PSTNet: Enhanced Polyp Segmentation With Multi-Scale Alignment and Frequency Domain Integration.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of colorectal polyps is vital for diagnosing and managing colorectal cancer (CRC).
• Current deep learning methods struggle with limited RGB data and feature misalignment in multi-scale analysis.
• Existing approaches face challenges in precisely identifying polyps due to these limitations.

Purpose of the Study:

• To develop a novel deep learning model for enhanced polyp segmentation in colonoscopy images.
• To address limitations of existing methods by integrating frequency domain information with RGB data.
• To improve the accuracy and efficiency of computer-assisted polyp detection for CRC management.

Main Methods:

• Proposed the Polyp Segmentation Network with Shunted Transformer (PSTNet).
• Integrated RGB and frequency domain cues using three key modules: FCAM, FSAM, and CPM.
• FCAM extracts frequency cues, FSAM aligns semantic information, and CPM synergizes frequency and semantic data.

Main Results:

• PSTNet demonstrated significant improvements in polyp segmentation accuracy across various metrics.
• The model consistently outperformed existing state-of-the-art methods on challenging datasets.
• Integration of frequency domain cues led to superior performance compared to RGB-only methods.

Conclusions:

• PSTNet effectively enhances polyp segmentation by leveraging both RGB and frequency domain information.
• The novel architectural design of PSTNet advances computer-assisted polyp segmentation.
• This approach facilitates more accurate diagnosis and management of colorectal cancer.