PGMNet: a polyp segmentation network based on bit-plane slicing and multi-scale adaptive fusion.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate polyp detection and segmentation in colonoscopy are crucial for early colorectal cancer prevention and treatment.
• Variations in polyp size, shape, and blurred boundaries present significant challenges for current deep learning (DL) segmentation methods, leading to unstable and unsatisfactory results.

Purpose of the Study:

• To develop an accurate and efficient deep learning network, PGMNet, for improved polyp segmentation in colonoscopy images.
• To address the limitations of existing DL methods in handling polyp variations and boundary ambiguity.

Main Methods:

• PGMNet utilizes a PVTv2 encoder for capturing fine-grained details and global semantic information.
• The network incorporates a Global-Local Interactive Relation Module (GLIRM) for multi-scale information fusion and noise suppression.
• A Multi-stage Feature Aggregation Module (MFAM) with a gating mechanism efficiently aggregates features to enhance prediction quality.

Main Results:

• PGMNet demonstrated promising performance across five public polyp datasets, showing strong segmentation accuracy and generalization ability.
• On the challenging ETIS dataset, PGMNet achieved a mean Dice coefficient (mDice) of 82.33% and a mean Intersection over Union (mIoU) of 74.29%.

Conclusions:

• PGMNet offers a superior solution for polyp segmentation compared to existing methods.
• The proposed network shows significant potential for improving early colorectal cancer diagnosis through enhanced colonoscopic polyp detection and segmentation.