Two-stage deep-learning-based colonoscopy polyp detection incorporating fisheye and reflection correction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colonoscopy is crucial for diagnosing colorectal diseases.
• Computer-aided systems aid polyp detection in colonoscopy images.
• Image distortions (fisheye lens, reflections) hinder polyp detection.

Purpose of the Study:

• To propose a two-stage deep learning model for correcting colonoscopy image distortions and reflections.
• To improve the accuracy of polyp detection in colonoscopy images.

Main Methods:

• A two-stage deep learning model was developed.
• Stage 1: Convolutional Neural Network (CNN) for fisheye distortion correction and polyp detection.
• Stage 2: Generative Adversarial Networks (GANs) for reflection correction, followed by CNN for polyp detection.

Main Results:

• The model achieved higher accuracy with corrected images (96.8%) compared to uncorrected images (90.8%).
• Polyp detection accuracy on the Kvasir-SEG dataset reached 96%.
• Area under the ROC curve was 0.94.

Conclusions:

• The proposed model effectively corrects distortions and reflections in colonoscopy images.
• This facilitates clinical diagnosis of colorectal polyps and enhances colonoscopy quality.