Multi-scale nested UNet with transformer for colorectal polyp segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Polyp detection and localization are critical for colonoscopy.
• Convolutional Neural Networks (CNNs) like U-Net excel at segmentation but struggle with long-range dependencies.
• Limited receptive fields hinder precise polyp segmentation.

Purpose of the Study:

• To develop a novel architecture for polyp segmentation.
• To integrate local feature extraction with long-range dependency modeling.
• To improve the accuracy and generalization of polyp segmentation models.

Main Methods:

• A hybrid CNN-Transformer architecture was developed, integrating a multi-scale nested U-Net with transformer layers.
• Transformer layers were embedded between the encoder and decoder to capture global context and long-range semantic information.
• A Multi-Scale Feature Fusion (MSFF) unit was introduced to handle variations in polyp sizes by fusing multi-resolution features.

Main Results:

• The proposed model achieved high accuracy on multiple datasets (Kvasir-SEG: 0.942 mean Dice score, CVC-ClinicDB: 0.950 mean Dice score).
• Cross-dataset validation demonstrated superior generalization capability compared to state-of-the-art methods.
• Ablation studies confirmed the effectiveness of individual model components.

Conclusions:

• The novel hybrid model significantly improves polyp segmentation accuracy over existing methods.
• The model's ability to segment polyps of varying sizes indicates strong potential for clinical applications in colonoscopy.