Low dose CT denoising with residual convolutional networks and deep learned perceptual loss

Congress:

ECR 2020

Poster Number:

C-03463

Type:

Scientific Exhibit

Keywords:

Not applicable, Observational, Artifacts, Computer Applications-General, CT, eHealth, Computer applications, Artificial Intelligence, Artificial Intelligence and Machine Learning

Authors:

M. Patwari¹, R. Gutjahr², R. Raupach³, A. Maier¹; ¹Erlangen/DE, ²Nürnberg, Deutschland/DE, ³Forchheim/DE

DOI:

10.26044/ecr2020/C-03463

DOI-Link:

https://dx.doi.org/10.26044/ecr2020/C-03463

Fig. 1: This image shows the structure of the neural network used for denoising. The...

Fig. 3: This image shows the extraction of perceptual information using the quality...

Fig. 2: This image shows the reconstruction quality network used to extract features....

Fig. 4: This image shows noisy images denoised using 2D and 3D networks trained with...

Purpose

Due to the use of CT as a crucial medical tool for diagnostic and interventional purposes, the number of CT scans has been increasing. However, CT scans emit radiation which is harmful to the patient [1]. In accordance with the ALARA principle, we aim to reduce clinical CT dose to improve patient safety. However, reducing the radiation dose increases quantum noise in the reconstructed images [2]. Quantum noise in clinical CT is usually removed using nonlinear filters such as the bilateral filter [3], the median...

Methods and materials

Data:Ten CT body scans were reconstructed using weighted filtered backprojection [13]. Each volume was reconstructed with the standard clinical dose and with doses of 50%, 25%, 10% and 5% of the standard clinical dose to simulate reduced dose acquisitions. The images reconstructed with 100% of the standard dose were treated as ground truth images. Each image had a resolution of 256 x 256. The testing dataset consisted of three body CT scans which were not present in the training dataset. The test dataset was reconstructed...

Results

The use of a perceptual loss function improved both feature preservation and denoising performance in 2D denoising networks (SSIM = 0.7067, PSNR = 44.5189) compared to mean squared error denoising (SSIM = 0.7065, PSNR = 44.4888). Perceptual loss functionsimproved both feature preservation and denoising performancein 3D denoising networks (SSIM = 0.6944, PSNR = 45.9397) compared to mean squared error denoising (SSIM = 0.6871, PSNR = 45.6718). Additionally, 3D networks were better at noise removal compared to 2D networks. However, the 2D networks improved feature preservation...

Conclusion

The purpose of this study was to evaluate the performance of denoising networks and the impact of including perceptual loss. We also studied whether the use of additional slices can help to improve the denoising performance of the network. 3D networks significantly outperform 2D networks in noise removal, however, 2D networks make up for it by better feature preservation after denoising. This indicates that using neighbouring slices for denoising helps in noise removal, however, it does not preserve features. Using perceptual loss improves feature preservation...

Personal information and conflict of interest

R. Gutjahr; Nuremberg, DEUTSCHLAND/DE - Employee at Siemens Healthcare GmbH M. Patwari; Erlangen/DE - Employee at Siemens Healthcare GmbH R. Raupach; Forchheim/DE - Employee at Siemens Healthcare GmbH A. Maier; Erlangen/DE - Research/Grant Support at Siemens Healthcare GmbH

References

1. Monga M. Computed tomography - An increasing source of radiation exposure: Editorial comment. Int Braz J Urol. 2007;33(6):855. 2. Oppelt A. Noise in Computed Tomography. In: Aktiengesselschaft S, editor. Imaging Systems for Medical Diagnostics [Internet]. 2nd ed. Publicis Corporate Publishing; 2005. p. 996. 3. Manduca A, Yu L, Trzasko JD, Khaylova N, Kofler JM, McCollough CM, et al. Projection space denoising with bilateral filtering and CT noise modeling for dose reduction in CT. Med Phys. 2009;36(11):4911–9. 4. Manhart M, Fahrig R, Hornegger J, Doerfler...