MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Incorporating L0-Norm Regularization Into Cycle-Consistent Generative Adversarial Network (cGAN) for MR-To-CT Translation

H Kim1*, H Lee2 , J Kwak1 , C Jeong1 , B Cho1 , (1) Department of Radiation Oncology, Asan Medical Center, Seoul, Seoul, (2) Korea Advanced Institute of Science and Technology, Yuseong-gu, Daejeon

Presentations

(Monday, 7/30/2018) 4:30 PM - 6:00 PM

Room: Karl Dean Ballroom C

Purpose: Combining cycle-consistency with generative adversarial networks (cGAN) enhanced the accuracy of MR-to-CT image translation possibly for MR-only radiotherapy, relative to conventional GAN consisting of a generator and a discriminator. This work further promotes the translational accuracy in cGAN by modifying the loss function based on the notion of L0-norm minimization.

Methods: Though GAN proposed a salient way to create a synthetic image that looks real for image-to-image (X-to-Y) translation, the application to MR-to-CT translation could be limited due to possibly misaligned, unpaired MR-to-CT image datasets. Recently, cGAN was developed to overcome such a drawback by configuring two generative convolutional neural networks (CNNs) (F and G) that encourage the cycle-consistency (F(G(X))≈X and G(F(Y))≈Y). The goal was achieved by a regularizing term defined as the sum of absolute difference, mathematically based on L1-norm minimization, between the resulting images from two generative networks and original images. Our attempt to enhance translational accuracy is to incorporate theoretically ideal operator for sparse signal processing, L0-norm that minimizes the number of non-zero elements in image difference. To prevent divergence, this work counted the number of elements beyond a certain threshold in image difference during training. The training dataset consists of 2-fold augmented 2D MR and CT images from 50 patients trained by 200 epochs and 10-4 learning rate, and test dataset has images from 10 patients.

Results: The training from the proposed regularization with L0-norm yielded variational results depending on the threshold values, which provided the smallest error at a value of 0.75. With such a condition, the network with L0-norm constrained loss lowered mean-absolute error from 101.06 to 92.75 and root-mean-squared error from 0.873 to 0.794, relative to that with L1-norm regularized loss function.

Conclusion: Incorporating the L0-norm regularization into adversarial term of cGAN was demonstrated to improve the accuracy of MR-to-CT synthesis.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Research Foundation of Korea grant funded by the Korea government(2017R1D1A1B0403367).

Keywords

MR, CT

Taxonomy

IM/TH- MRI in Radiation Therapy: MRI for treatment planning

Contact Email