• Ignas Kupcikevicius
Researchers put a lot of efforts in order to reduce the noise of low-dose CT scans, which appear due to lowered X-ray radiation. Image post-processing methods for this purpose vary from traditional algorithms, like non-local means (NLM), k-SVD, block-matching 3D (BM3D) to more recent attempts with convolutional neural networks (CNN's). However, none of those studies tried to employ CNN in a way that it would simulate the performance of NLM algorithm. Moreover, there was no clear way of how to integrate the proposed model into already existing medical imaging platform, like Siemens Syngo.via Frontier. Therefore, the aim of this study is to develop the CNN which simulates NLM performance on the 3D low-dose CT data and to integrate this model into Syngo.via Frontier platform.
The model used the Lung Image Database Consortium (LIDC) dataset from 1018 patients. It was necessary to put the constraints on the dataset since it was high in resolution and large in size. The region of interest was selected and cut into the fixed 3D patches in order to make the data more manageable. Additionally, additive noise was applied on 6,000 patches which were denoised with NLM to make the ground truth images for the CNN. The encoder-decoder CNN with shortcut connections were used to train the final model, and tuned for the mean-square-error objective function. The integration schema was proposed which involves model development on Python, implementation of the Tensorflow and Keras into research prototyping platform MeVisLab, and Siemens MR Starter Kit to simulate Syngo.via Frontier environment.
The results showed that NLM remained superior over proposed CNN, comparing quantitatively and qualitatively. PSNR of NLM on the average was 39.56 dB, in contrary to 28.97 dB of CNN. Additionally, SSIM was calculated with values of 0.91 and 0.81 for NLM and CNN, respectively. In terms of denoising time, CNN showed a significant lead of 0.07s against 19.36s for NLM, to denoise one patch. The developed model was successfully integrated into simulated Syngo.via Frontier environment.
Even though CNN did not achieve the denoised image quality of NLM, it still showed denoising capabilities as a future alternative. Moreover, CNN showed much faster post-processing time on the trained network against NLM. However, the current CNN model needs further optimization involving more complex CNN architectures. Using proposed integration schema neural network model can be implemented and run on Siemens Syngo.via Frontier platform, that the platform can take advantage of the newest algorithms.
Publication date20 Dec 2018
Number of pages46
External collaboratorSiemens Healthineers AS
Scientific Research Scientist Anders B. Rodell anders.rodell@siemens-healthineers.com
Place of Internship
KeywordsLDCT, CNN, NLM, Syngo, Siemens
ID: 291954790