoising time was ~1 second for each sinogram. The average standard deviation of the network processed sinograms was significantly reduced to compared to the original sinograms. Moreover, most of photon starvation induced streaks in reconstructed images were removed.

Conclusion: The results demonstrate that the proposed residual learning method can achieve a remarkable denoising performance in terms of efficiency and image quality.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email