EMS-Net: Enhanced Multi-Scale Network for Polyp Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Polyp segmentation is vital for colorectal cancer diagnosis and treatment.
• Challenges include variations in polyp size, shape, and boundary clarity.
• Multi-scale contextual information can improve segmentation accuracy.

Purpose of the Study:

• To develop an enhanced multi-scale network for accurate polyp segmentation.
• To leverage multi-scale features for improved diagnostic capabilities.
• To address the limitations of existing polyp segmentation methods.

Main Methods:

• Proposed an enhanced multi-scale network integrating U-Net+++, Res2Net, Receptive Field Block (RFB) modules, and Local Context Attention (LCA) modules.
• Utilized multi-scale skip connections to aggregate features from various layers.
• Evaluated the model on EndoScene, CVC-ClinicDB, and Kvasir-SEG datasets.

Main Results:

• Achieved state-of-the-art (SOTA) performance compared to existing methods.
• Demonstrated superior segmentation accuracy on challenging datasets.
• The model was the only one to exceed 0.900 mean Dice score on EndoScene and CVC-ClinicDB.

Conclusions:

• The proposed enhanced multi-scale network significantly improves polyp segmentation accuracy.
• This advancement holds promise for more effective colorectal cancer diagnosis and treatment planning.
• The model's performance highlights the importance of multi-scale feature aggregation in medical image analysis.