ECC-PolypDet: Enhanced CenterNet With Contrastive Learning for Automatic Polyp Detection.

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Accurate polyp detection is vital for early colorectal cancer (CRC) diagnosis.
• Challenges remain due to the complex colon environment and polyps with unclear boundaries.
• Existing methods struggle with computationally expensive context aggregation or lack prior polyp modeling.

Purpose of the Study:

• To develop a robust polyp detection framework addressing limitations of current methods.
• To improve the detection of concealed and small polyps in colonoscopy images.
• To enhance early diagnosis of colorectal cancer through advanced AI.

Main Methods:

• Proposed Enhanced CenterNet with Contrastive Learning (ECC-PolypDet) framework.
• Utilized Box-assisted Contrastive Learning (BCL) for better feature discrimination.
• Introduced Semantic Flow-guided Feature Pyramid Network (SFFPN) and Heatmap Propagation (HP) module for multi-scale feature aggregation and attention.
• Implemented IoU-guided Sample Re-weighting (ISR) for fine-tuning on hard samples.

Main Results:

• ECC-PolypDet demonstrated superior performance compared to state-of-the-art detectors.
• The model effectively captures concealed polyps by minimizing intra-class differences and maximizing inter-class differences.
• Enhanced recognition of small polyps through SFFPN and HP module.
• Fine-tuning with ISR improved robustness on challenging cases.

Conclusions:

• ECC-PolypDet offers a significant advancement in automated polyp detection.
• The proposed methods address key challenges in colonoscopy image analysis for CRC screening.
• This framework holds promise for improving the accuracy and efficiency of colorectal cancer diagnosis.