DYNAFormer: Enhancing transformer segmentation with dynamic anchor mask for medical imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of colorectal polyps, particularly sessile and pedunculated types, is crucial for cancer diagnosis and risk assessment.
• Existing datasets and models have limitations in segmenting these specific polyp shapes, hindering diagnostic accuracy.

Purpose of the Study:

• To introduce PolypDB_INS, a comprehensive dataset for polyp segmentation.
• To propose DYNAFormer, a novel transformer-based model for enhanced polyp segmentation.
• To improve the accuracy of colorectal polyp diagnosis and cancer risk assessment through advanced AI.

Main Methods:

• Development of PolypDB_INS dataset with 4403 images and 4918 annotated sessile and pedunculated polyps.
• Proposal of DYNAFormer, a transformer model employing anchor mask-guided mechanisms, cross-attention, dynamic query updates, and query denoising.
• Evaluation using standard instance and semantic segmentation metrics on the PolypDB_INS dataset.

Main Results:

• DYNAFormer demonstrated significant performance improvements over state-of-the-art methods in polyp segmentation.
• Ablation studies validated the effectiveness of the individual components within the DYNAFormer model.
• The model's robustness in segmenting complex polyp structures was confirmed, aiding in diagnostic capabilities.

Conclusions:

• The PolypDB_INS dataset and DYNAFormer model offer a robust solution for colorectal polyp segmentation.
• The proposed methods enhance the precision of polyp diagnosis, contributing to better colorectal cancer risk assessment.
• The developed AI tools show promise for clinical application in gastroenterology and oncology.