APRNet: Alternative Prediction Refinement Network for Polyp Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Colorectal cancer is a leading cause of cancer death, necessitating effective screening methods.
• Automatic polyp segmentation in colonoscopy is crucial for early detection but challenging due to polyp variability and ambiguous boundaries.

Purpose of the Study:

• To develop a novel deep learning network for accurate automatic polyp segmentation in colonoscopy.
• To improve the performance of polyp detection in screening systems.

Main Methods:

• Proposed the Alternative Prediction Refinement Network (APRNet), based on the UNet architecture.
• Utilized an encoder-decoder structure with alternative feature leveraging.
• Introduced Prediction Residual Refinement (PRR) modules for progressive segmentation refinement.

Main Results:

• Achieved state-of-the-art performance on benchmark datasets.
• Obtained a Dice score of 91.33% and 97.31% accuracy on the Kvasir-SEG dataset.
• Achieved a Dice score of 86.33% and 97.12% accuracy on the EndoScene dataset.

Conclusions:

• The APRNet demonstrates superior accuracy in polyp segmentation compared to existing methods.
• This algorithm has the potential to serve as an assistive tool in colonoscopy procedures for polyp identification.