InterNet: Detection of Active Abdominal Arterial Bleeding Using Emergency Digital Subtraction Angiography Imaging With Two-Stage Deep Learning

Xiangde Min, Zhaoyan Feng, Junfeng Gao, Shu Chen, Peipei Zhang, Tianyu Fu, Hong Shen, Nan WangFrontiers in Medicine 2022 · DOI

TL;DR

InterNet is a novel two-stage deep learning framework designed to automatically detect active abdominal arterial bleeding in emergency Digital Subtraction Angiography (DSA) images. By localizing the angiographic region before identifying specific bleeding sites, the system nearly halves the time required for clinical interpretation while significantly improving diagnostic sensitivity.

What Problem Does This Paper Address?

Active abdominal arterial bleeding is a high-stakes medical emergency that can lead to hemorrhagic shock if not treated immediately. The gold standard for diagnosis and treatment is Digital Subtraction Angiography (DSA)—a type of fluoroscopy used to visualize blood vessels.

However, interpreting DSA sequences in real-time is grueling. Radiologists often face high workloads and sleep deprivation, and subtle bleeding can be easily missed, especially by junior clinicians. Currently, physicians must review captured video sequences "offline" before performing an intervention. This paper introduces InterNet, an AI assistant designed to move this workflow toward a real-time procedure by automating the detection of bleeding foci.

Methods

The authors proposed a "cascade" architecture called InterNet, which breaks the problem into two distinct stages:

  1. Region Localization Stage (RLS): This stage identifies the specific angiographic region within the full original frame. By "cropping" the search area, the model reduces noise and computational redundancy.
  2. Bleeding Site Detection Stage (BSDS): Once the region is localized, this stage performs the heavy lifting of identifying the actual bleeding points.

Technical Architecture: * Backbone: The system uses a ResNet-50 architecture. * Multi-scale Detection: The researchers integrated Feature Pyramid Networks (FPN) to handle the fact that bleeding sites vary significantly in size and shape. * Detection Framework: They employed Faster R-CNN (Region-based Convolutional Neural Networks) with Region Proposal Networks (RPN) to classify targets and refine bounding boxes. * Preprocessing: To handle the low contrast of medical images, they used CLAHE (Contrast-Limited Adaptive Histogram Equalization) and various data augmentation techniques (random shifts and rotations) to prevent overfitting.

Datasets

The study utilized a retrospective dataset from a single center:

  • RLS Development: 546 sequences from 450 patients.
  • BSDS/InterNet Development: 160 consecutive patients with confirmed active bleeding.
  • Independent Testing: A balanced set of 100 patients (50 with active bleeding, 50 without) was used to evaluate the final "InterNet" system and compare its performance against human doctors.
  • Annotation: All data were manually annotated and segmented by two radiologists, with a third experienced radiologist resolving any disagreements.

Key Insights and Ideas

  • The Power of the Two-Stage Approach: The authors found that localizing the angiographic region first (RLS) was a "practical benefit." It allows the model to ignore irrelevant parts of the image, which is crucial given that bleeding sites are often sparse and subtle.
  • FPN is Essential for Medical Imaging: Adding the Feature Pyramid Network improved the Average Precision (AP) from 58.1% to 60.3%. This confirms that "one size does not fit all" in medical object detection; the model needs to "see" features at multiple scales to catch both tiny leaks and large hemorrhages.
  • Prioritizing Recall over Precision: In an emergency setting, a "false alarm" (low precision) is annoying, but a "missed bleed" (low recall) is fatal. The researchers intentionally tuned InterNet to a recall of 0.7, accepting a precision of 53% to ensure that the most critical cases are flagged for the physician.

Key Results

The results demonstrate a clear "AI-plus-human" advantage:

  • Diagnostic Accuracy: The Area Under the Curve (AUC) for a doctor assisted by InterNet was 0.803, compared to 0.759 for a doctor working alone.
  • Sensitivity Boost: Sensitivity for the doctor jumped from 73.13% to 88.06% when using the AI assistant.
  • Workflow Efficiency: Most impressively, the time required to interpret a DSA sequence was slashed by nearly 50%, dropping from 84.88 seconds to 43.78 seconds.
  • Model Performance: The RLS stage was nearly perfect, with an AP of 0.99, ensuring the second stage always had the right "crop" to work with.

Limitations and Future Work

While promising, the authors acknowledge several hurdles:

  • Single-Center Study: The data came from one hospital and one equipment manufacturer. The model's "generalizability" to different types of DSA machines or different hospital protocols remains unproven.
  • Sample Size: While 160 bleeding patients is substantial for this niche, deep learning typically thrives on much larger datasets.
  • Real-Time Integration: While the model is fast (approaching 4-5 frames per second), it hasn't yet reached the 6 frames per second required for true real-time synchronization with DSA hardware. Future work will likely involve optimizing the code (e.g., moving from Python to C++) to achieve that "live" overlay during surgery.

Summary updated

Abstract

Our InterNet system could assist interventional radiologists in identifying bleeding foci quickly and may improve the workflow of the DSA operation to a more real-time procedure.
BibTeX
@inproceedings{min2022internet,
  author = {Xiangde Min and Zhaoyan Feng and Junfeng Gao and Shu Chen and Peipei Zhang and Tianyu Fu and Hong Shen and Nan Wang},
  title = {{InterNet: Detection of Active Abdominal Arterial Bleeding Using Emergency Digital Subtraction Angiography Imaging With Two-Stage Deep Learning}},
  booktitle = {Frontiers in Medicine},
  year = {2022},
  doi = {10.3389/fmed.2022.762091},
}
Figure