---
title: 'BleedOrigin-Bench: ESD Bleeding Source Benchmark'
url: https://www.emergentmind.com/topics/bleedorigin-bench
type: topic
---

# BleedOrigin-Bench: ESD Bleeding Source Benchmark

BleedOrigin-Bench is a comprehensive, large-scale benchmark dataset designed for bleeding source detection and tracking in Endoscopic Submucosal Dissection (ESD). Developed to address the need for precise, real-time localization and temporal tracking of bleeding sources during ESD—where intraoperative bleeding presents significant risks—BleedOrigin-Bench provides high-fidelity expert annotations, extensive pseudo-labeling, robust coverage of anatomical and clinical variability, and clearly defined evaluation protocols. The dataset enables systematic development and validation of Artificial Intelligence (AI) methods for dynamic bleeding source localization, a task distinct from the more commonly addressed bleeding region segmentation.

## 1. Dataset Structure and Composition

BleedOrigin-Bench comprises data sourced from 44 distinct ESD procedures, encompassing a total of 33.54 hours of raw endoscopic video originally sampled at 30 frames per second (FPS) and down-sampled to 1 FPS for annotation. After preprocessing (overlay/margin removal), the dataset contains 106,222 frames. Key annotated components include 1,771 frames with expert-annotated bleeding source coordinates and 39,755 frames with pseudo-labeled source coordinates generated for tracking supervision purposes. A total of 485 bleeding clips were extracted, segmented into detection and tracking subsets.

| Component                 | Quantity     | Description                                  |
|---------------------------|-------------|----------------------------------------------|
| Total patients            | 44          | Unique ESD procedures                        |
| Total frames              | 106,222     | Down-sampled, preprocessed                   |
| Expert-labeled frames     | 1,771       | "Ground-truth" source point per frame        |
| Pseudo-labeled frames     | 39,755      | Tracking supervision (see §3)                |
| Detection clips           | 485×150     | 66,896 frames, one onset per clip            |
| Tracking set (same clips) | 485         | Covers both expert and pseudo labels         |

Clips are systematically organized into detection (fixed onset, 150 frames/clip) and tracking sets (variable length, 31–731 frames/clip). The detection subset provides a tightly controlled scenario for frame-level onset identification, while the tracking subset enables evaluation of spatial-temporal localization performance.

## 2. Coverage of Anatomical Sites and Clinical Scenarios

To ensure generalizability and clinical relevance, BleedOrigin-Bench spans a broad spectrum of anatomies and operative challenges encountered in ESD. The dataset includes frames from the following eight anatomical sites: gastric antrum, gastric body, lesser curvature, cardia, duodenum, esophagus, ascending colon, and sigmoid colon.

Six clinically challenging scenarios are explicitly represented:
1. Clear bleeding view
2. Obscured bleeding view (e.g., blood pooling, tissue motion)
3. Camera jitter (endoscope hand motion)
4. Light-reflection glare
5. Water-flushing (intermittent field washouts)
6. Instrument interference (forceps, knives present)

Each clip may exhibit multiple artifacts, reflecting the diverse, rapidly changing conditions typical in ESD and supporting robust model development and benchmarking across real-world operational complexities.

## 3. Annotation Process and Pseudo-Label Generation

### Expert Annotation

Manual bleeding source annotation was conducted by four board-certified endoscopists using LabelMe. Each suspected source frame (sampled at 1 FPS) was annotated through a multi-stage consensus protocol:
- Two independent clinician teams marked all candidate frames.
- Majority-vote consensus resolved inter-annotator disagreements.
- Researcher-led review filtered ambiguous or low-quality labels.
This yielded 1,771 expert-generated “ground-truth” points. The bleeding source is strictly defined as the 2D center of the vessel or orifice from which blood first emerges immediately following a flush-cleared view.

