U-Net for SPECT Image Denoising

Reymann M, Würfl T, Stimpel B, Ritt P, Cachovan M, Vija AH, Maier A (2019)

Publication Language: English

Publication Type: Conference contribution, Abstract of a poster

Publication year: 2019

Conference Proceedings Title: 2019 IEEE Nuclear Science Symposium and Medical Imaging Conference Proceedings (NSS/MIC)

Event location: Manchester GB

DOI: 10.1109/nss/mic42101.2019.9059879


Single Photon Emitted Computed Tomography (SPECT) is characterized by low photon counts and high degrees of image noise. In this work we investigate deep image postprocessing methods for improving image quality in SPECT projections and propose a new architecture. Specifically, the residual U-Net proposed by Heinrich et al. for low-dose Computed Tomography (CT) and the Convolutional Denoising Autoencoder (CNN DAE) by Gondara et al. were tested for their performance for denoising SPECT projections. Reference images without noise and scatter were obtained from SPECT Monte Carlo simulations of 24 different XCAT phantoms of brains, lungs, livers and skeletons using the SIMIND software. These clean images were then used as training target for the neural networks. Additionally, we propose a U-Net architecture that is more suited for SPECT image denoising, consisting of 4 layers, each with 2 blocks of convolution, batch normalization and ReLU. Our results showed that the proposed U-Net outperforms the residual U-Net and the CNN DAE with a PSNR of 38.25 dB, compared to 27.95 dB and 28.25 dB, respectively. It was shown that the characteristics of the SPECT data posed a challenge to the neural networks, as there are only few distinct photon count values at the detector available and these are superimposed with a high degree of image noise, compared to low-dose CT. Our proposed network was able to decrease the noise within the phantom and additionally decreased the background noise in the image. The image quality was improved by 14.15 dB compared to the noisy input image, demonstrating the suitability of our network. We believe that with increased size and diversity of our dataset we can further boost the performance.

Authors with CRIS profile

Related research project(s)

How to cite


Reymann, M., Würfl, T., Stimpel, B., Ritt, P., Cachovan, M., Vija, A.H., & Maier, A. (2019). U-Net for SPECT Image Denoising. Poster presentation at 2019 IEEE Nuclear Science Symposium (NSS) and Medical Imaging Conference (MIC), Manchester, GB.


Reymann, Maximilian, et al. "U-Net for SPECT Image Denoising." Presented at 2019 IEEE Nuclear Science Symposium (NSS) and Medical Imaging Conference (MIC), Manchester Ed. IEEE, 2019.

BibTeX: Download