TLDR: A new AI system, BleedOrigin-Net, combined with a specialized dataset, BleedOrigin-Bench, accurately detects the onset and continuously tracks bleeding sources during Endoscopic Submucosal Dissection (ESD) procedures. This dual-stage framework uses advanced techniques to overcome visual challenges in surgery, significantly improving precision and potentially enhancing patient safety by enabling faster intervention.
Endoscopic Submucosal Dissection (ESD) is a groundbreaking procedure for treating early gastrointestinal cancers, allowing precise removal while preserving organ function. However, a significant challenge during these operations is intraoperative bleeding. This bleeding can obscure the surgical field, forcing surgeons to repeatedly flush the area to gain a momentary view of the bleeding source. This process is inefficient, prolongs the procedure, and increases risks for patients.
Current Artificial Intelligence (AI) methods have largely focused on identifying general bleeding regions, but they often miss the crucial need for precise detection and continuous tracking of the actual bleeding source in the dynamic and often obstructed environment of ESD. This gap is further widened by a lack of specialized datasets for this specific challenge.
To address these critical issues, researchers have introduced a comprehensive framework called BleedOrigin. This initiative includes two major components: BleedOrigin-Bench, the first large-scale dataset specifically for ESD bleeding sources, and BleedOrigin-Net, a novel AI system designed for dual-stage bleeding source localization and tracking.
Introducing BleedOrigin-Bench: A New Dataset for Surgical AI
The BleedOrigin-Bench dataset is a pioneering effort, comprising 1,771 expert-annotated bleeding sources across over 100,000 frames from 44 ESD procedures. It also includes nearly 40,000 pseudo-labeled frames, which are AI-generated labels that help train the system. This dataset is unique because it covers 8 different anatomical sites and 6 challenging clinical scenarios, such as obscured views, camera jitter, light reflection, water flushing, and instrument interference. This comprehensive data is vital for developing robust AI systems that can perform reliably in real-world surgical settings.
BleedOrigin-Net: A Dual-Stage Approach to Bleeding Management
BleedOrigin-Net is designed to handle the entire workflow of bleeding source localization, from detecting the very first sign of bleeding to continuously tracking its exact location. It operates in two main stages:
1. Initial Bleeding Detection
Detecting the exact moment bleeding begins and pinpointing its initial source is crucial. Bleeding often starts subtly and gradually spreads, making it hard to identify the onset with single-frame analysis. BleedOrigin-Net addresses this by using a Multi-Domain Confidence-based Frame Memory (MDCFM) module. This module leverages information from multiple visual aspects—like color (RGB and HSV) and motion (optical flow)—to maintain a robust temporal context. It selectively remembers ‘clean’ views from the past, helping the system distinguish true bleeding onset from transient visual disturbances like smoke or instrument movements. Additionally, a Multi-Domain Gated Attention (MDG) module helps the system focus on the most informative visual cues in challenging scenarios.
2. Continuous Bleeding Source Tracking
Once the bleeding source is identified, the system continuously tracks it, even when it’s temporarily obscured by blood flow, water flushing, or surgical instruments. This is achieved using a sophisticated pseudo-label enhanced strategy. Since manually annotating every frame for tracking is impractical, the system generates ‘pseudo-labels’ by matching features between sparsely annotated frames and then uses Kalman filtering to smooth out the predicted trajectory, making it more robust against visual noise and drift. The tracking model, based on a transformer architecture, is fine-tuned using a technique called Low-Rank Adaptation (LoRA), which efficiently adapts the model to the unique challenges of surgical videos without requiring extensive data or computational resources.
Also Read:
- EndoControlMag: Enhancing Vascular Visualization in Endoscopic Surgery
- SynDiff: Enhancing Medical Image Segmentation with Text-Guided Synthetic Data and Single-Step Diffusion
Impressive Performance and Clinical Impact
The BleedOrigin-Net framework has demonstrated state-of-the-art performance. It achieves 96.85% accuracy in detecting the initial bleeding frame (within an 8-frame tolerance), 70.24% pixel-level accuracy for initial source detection (within 100 pixels), and 96.11% pixel-level accuracy for continuous tracking (within 100 pixels). These results significantly outperform existing object detection models and even advanced multimodal large language models like ChatGPT-4o, Claude-3.5, Gemini, and Qwen2.5-VL, which struggled with the precision required for this task.
The clinical significance of this work is substantial. By providing real-time, precise bleeding alerts and continuous localization, BleedOrigin-Net can enable surgeons to intervene more promptly, potentially reducing the need for repeated water flushing, shortening operation times, and ultimately improving patient safety and outcomes. Feedback from experienced endoscopists has been overwhelmingly positive, with clinicians expressing a strong willingness to adopt the system, highlighting its potential to provide consistent support, especially for less experienced surgeons.
While the current dataset is from a single institution, future work aims to expand it with multi-institutional data and integrate depth information for even greater spatial precision. This research marks a significant step towards a paradigm shift in surgical hemorrhage management, moving from reactive treatment to proactive prevention. For more technical details, you can refer to the research paper.
