CTNet: Contrastive Transformer Network for Polyp Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Polyp segmentation in colonoscopy is crucial for colorectal cancer diagnosis.
• Existing methods struggle with polyp camouflage and size variations, lacking stable results.
• Challenges include limited distinguishing features and high-level semantic details.

Purpose of the Study:

• To introduce a novel polyp segmentation framework, the contrastive Transformer network (CTNet).
• To address limitations in current polyp segmentation techniques, particularly regarding camouflage and size variability.
• To improve the accuracy and stability of polyp segmentation in colonoscopy images.

Main Methods:

• Proposed CTNet framework with three components: contrastive Transformer backbone, self-multiscale interaction module (SMIM), and collection information module (CIM).
• Utilized contrastive Transformer for long-range dependence and structured feature maps to handle camouflaged polyps.
• Employed SMIM and CIM to integrate multiscale information and high-resolution semantic features for accurate segmentation of diverse polyp sizes.

Main Results:

• CTNet demonstrated significant performance gains over the PraNet method on multiple benchmark datasets (Kvasir-SEG, CVC-ClinicDB, Endoscene, ETIS-LaribPolypDB, CVC-ColonDB).
• Achieved percentage gains of 2.3%, 3.7%, 3.7%, 18.2%, and 10.1% on the respective datasets.
• Showcased advantages in camouflaged object detection and defect detection tasks.

Conclusions:

• CTNet offers excellent learning and generalization abilities for polyp segmentation.
• The framework effectively localizes camouflaged polyps and accurately segments polyps of varying sizes.
• CTNet represents a significant advancement in automated polyp detection and segmentation for improved colorectal cancer screening.