### Pseudo-Labeling for Tracking

Pseudo-label generation leverages feature-based propagation and transformer-based tracking:
- For each labeled frame pair $(I_t, I_{t+30})$, features within a 50 px radius of the ground-truth are extracted using XFeat, retaining matches with confidence above 0.7 (typically ~36 keypoints).
- Each pseudo keypoint is propagated over 30 subsequent frames using a pre-trained transformer tracker (Track-On).
- Trajectories are backward smoothed with a Kalman filter, yielding final pseudo-labeled points across 39,755 frames.
These pseudo-labels substantially augment sparse GT, enabling supervised training of tracking models within or across longer temporal windows.

## 4. Dataset Splits and Format

To avoid patient-wise information leakage, all dataset splits are performed at the patient level in a 4:1:1 ratio (train:val:test). This equates to 25 patients/319 clips for training, 6 patients/68 clips for validation, and 8 patients/98 clips for testing. The detection subset fixes the bleeding onset to frame 120 within each 150-frame clip; annotation is provided only at the onset frame. During training, the onset position is randomized within a 0–60 frame offset.

Tracking subset clips vary in length, annotated either as “short-clips” (31 consecutive frames, GT at endpoints) or “long-clips” (full video, sparse labeling every 30 frames). Data formats include:
- Images: 1240×1080 or 780×670 (JPEG/PNG)
- Annotation: Per-clip CSV/txt files listing (clip_id, frame_idx, x, y) for both GT and pseudo labels.

All labels are centralized in format-compatible text/CSV files, which are required for submissions to public benchmarks.

## 5. Evaluation Protocols

BleedOrigin-Bench defines standardized metrics and protocols for fair and reproducible evaluation:
- **Frame-Level Bleeding Onset Detection:** Prediction is correct if $|t_{pred} - t_{gt}| \leq k$, with accuracy reported at $k = 0, 1, 2, 4, 8$, and mean error ($Err_{avg}^{ibf}$) across all clips.
- **Pixel-Level Source Localization and Tracking:** Spatial accuracy is evaluated as the proportion of predictions within a given Euclidean distance $d$ (reported at 10, 25, 50, 75, 100 px), with mean error ($Err_{avg}$) in pixels.

Models are explicitly instructed to initialize tracking at the ground-truth point, and to adhere to specified frame and pixel tolerances, ensuring benchmarking uniformity. For long tracking clips, a memory refresh (tracker re-initialization) every 60 frames is recommended to control drift.

## 6. Usage Guidelines and Recommendations

Researchers are advised to maintain strict patient-level data separation during development and evaluation to prevent overfitting and leakage. Usage recommendations include:
- For detection testing, always use the GT onset frame for coordinate metrics (do not cascade detection output when replicating published results).
- For tracking, initialize with the GT point at the bleeding onset frame.
- Leverage both expert and pseudo labels in training to balance spatial precision and long-term tracking robustness, optionally integrating both short- and long-clip paradigms.
- A 60-frame sliding window is recommended for onset detection, aligning with clinical temporal context.
- Maintain the provided image resolution and avoid overlay artifacts to ensure reproducibility.

## 7. Significance and Comparative Context

BleedOrigin-Bench is the first dedicated, large-scale ESD bleeding source dataset with explicit support for temporal source tracking and onset detection under varied practical conditions [2507.15094]. Its depth and rigor support end-to-end benchmarking for bleeding source localization frameworks, as illustrated by results from BleedOrigin-Net—a dual-stage detection-tracking model—which achieved 96.85% frame-level accuracy (within $\pm8$ frames), 70.24% pixel-level accuracy ($\leq100$ px) for initial detection, and 96.11% pixel-level accuracy ($\leq100$ px) for tracking. The dataset's explicit differentiation from general bleeding region segmentation, its inclusion of real-world operational artifacts, and its clear annotation/metric protocols make it a primary reference for researchers developing AI systems in ESD guidance or related medical visual tracking contexts.

Source: https://www.emergentmind.com/topics/bleedorigin-bench