Selective denoising technique in optical diffraction tomography images of red blood cell

Abstract

We present a selective denoising technique in optical diffraction tomography (ODT) red blood cell images. In optical diffraction tomography images of red blood cells (RBCs), similar artifacts are found at the same locations in different cell images. These artifacts are independent from the data and it is considered as a fixed noise which should be removed from the original image. Its size and pattern are different from Gaussian noise which is generally handled in most of the previous denoising methods. Thus, noise reduction algorithm which encounters the selected region which contains the fixed noise artifact and does not harm the structure of the cell is needed. We propose a denoising strategy that automatically detects the fixed noise artifact and denoises in the selected areas. At first, we manually annotate boxes which contain microscopy-specific artifact in the cell images. By making a training dataset as pairs of boxes and images, we train a convolutional neural network (CNN) model which detects the artifacts in given ODT red blood cell images. After detecting artifacts, we apply proposed method to detected boxes which is median filter with window size quarter of the box containing the artifact. Finally, we have an automatic framework which gives an image with denoised selected region. This technique can be also applied to the ODT 3D volume data as the application. The main contribution of this work is that we made a framework that automatically detects the fixed noise and denoise the selected region with the proposed method in ODT images.