Click here to


Are you sure ?

Yes, do it No, cancel

BEST IN PHYSICS (IMAGING): Comparison of Loss Functions in Dual-Domain Convolutional Neural Networks for Low-Dose CT Enhancement

KJ Chung1-3*, R Souza4,5, R Frayne4,5, TY Lee1-3, (1) University of Western Ontario, London, ON, CA, (2) Robarts Research Institute, London, ON, CA, (3) Lawson Health Research Institute, London, ON, CA, (4) University Of Calgary, Calgary, AB, CA, (5) Foothills Medical Centre, Calgary, AB, CA


(Sunday, 7/12/2020) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Room: Track 1

Purpose: To compare the performance of denoising low-dose CT images using convolutional neural networks (CNNs) trained with mean squared error (MSE) or a perceptual loss function.

Methods: The CNN proposed for low-dose CT enhancement is a cascade of U-nets operating in the spatial frequency domain first then the image domain. Low-dose inputs were transformed by a 2D Fourier Transform (FT) and processed by the first U-net. Following an inverse FT operation, the intermediate spatial frequency-domain result was then processed by the second image-domain U-net, resulting in the denoised low-dose CT image. The perceptual loss function was defined as a weighted sum of multi-scale structural similarity (SSIM) and mean absolute error. Loss was computed on the intermediate and final results in their respective domains with greater weight given to the final output’s loss. We compared the performance of networks trained with (1) MSE-only, (2) frequency-domain MSE and image-domain perceptual loss, and (3) perceptual loss-only. Routine- and quarter-dose contrast-enhanced CT images of the abdomen of 10 patients from the AAPM Low-dose Grand Challenge were used.

Results: Deep learning processing substantially suppressed noise compared to low-dose inputs. Networks involving an MSE loss had poorer image contrast and contrast-enhancement in the liver appeared less distinct compared to the perceptual loss-only network. SSIM and peak signal-to-noise ratio improved from 0.899±0.047 to 0.953±0.023, 0.953±0.020, and 0.961±0.019; and 38.62±3.08 dB to 42.66±2.78 dB, 42.35±2.34 dB, and 43.39±2.68 dB in the order of unprocessed and MSE-only, MSE/perceptual, and perceptual-only network processed images, respectively.

Conclusion: Perceptual loss for spatial frequency- and image-domain processing was superior to MSE for low-dose CT denoising. High frequency noise could be filtered while remaining sensitive to true high frequency image features. The investigated perceptual loss function is a versatile tool with applications beyond spatial domain optimization.

Funding Support, Disclosures, and Conflict of Interest: Operational funding: Natural Sciences and Engineering Research Council of Canada (NSERC); Cloud computing services: Amazon Web Services Cloud Credits for Research


Reconstruction, CT, Low-dose CT


IM- CT: Image Reconstruction

Contact Email