Introduction
In clinical colonoscopy, quality is often measured by the Adenoma Detection Rate (ADR) and Polyps Per Colonoscopy (PPC). While AI has made significant strides in frame-level polyp detection, accurately counting polyps throughout a full 20-to-60-minute procedure remains a formidable challenge. The primary obstacle is the "re-identification" (ReID) task: determining whether a polyp seen at one moment is the same one encountered earlier after camera movement or cleaning. This paper presents a novel framework that treats polyp counting as a ReID problem, significantly reducing "fragmentation"—the erroneous counting of a single polyp multiple times.
Methodology: A Two-Stage Pipeline
The proposed approach decouples the counting problem into two distinct stages: feature encoding and re-association via unsupervised clustering.
1. Feature Encoding
Visual encoders are trained using contrastive learning (SimCLR) to generate robust representations of polyp tracklets. Two architectures were explored:
- Single-Frame Encoder (SFE): Learns from individual bounding boxes.
- Multi-View Encoder (MVE): Utilizes sequences of frames to capture the 3D structure and varying perspectives of a polyp.
2. Re-association via Clustering
The core innovation lies in replacing simple similarity thresholds with global unsupervised clustering. By analyzing all tracklets within a video simultaneously, the model can better distinguish between unique polyps. Among the methods tested—including Hierarchical Clustering and HDBSCAN—Affinity Propagation emerged as the most effective. It automatically determines the number of clusters and identifies "exemplars" for each polyp, providing a more stable and accurate count than brittle threshold-based methods.
Experimental Setup
The framework was evaluated using the REAL-Colon dataset, the first open-access dataset of its kind, featuring 60 full-procedure videos. To focus specifically on the ReID challenge, the authors established a standardized baseline using tracklets derived directly from ground-truth annotations. A balanced split was defined for the community: 32 videos for training, 14 for validation, and 14 for testing.
Results and Discussion
The implementation of unsupervised clustering led to a dramatic improvement in counting accuracy. The proposed method achieved a state-of-the-art Fragmentation Rate (FR) of 6.30 on the test set, a nearly fourfold reduction compared to previous threshold-based approaches (FR of 24.60).
Key findings include:
- Clustering Superiority: Global clustering provides a "big picture" view that prevents the over-counting common in local thresholding. Without ReID, the Fragmentation Rate was as high as 56.33.
- Low False Positives: The system maintained a False Positive Rate (FPR) below 5%, ensuring it rarely misidentifies distinct polyps as the same entity.
- Encoder Performance: While the MVE is theoretically superior, it showed only marginal gains over the SFE on this dataset, likely due to the data-intensive nature of its Transformer-based architecture.
Limitations and Future Work
While successful, the current approach is an "offline" process, clustering the video after the procedure is complete. Future research will focus on developing "online" clustering methods for real-time feedback during colonoscopies. Additionally, integrating this ReID framework with real-time object trackers like ByteTrack will be essential for handling the noisy detections encountered in clinical practice.
Summary updated