ClearNet: auto-encoder based denoising model for endoscopy images

Abstract

Gastrointestinal (GI) endoscopy images play a crucial role in the detection and diagnosis of diseases within the digestive tract. However, the development of effective computer vision models for automated analysis and denoising of endoscopy images faces challenges arising from the diverse nature of abnormalities and the presence of image artefacts. In this work, the utilization of an encoder-decoder network for denoising GI endoscopy images using the HyperKvasir dataset has been analyzed. This approach involves training a custom encoder-decoder model on this extensive multiclass endoscopy image dataset and assessing its performance across 23 prevalent classes of digestive tract issues. Here experiments showcase the model’s ability to learn robust visual representations from endoscopic data, enabling accurate disease prediction. The achieved promising results highlight the potential of encoder-decoder architectures as a foundational framework for computer-aided endoscopy analysis with a specific focus on denoising applications. Our model manages to increase the peak signal-tonoise ratio (PSNR) of original-noisy pair from 19.118954 to 69.892631 for original-reconstructed pair showcasing almost perfect reconstruction.