UM-Net: Rethinking ICGNet for polyp segmentation with uncertainty modeling.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Colorectal Cancer Detection

Background:

• Automatic polyp segmentation in colonoscopy is crucial for early colorectal cancer diagnosis.
• Previous models like ICGNet struggled with inconsistent image color distribution and lacked uncertainty measures.
• Inconsistent color distribution across datasets leads to overfitting and reduced focus on relevant polyp features.

Purpose of the Study:

• To develop an improved segmentation network (UM-Net) addressing color inconsistency and providing prediction uncertainty.
• To enhance the reliability and clinical applicability of automated polyp detection in colonoscopy.
• To make the model more robust to variations in imaging equipment and polyp characteristics.

Main Methods:

• Implemented a color transfer operation to make the model focus on polyp shape rather than color.
• Integrated an uncertainty measure, using variance to rectify it, to quantify prediction reliability.
• Extended the existing ICGNet architecture to incorporate these novel features.

Main Results:

• UM-Net demonstrated competitive performance across five polyp datasets.
• The color transfer method improved the model's focus on polyp morphology.
• The uncertainty quantification enhanced the trustworthiness of segmentation results.

Conclusions:

• UM-Net offers substantial practical value for polyp segmentation in colonoscopy.
• The developed method improves both learning ability and generalization capability compared to existing approaches.
• The inclusion of uncertainty measures aids physicians in making more informed diagnostic decisions.