A Siamese based Generative Adversarial Network for Separated Medical Image Enhancement

Due to the need for expensive and complex equipment, highly skilled personnel, and often postponed operations, obtaining high-resolution pictures in medical imaging is a difficult and expensive task. Super resolution methods based on deep learning may help us get rich information out of low-resolution images taken with current equipment. One of the major areas of focus in medical research recently has been the improvement of unpaired medical images. Medical picture improvement using deep learning has been very successful, however these systems still struggle with issues like small data sets and poor training sets quality. This article proposes a Siamese GAN with ResNet34-based dual-input mechanism image enhancement method that makes use of the target highlight structure (texture enhancement) while background balance structure (consistent background contrast) from uncoupled low-quality as well as high-quality medical images. In addition, the suggested approach incorporates the generative adversarial network methodology to accomplish structure-preserving augmentation via concurrent adversarial learning iterations. By comparing the suggested Siamese GAN plus ResNet34 to other state-of-the-art methods, experiments show how well it performs in unpaired picture improvement. Experiments are conducted using three datasets: (1) pictures from retinal fundoscopy (DRIVE and STARE), (2) brain MRI (BraTS), and (3) dermoscopy images from skin cancer (ISIC) patients.