Xie, J., Xu, L., & Chen, E. (2012). Image denoising and inpainting with deep neural networks. In Advances in neural information processing systems (pp. 341-349).

Dong, W., Wang, P., Yin, W., Shi, G., Wu, F., & Lu, X. (2018). Denoising Prior Driven Deep Neural Network for Image Restoration. arXiv preprint arXiv:1801.06756.

Lucas, A., Iliadis, M., Molina, R., & Katsaggelos, A. K. (2018). Using Deep Neural Networks for Inverse Problems in Imaging. IEEE SIgnal ProcESSIng MagazInE, 1053(5888/18).

You can use the VGG, which has a better performance for image denoising. We have a paper about image deblurring and image denoising using a deeper network, which will be public soon. If you are interested, you can follow us work.

You can increase your data size by transforming the same data with some simple transformations.

You may use VGG, AlexNet or GoogleNet. I suggest you to follow the schema: Pre-trained CNN (such as GoogleNet or the other ones) + Transfer Learning (for a fine tuning on the application domain) + Data Augmentation (you can adopt simple geometric transform like rotation and scaling).

Thank you for your answers . I have 60 subjects with CT scans. I know I can do augmentation with geometric transformations. I just thought my data is too small for fine tuning.

You are correct that your data is too small other than to do very simple estimates. If there was some network that had been tuned on precisely this type of imagery, gathered for your current purpose, using essentially the same equipment, and its characteristics are well-documented, then you could use some pre-trained network.

Whether you are using a "neural network" or some other method that makes inferences from sampled data, you are doing statistics. Particularly in a mediacally-related application you need to be able to make reliable statements about the images. A pre-traiined network using data you do not have access to, or do not have time to thoroughly review, represents a nonparametric model with thousands of free parameters of unknown provenance. When the image is transformed, you cannot make any definitive statements about how your outcome is affected.

The first thing to do is look at the data carefully, and ascertain up front the characteristics you wish to enhance or remove (Exploratory Data Analysis, or EDA).

Your best bet then is probably to review both recent procedural literature in your field and the documentation for the R programming language to identify a well-documented method (e.g. kernel smoothing of a particular sort) that are well suited to the type of enhanceement you need..

Usually, the hard part in these kind of problems is "the vocabulary hunt:" you know what you want to do, a bunch of other people have worked on similar problems and written the code, etc, but they don't call it the same thing you do. So you probably need to spend a day or two chasing down leads on the internet until you have matched your jargon with that of the party who has a helpful statistical method to offer..

In the end you will have some analytic method that may or may not use adaptive computing. But you will know exactly why it does what it does, and have a body of knowledge to fall back on if mysterious problems arise, and you will have some confidence in conclusions you make that were based in part from the results of the software.

@ Nastaran Emaminejad , 60 is too small that is the catch with CNN .

So far we have used 1000's ...

Go for Generative Adversarial Network (GAN). Here is the example paper:

https://ieeexplore.ieee.org/abstract/document/8340157/

Xie, J., Xu, L., & Chen, E. (2012). Image denoising and inpainting with deep neural networks. In Advances in neural information processing systems (pp. 341-349).

Dong, W., Wang, P., Yin, W., Shi, G., Wu, F., & Lu, X. (2018). Denoising Prior Driven Deep Neural Network for Image Restoration. arXiv preprint arXiv:1801.06756.

Lucas, A., Iliadis, M., Molina, R., & Katsaggelos, A. K. (2018). Using Deep Neural Networks for Inverse Problems in Imaging. IEEE SIgnal ProcESSIng MagazInE, 1053(5888/18).

Helleo Nastaran Emaminejad

For this task transfer learning will be best you can pick already trained model related domain after you transfer the knowledge from already trained model to your targeted model .

Following link will direct you to some of deep-learning pretrained models available online. - https://github.com/wenbihan/reproducible-image-denoising-state-of-the-art