Stereo-Correlation and Noise-Distribution Aware ResVoxGAN for Dense Slices Reconstruction and Noise Reduction in Thick Low-Dose CT
The low-dose computed tomography (CT) scan with thick slice thickness (3 mm) dramatically improves the imaging efficiency and reduces the radiation risk in clinical. However, the low-dose CT acquisition inherently compromises the signal-to-noise ratio, and the sparse sampled thick slices poorly reproduce the coronal/sagittal anatomy. We propose a Residual Voxel Generative Adversarial Nets (ResVoxGAN), the first powerful work to densely reconstruct slices into the thin thickness (1 mm), and simultaneously denoise the CT image into the more readable pattern, directly from the widely accessible thick low-dose CT. The framework is achieved in a voxel-wise conditional GAN constituted by the following: (1) a generator is composed of consecutive 3D multi-scale residual blocks that richly extracts multi-scale stereo feature for fine-granted and latent spatial structure mining from the noisy volume, and a followed Subpixel Convnet further interpretively reconstructs dense slices from the features for high-resolution and denoising volume; (2) a stereo-correlation constraint elegantly penalizes gradient deviation in voxel adjacent region (i.e., 3D 26-neighborhoods) to guide structural detail, together with a image-expression constraint on perceptual feature representations transformed from a pretrained deep convolution autoencoder to keep scene content; and (3) a pair of coupled discriminators advantageously fuse the prior-knowledge from the thick low-dose CT with the generated image and residual noise via self-learning to drive the generation towards into both realistic anatomic structure distribution and valid noise-reduction distribution. The experiment validated on Mayo dataset shows that the ResVoxGAN successfully reconstruct the low-dose CT of 3 mm thickness into 1 mm, and meanwhile keeps the with peak signal to noise ratio of 40.80 for noise reduction, and structural similarity index of 0.918 for dense slices reconstruction. These advantages reveal our method a great potential in clinical CT imaging.
This study was supported by the Postgraduate Research & Practice Innovation Program of Jiangsu Province (No. KYCX17\(\_\)0104), the China Scholarship Council (No. 201706090248), the State’s Key Project of Research and Development Plan (No. 2017YFA0104302, No. 2017YFC0109202 and No. 2017YFC0107900), the National Natural Science Foundation (No. 81530060 and No. 61871117).
- 5.Bae, W., Lee, S., Park, G., Park, H., Jung, K.H.: Residual CNN-based image super-resolution for CT slice thickness reduction using paired CT scans: preliminary validation study. In: Medical Imaging with Deep Learning (2018)Google Scholar
- 8.Liu, J., et al.: Discriminative feature representation to improve projection data inconsistency for low dose CT imaging. IEEE TMI 36(12), 2499–2509 (2017)Google Scholar
- 9.Chen, H., et al.: Low-dose CT with a residual encoder-decoder convolutional neural network. IEEE TMI 36(12), 2524–2535 (2017)Google Scholar
- 11.Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of Wasserstein GANs. In: NIPS, pp. 5767–5777 (2017)Google Scholar
- 12.Shi, W., et al.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR, pp. 1874–1883 (2016)Google Scholar
- 15.Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR, pp. 1125–1134 (2017)Google Scholar