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

• DOI: 10.48550/arxiv.2401.04961
• Publication Date: 2024-01-01
• Authors: Yuncheng Jiang; Zixun Zhang; Yiwen Hu; Guanbin Li; Xiang Wan; Song Wu; Shuguang Cui; Silin Huang; Zhen Li
• Keywords: Minimum bounding box; Feature (linguistics); Pyramid (geometry); Bounding overwatch

Accurate polyp detection is critical for early colorectal cancer diagnosis. Although remarkable progress has been achieved in recent years, the complex colon environment and concealed polyps with unclear boundaries still pose severe challenges this area. Existing methods either involve computationally expensive context aggregation or lack prior modeling of polyps, resulting poor performance challenging cases. In paper, we propose Enhanced CenterNet Contrastive Learning (ECC-PolypDet), a two-stage training \& end-to-end inference framework that leverages images bounding box annotations to train general model fine-tune it based on score obtain final robust model. Specifically, conduct Box-assisted (BCL) during minimize intra-class difference maximize inter-class between foreground backgrounds, enabling our capture polyps. Moreover, enhance recognition small design Semantic Flow-guided Feature Pyramid Network (SFFPN) aggregate multi-scale features Heatmap Propagation (HP) module boost model's attention targets. fine-tuning stage, introduce IoU-guided Sample Re-weighting (ISR) mechanism prioritize hard samples by adaptively adjusting loss weight each sample fine-tuning. Extensive experiments six large-scale colonoscopy datasets demonstrate superiority compared previous state-of-the-art detectors.