Improved dual-aggregation polyp segmentation network combining a pyramid vision transformer with a fully...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate polyp segmentation in colonoscopy is crucial for early colorectal cancer diagnosis and treatment.
• Challenges include variations in polyp size, morphology, and indistinct boundaries with surrounding mucosa.

Purpose of the Study:

• To propose Dua-PSNet, an improved dual-aggregation network for precise automatic polyp segmentation.
• To enhance the discrimination of polyp features and improve the accuracy of full-size segmentation maps.

Main Methods:

• A parallel dual-branch network combining a transformer (PVTv2-B3) and a fully convolutional network (FCN).
• Incorporated a multi-stage feature aggregation decoder (MFAD) with adaptive feature aggregation (AFA) and ResidualBlock modules.
• Utilized a selective global-to-local fusion head (SGLFH) for integrating detailed boundary cues with global semantic features.

Main Results:

• Dua-PSNet demonstrated powerful learning and generalization capabilities on five challenging polyp segmentation benchmarks.
• Achieved state-of-the-art segmentation performance, outperforming existing cutting-edge methods.
• Successfully addressed challenges related to polyp scale variations and indistinct boundaries.

Conclusions:

• Dua-PSNet offers a promising solution for practical polyp segmentation tasks with diverse data variations.
• The dual-aggregation approach effectively enhances feature discrimination and segmentation accuracy.
• The method holds significant potential for improving colorectal cancer screening and management.