GFANet: Global Feature Attention Network for Polyp Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal polyps are crucial indicators of colorectal cancer, necessitating precise segmentation for diagnosis and treatment.
• Current deep learning models struggle with polyp segmentation due to variations in size, shape, color, and indistinct boundaries.

Purpose of the Study:

• To develop an advanced deep learning network, GFANet, for improved automatic segmentation of colorectal polyps.
• To address limitations in existing methods, including poor handling of geometric features, low sensitivity to small polyps, and insufficient multi-scale information fusion.

Main Methods:

• Introduction of GFANet, incorporating a global feature direction encoder (GFDE), feature attention module (FAM), and multi-scale information aggregation (MIA).
• GFDE enhances localization of polyps with challenging visual characteristics.
• FAM refines feature representation and suppresses background noise.
• MIA aggregates multi-scale and semantic features for comprehensive segmentation.

Main Results:

• GFANet demonstrated superior performance compared to ten state-of-the-art methods across five datasets.
• Achieved 90.2% mDice and 83.5% mIoU on the CVC-300 dataset, outperforming existing approaches.
• Significantly outperformed PraNet by 17.2% in mDice on the ETIS-LaribPolypDB dataset, showing strong generalization.

Conclusions:

• GFANet effectively addresses key challenges in colorectal polyp segmentation, offering enhanced accuracy and sensitivity.
• The network's innovative modules contribute to superior performance in identifying polyps of various sizes and complexities.
• GFANet shows significant potential for clinical application in colorectal cancer diagnosis and management.