Deep Learning Based Focus lnterpolation for Whole Slide Images

Abstract: Whole slide imaging is a crucial component of digital pathology, which emulates conventional microscopy by scanning the entire microscope slide. It results in a large digital image that can be examined by a cytopathologist or used in further computer-assisted image analysis. However, due to the optical limitations of the imaging system, only a thin slice of the three-dimensional sample is in focus in the captured image. Consequently, the cytopathologist looses the possibility to smoothly adjust the focus and get a feel for the shape of the sample at different depths. A common solution to this problem is to capture multiple images of the same slide at different heights, i.e., a so called z-stack (or focus stack). However, it still allows to inspect only a limited number of discretely positioned slices of the sample, and it dramatically increases the imaging time and captured data volume. This project proposes a focus interpolation method that allows to take only a few manually acquired images and interpolate the rest on demand. The proposed method relies on a convolutional neural network based on the U-Net architecture, which takes two different focus images as inputs and outputs an image corresponding to the intermediate focus level. The proposed method shows better results than no interpolation or linear interpolation and can be generalized to different datasets.