Evaluation of super resolution technology for digestive endoscopic images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Digestive endoscopy generates numerous images crucial for diagnosis.
• Low-resolution images can limit diagnostic accuracy and hinder automated analysis.
• Super resolution (SR) technology offers potential for image quality enhancement.

Purpose of the Study:

• To evaluate the effectiveness of SR technology in improving the quality of digestive endoscopic images.
• To assess the impact of SR on both human interpretation and computer vision algorithms.

Main Methods:

• Retrospective study using two SR models: SwinIR and ESRGAN.
• Training on 12,212 high-resolution images and evaluation on 2,566 low-resolution images (HyperKvasir dataset).
• Endoscopist assessment via a 5-point Likert scale and evaluation on two endoscopic image classification tasks.

Main Results:

• SwinIR achieved high performance metrics (PSNR: 32.60, SSIM: 0.90, VIF: 0.47).
• 90% of endoscopists found SR preprocessing improved image readability.
• SR enhanced convolutional neural network performance in classifying Barrett's esophagus and Mayo Endoscopy Score (F1-score improvement: 0.04).

Conclusions:

• SR technology can effectively generate high-resolution digestive endoscopic images from low-resolution inputs.
• SR preprocessing improves clinical readability and enhances the performance of computer vision models for endoscopic image analysis.