Despite its unique capabilities, diffusion-weighted imaging (DWI) in prostate is inherently limited by low signal-to-noise ratio (SNR). Currently, gains in SNR of high b-value images are achieved through increase in the number of excitations (NEX), at the cost of increase in total acquisition time. We demonstrate feasibility of improving prostate DWI image quality by leveraging denoising convolutional network. Using pairs of "noisy" NEX4 and "clean" NEX16 DWI images, reconstructed from raw data, CNN was trained to denoise prostate DWI images. Denoising of images significantly improved SNR and increased overall image quality, reviewed by two experienced genitourinary radiologists.
Network: In DnCNN8 (Fig. 1), noisy observation was defined as y = x + v, where x is a “clean” image, and v is noise. The goal was to train residual mapping R(y) ≈ v, so that denoised image is found as x = y – R(y). Loss function was averaged mean squared error between the “clean” and estimated denoised input images. Network depth was increased to 21, and patch size was increased to 60 × 60.
Data: Field-of-view Optimized and Constrained Undistorted Single shot (FOCUS)9 DWI is used in the standard prostate MRI protocol at our institution. Raw FOCUS data (N = 53: training-43, evaluation-5, testing-5) were retrospectively collected from six 3 T MRI scanners (GE Healthcare, Waukesha, WI) with IRB approval. Images were acquired with 3 diffusion directions, with b-valueS of 0 and 1000 s/mm2. For each direction, NEX for low b-value was 2 - 4, and for high b-value – 12 - 16.
Data pre-processing (Fig 2): 1) For high b-value images, raw data were reconstructed to generate DWI images for NEX of 4, 8 and 16; low b-value images were reconstructed using all NEX and were only used for apparent diffusion coefficient (ADC) map generation. 2) For training, pairs of “noisy", DWI NEX8, and "clean”, DWI NEX16, 116 × 246 images were created. All slices were used.
Training: Using 1262 image pairs (43 patients from 5 scanners), 367,360 patches were generated. "Noisy" and "clean" images had average apparent SNR of 16.5 and 20.7. Number of epochs was 40, with 2870 iterations. The Amazon Cloud with NVIDIA K80 GPU was used.
Analysis: Trained model was applied to 5 test sets to generate denoised images for NEX of 4,8 and 16. Using unprocessed and denoised DWI images, ADC maps were computed. Apparent SNR was measured in all images. In ADC maps, the number of pixels with ADC < 0 s/mm2 was recorded. Negative ADC ratio was computed by normalizing all pixel counts by the number of negative ADC pixels in unprocessed NEX16 image, original DWI image. Two-tail paired t-test was used to compare quantitative metrics between unprocessed and denoised images. Additionally, two experienced genitourinary radiologists independently reviewed 25 randomly-paired-up unprocessed and denoised image sets. Each recorded if one set of images had better overall diagnostic quality compared to the other or if there was no perceivable difference.
1. Ning, P.; Shi, D.; Sonn, G. A.; Vasanawala, S. S.; Loening, A. M.; Ghanouni, P.; Obara, P.; Shin, L. K.; Fan, R. E.; Hargreaves, B. A., The impact of computed high b-value images on the diagnostic accuracy of DWI for prostate cancer: A receiver operating characteristics analysis. Scientific reports 2018, 8 (1), 3409. 2. Stocker, D.; Manoliu, A.; Becker, A. S.; Barth, B. K.; Nanz, D.; Klarhöfer, M.; Donati, O. F., Image Quality and Geometric Distortion of Modern Diffusion-Weighted Imaging Sequences in Magnetic Resonance Imaging of the Prostate. Investigative radiology 2018, 53 (4), 200-206.
3. Veraart, J.; Novikov, D. S.; Christiaens, D.; Ades-Aron, B.; Sijbers, J.; Fieremans, E., Denoising of diffusion MRI using random matrix theory. NeuroImage 2016, 142, 394-406.
4. You, C.; Yang, Q.; Gjesteby, L.; Li, G.; Ju, S.; Zhang, Z.; Zhao, Z.; Zhang, Y.; Cong, W.; Wang, G., Structurally-Sensitive Multi-Scale Deep Neural Network for Low-Dose CT Denoising. IEEE Access 2018, 6, 41839-41855.
5. Kadimesetty, V. S.; Gutta, S.; Ganapathy, S.; Yalavarthy, P. K., Convolutional Neural Network based Robust Denoising of Low-Dose Computed Tomography Perfusion Maps. IEEE Transactions on Radiation and Plasma Medical Sciences 2018.
6. Xie, D.; Bai, L.; Wang, Z., Denoising Arterial Spin Labeling Cerebral Blood Flow Images Using Deep Learning. arXiv preprint arXiv:1801.09672 2018.
7. Gong, E.; Pauly, J. M.; Wintermark, M.; Zaharchuk, G., Deep learning enables reduced gadolinium dose for contrast‐enhanced brain MRI. Journal of Magnetic Resonance Imaging 2018.
8. Zhang, K.; Zuo, W.; Chen, Y.; Meng, D.; Zhang, L., Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing 2017, 26 (7), 3142-3155. 9. 9. Saritas, E. U.; Cunningham, C. H.; Lee, J. H.; Han, E. T.; Nishimura, D. G., DWI of the spinal cord with reduced FOV single‐shot EPI. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 2008, 60 (2), 468-473.