(Tuesday, 7/16/2019) 4:30 PM - 6:00 PM
Room: 225BCD
Purpose: CT dose reduction is important to reduce x-ray induced cancer risks. As dose is reduced, reconstructed image quality is degraded due to amplified noise in sinogram. Despite remarkable success achieved in image domain noise reduction, sinogram denoising remains a challenge, as data inconsistency after denoising is amplified by the reconstruction process, deteriorating image quality. To overcome this challenge and achieve effective noise reduction in sinogram, we develop a reconstruction–aware deep U-net based sinogram denoising framework that incorporates the subsequent CT reconstruction task in the denoising process.
Methods: The proposed denoising network utilizes the standard U-net structure that maps a noisy sinogram into a clean one. To make denoising be aware of the reconstruction task, the filter back projection (FBP) step is attached to the output layer of the denoising U-net as a layer of a known operation. A fused loss combining errors in both reconstructed image domain and sinogram domain is propagated back to train the parameters of the denoising U-net. A larger weight is assigned to the image domain error to emphasize the reconstruction quality in network training. To demonstrate the effectiveness of the proposed approach, we compare its performance to that of another U-net with identical structure but trained based on only the loss in the sonogram domain.
Results: On both training and testing data, average error of the reconstructed image using the sinogram generated by the proposed method is ~28 HU, lower than the error of ~32 HU for the denoising using a U-net not aware of the reconstruction task. Mean peak signal to noise ratio value is improved by 1.73dB and 2.08dB for training and testing data, respectively.
Conclusion: The deep learning-based reconstruction-aware framework for sinogram denoising can effectively reduce image noise and achieve superior performance to the standard deep learning-based sinogram denoising method.
Keywords
Taxonomy
Not Applicable / None Entered.
Contact Email
