Multi-scale aggregation network for colonoscopic polyp segmentation via frequency domain decoupling.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of colorectal polyps is crucial for early colorectal cancer detection and treatment.
• Challenges in colonoscopy image analysis include diverse polyp morphologies and variable illumination, hindering precise segmentation and edge extraction.

Purpose of the Study:

• To develop an advanced deep learning network, FDANet, for robust and accurate automated segmentation and edge extraction of colorectal polyps.
• To improve the performance of polyp segmentation by leveraging frequency domain analysis and multi-scale feature aggregation.

Main Methods:

• Proposed FDANet utilizes wavelet transform to decompose spatial features into frequency sub-bands, separating low-frequency and high-frequency components.
• Incorporated Low-Frequency Attention Enhancement Module (LAEM) to suppress noise and enhance foreground features.
• Integrated High-Frequency Multi-Scale Aggregation Module (HMAM) with directional convolutions and an edge loss function for fine-grained edge detection and multi-scale polyp representation.

Main Results:

• FDANet demonstrated superior segmentation performance on the CVC-ClinicDB and Kvasir-SEG datasets.
• The proposed method outperformed existing advanced segmentation techniques in accuracy and boundary localization.
• Frequency domain processing effectively handled variations in polyp morphology and illumination.

Conclusions:

• FDANet offers a promising approach for automated colorectal polyp segmentation, enhancing early cancer screening capabilities.
• The frequency domain decoupling and multi-scale feature aggregation strategy effectively addresses segmentation challenges in colonoscopy images.
• This method contributes to more accurate polyp detection and boundary localization, aiding clinical decision-making.