Polyp image segmentation based on parallel dilated convolution and dual attention mechanisms.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Colorectal cancer often originates from benign polyps that can become malignant if untreated.
• Early detection of polyps is crucial for improving patient survival rates.
• Existing methods struggle with localizing small polyps and segmenting indistinct polyp boundaries.

Purpose of the Study:

• To develop an advanced polyp image segmentation model, MSFNet (Multi-Scale Feature Unet), for enhanced colorectal cancer diagnosis.
• To address challenges in small polyp localization and boundary segmentation in medical images.

Main Methods:

• Proposed MSFNet model utilizing multi-scale contextual feature extraction and channel attention modules.
• Incorporated channel-by-channel convolution and dilated convolution for efficient feature exploration with a low parameter count.
• Introduced the Multi-Scale Big Dilated Conv kernel (MBDCK) module to enhance multi-scale context and receptive field.
• Implemented skip connections for multi-scale feature fusion to improve global feature extraction.

Main Results:

• MSFNet achieved a Dice score of 0.892 and an mIoU of 0.926 on the CVC-ClinicDB dataset.
• Demonstrated superior performance compared to U-Net, PraNet, and SwinUNet, with approximately 10% higher mIoU than U-Net.
• Exhibited excellent segmentation accuracy and generalization ability with fewer parameters than existing advanced models.

Conclusions:

• MSFNet effectively overcomes limitations in small polyp localization and boundary segmentation.
• The proposed model offers a promising solution for accurate and efficient colorectal polyp segmentation in medical imaging.
• MSFNet provides a strong balance between model complexity and segmentation performance, beneficial for clinical applications.