Deep Image Denoising in SPECT

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

Publication Language: English

Publication Type: Conference contribution, Abstract of lecture

Publication year: 2019

Publisher: Springer Berlin Heidelberg

City/Town: Berlin, Heidelberg

Pages Range: 219-219

Conference Proceedings Title: European Journal of Nuclear Medicine and Molecular Imaging (2019) 46 (Suppl 1): S1–S952, 10.1007/s00259-019-04486-2

Event location: Barcelona

URI: https://link.springer.com/article/10.1007/s00259-019-04486-2

DOI: 10.1007/s00259-019-04486-2

Abstract

Aim/Introduction: Single Photon Emitted Computed Tomography (SPECT) images are characterized

by a high degree of image noise, which is due to the low photon count at the detector. Recently,

several deep learning-based approaches for image denoising have been proposed for low-dose

Computed Tomography (CT). We investigate their suitability for SPECT denoising.

Materials and Methods: In order to have a groundtruth for training the networks, a Monte Carlo

Simulation was set up with the SIMIND software, where in total 24 SPECT acquisitions of brains,

lungs, livers and skeleton were simulated with the XCAT phantom. Pairs of noisy and clean SPECT

images were then used for training the neural networks. The networks tested were the U-Net as

proposed by Heinrich et al. [1], the convolutional denoising autoencoder (CNN DAE) proposed by

Gondara et al. [2] and a custom U-Net that had a depth of 4 layers with 2 convolutions, Batch

Normalization and ReLU in each layer. Training of the networks was done with the hyperparameters

as described in the original publications.

Results: The best denoising performance of the neural networks was achieved by either the CNN DAE

when measured with the Structural Similarity (SSIM) index of 0.933 or by the custom U-Net with a

Peak Signal to Noise Ratio (PSNR) of 30.03 dB. This corresponded to a signal quality improvement of

0.09 compared to the noisy input when measured with the SSIM, or 5.9 dB when measured with the

PSNR. The denoised SPECT images showed improved visual appearance, however, gain in image

quality was not as distinct as on low-dose CT data. We assume this was due to the image

characteristics of SPECT images, which are characterized by few distinct image values and a more

heterogeneous pixel intensity distribution compared to CT data.

Conclusion: Image characteristics in SPECT pose a special challenge to deep learning-based image

denoising methods, as the signal is sparse and contains a high degree of noise. We believe this, in

conjunction with the limited number of samples explains the weaker benefit in performance in SPECT

when compared to low-dose CT data. We assume that with a more diverse training set and

specialized loss function suitable for SPECT data , we can further increase performance.

Authors with CRIS profile

Related research project(s)

How to cite

APA:

Reymann, M., Würfl, T., Ritt, P., Cachovan, M., Stimpel, B., & Maier, A. (2019, September). Deep Image Denoising in SPECT. Paper presentation at Annual Congress of the European Association of Nuclear Medicine, Barcelona, ES.

MLA:

Reymann, Maximilian, et al. "Deep Image Denoising in SPECT." Presented at Annual Congress of the European Association of Nuclear Medicine, Barcelona Ed. Springer-Verlag GmbH Germany, part of Springer Nature 2019, Berlin, Heidelberg: Springer Berlin Heidelberg, 2019.

BibTeX: Download