BleedOrigin GitHub
arXiv · Computer Vision 2025

BleedOrigin

Dynamic Bleeding Source Localization in Endoscopic Submucosal Dissection via Dual-Stage Detection and Tracking

Mengya Xu* Rulin Zhou* An Wang* Chaoyang Lyu Zhen Li Ning Zhong Hongliang Ren†

* Equal contribution    † Corresponding author

The Chinese University of Hong Kong · CUHK Shenzhen Research Institute · Qilu Hospital of Shandong University

PDF Paper α arXiv </> Code CITE BibTeX

From onset to hemostasis

Detect the first bleed.
Keep the source in sight.

BleedOrigin connects two tasks that are usually studied separately: prompt bleeding-onset detection and continuous source tracking.

Predicted source Bleeding alert
Clear bleeding view Real-time source localization
106,222 Endoscopic frames
44 ESD procedures
8 Anatomical sites
6 Clinical challenges

TL;DR

BleedOrigin gives endoscopists an early warning when bleeding starts, identifies its exact origin, and keeps tracking that point through blood, flushing, reflection, camera motion, and instrument interference.

Clinical problem

A few milliseconds before the source disappears.

During ESD, blood can rapidly obscure the field. Endoscopists often flush repeatedly to regain only a brief view of the true source.

Clinical motivation and BleedOrigin detection-tracking workflow
The system identifies the initial bleeding source and maintains its location until hemostasis.
01 Bleeding begins

The source is briefly visible in a dynamic endoscopic scene.

02 View becomes obscured

Blood, smoke, reflection, or instruments hide visual evidence.

03 Repeated flushing

The clinician repeatedly clears the field to relocate the point.

04 BleedOrigin assists

An alert and a persistent location guide the next intervention.

BleedOrigin-Bench

A benchmark designed around the complete bleeding event.

From the first bleeding frame to long-term point tracking, the benchmark captures both when bleeding begins and where it comes from.

BleedOrigin-Bench dataset composition and clinical diversity
A

Onset detection

66,896 frames

485 clinically annotated bleeding time points.

B

Source tracking

1,771 expert labels

Plus 39,755 dense pseudo-labeled frames.

C

Clinical diversity

8 anatomical sites

Collected from 485 clips across 44 procedures.

BleedOrigin-Net

One pipeline, two coordinated stages.

The first stage recognizes the onset and localizes the source. The second maintains that location through the rest of the event.

BleedOrigin dual-stage detection and tracking framework
Stage 01

Detection

Recognize onset. Pinpoint origin.

Multi-Domain Confidence-based Frame Memory builds temporal context from complementary RGB, HSV, and optical-flow cues.

  • MDCFM multi-domain frame memory
  • Multi-Domain Gated Attention
  • Frame-level onset and pixel-level source prediction
Stage 02

Tracking

Follow the point through change.

Sparse expert annotations are expanded into dense supervision with feature matching, trajectory prediction, and Kalman filtering.

  • Pseudo-label enhanced training
  • Memory refresh for re-identification
  • Parameter-efficient LoRA fine-tuning

State-of-the-art performance

Accurate at the moment of onset—and afterward.

96.85%

Onset detection

Frame-level accuracy within ±8 frames.

70.24%

Initial localization

Pixel-level accuracy within 100 pixels.

96.11%

Continuous tracking

Pixel-level accuracy within 100 pixels.

Evaluated against YOLOv11/v12, multimodal large language models, and point-tracking baselines.

Six surgical scenarios

Localization when the visual evidence changes.

Select a scenario to inspect predictions under common ESD disruptions.

Scenario 01

Clear bleeding view

A visible source provides the reference case for prompt onset localization and stable tracking.

Prediction Ground truth

Clinical deployment study

Alert first. Re-identify when evidence returns.

In high-difficulty deployment footage, copious flushing may cause temporary tracking drift. The memory-refresh strategy supports source re-identification once visual features reappear.

Real-time bleeding state Pulsating source alert Memory-based recovery
High-difficulty deployment example

What this work contributes

From a segmentation problem to a clinical workflow.

01

A new benchmark

Expert labels and dense pseudo-labels for onset, origin, and continuous tracking across diverse ESD cases.

02

A dual-stage network

Multi-domain temporal memory for detection, followed by efficient point tracking with recovery.

03

A deployment-oriented task

The system is designed around the path from early warning to hemostatic intervention, not a single isolated frame.

Publication

BleedOrigin

Dynamic Bleeding Source Localization in Endoscopic Submucosal Dissection via Dual-Stage Detection and Tracking

Year
2025
Archive
arXiv
ID
2507.15094
Class
cs.CV
arxiv.org/abs/2507.15094
BibTeX 
@misc{xu2025bleedorigin,
  title  = {BleedOrigin: Dynamic Bleeding Source
            Localization in Endoscopic Submucosal
            Dissection via Dual-Stage Detection
            and Tracking},
  author = {Xu, Mengya and Zhou, Rulin and Wang, An
            and Lyu, Chaoyang and Li, Zhen and
            Zhong, Ning and Ren, Hongliang},
  year   = {2025},
  eprint = {2507.15094},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}