A residual recurrent convolutional neural network for image superresolution with whole slide images

Abstract

Presented is a deep learning based computational approach to solve the problem of enhancing the resolution of images gained from commonly available low magnification scanners, also known as the image super-resolution (SR) problem. The given class of scanner produces microscopic images relatively fast and has the advantage of storage efficiency. However, those scanners generate comparatively low quality images compared to images from complex and sophisticated higher cost, lower availability scanners and do not have the necessary resolution for diagnostic or clinical research. Therefore, low resolutions scanners are not in demand for these purposes. The motivation of this research is to determine whether an image with low resolution could be enhanced by applying a deep learning framework, resulting in an image that could serve the same diagnostic purposes as a high resolution image from expensive scanners or microscopes. Here, proposed are various models built onto a Recurrent Convolutional Neural Network (RCNN), with primary emphasis placed on a Residual Recurrent Convolutional Neural Network (RRCNN). The RRCNN created is a supervised machine learning method to process images that has properties of convolutional, residual, and recurrent neural networks. These models are specifically trained to take a low-resolution microscopic image from one of two tissue micro-arrays (TMAs) and transform it into a high-resolution image. Validation of these resolution improvements with computational analysis is done to show quantitative results for reconstructed images. The experiments completed demonstrate that some of the models produce images which are of similar quality to images from high resolution scanners, opening up new possibilities for research or clinical use.