Super Resolution in Medical Imaging

Medical imaging is a technique and a procedure that involves imaging of the interior portion of a human body for clinical evaluation, medical intervention, and to show how certain organs or tissues are functioning (physiology). MRI scan images of any patient can degrade in quality if they are shared over the internet or in hardcopy. This paper intends to provide an application to obtain high-resolution medical images from lower resolution images. The healthcare sector requires services and data of the best possible quality and this application would help them in better diagnosis and treatment. This paper uses Deep Learning Models and Machine Learning techniques, particularly the Super Resolution Generative Adversarial Networks, to high-resolution Brain MRI images (128 x 128 dimensions) from low-resolution images (32 x 32 dimensions). A web application is developed using the Streamlit library that can also be accessed from a mobile (responsive User Interface) for the end-to-end implementation of the paper. The underlying architecture of the SRGAN model can be replicated to other use cases with suitable customization to achieve high accuracy in generation of super-resolution images. The future scope would include the implementation of a mobile application with real-time processing of generation of high-resolution images from lower resolution images.

I. Introduction

Medical field is one of the sectors where the data required to diagnose an issue is required to be as accurate as possible. It is also important that the data is readable to easily and correctly identify the presence of any abnormalities which in turn will assist the doctors to give correct treatment to their patients. Due to many factors, the readability of these data is not up to the mark and the data itself may undergo degradation over time that is in case of image data, the loss of clarity (resolution). This data degradation is applicable for both virtual and physical formats of data. With the help of deep learning technologies, and deep neural networks such as Super Resolution - Generative Adversarial Network, a low-resolution image can be converted into a high- resolution image. The generated high-resolution images can be helpful over the low-resolution images for better diagnosis of the medical images.