Shedding Light on Diagnostic Precision: GANs for Low Light Endoscopy Image Enhancements

Abstract Understanding and identifying polyps in their early stages is crucial, and computerized diagnostic systems can significantly boost the effectiveness of screening through a large volume of endoscopy images. To achieve this, it's essential to enhance the quality of these images. In our research, we employ a well-established model called CycleGAN, introduced by Zhu et al. , which utilizes Generative Adversarial Networks (GANs). The CycleGAN model is trained on a diverse natural image database called RAISE and then tested on low illuminance endoscopy images from the Kvasir dataset, specifically created for polyp segmentation. Since there's no established ground truth for low illuminance endoscopy images, we use the non-reference Naturalness Image Quality Evaluator (NIQE) parameter to evaluate the quality of the generated high illuminance images in comparison to their low illuminance counterparts. Our main focus is to showcase how this image enhancement technique can be a valuable preprocessing step to avoid missing polyp detection during subsequent image analysis. Both visual observations and quantitative results suggest that our proposed method has the potential to effectively enhance endoscopic images analysis, contributing to an overall improvement in accuracy when it comes to polyp detection.