Related Experiment Video
Updated: Aug 30, 2025
10:41
Analysis of Cerebral Vasospasm in a Murine Model of Subarachnoid Hemorrhage with High Frequency Transcranial Duplex Ultrasound
Published on: June 3, 2021
3.7K
Automatic Hemorrhage Detection From Color Doppler Ultrasound Using a Generative Adversarial Network (GAN)-Based
Jhimli Mitra1, Jianwei Qiu2, Michael MacDonald1,2
1General Electric Research Niskayuna NY 12309 USA.
IEEE Journal of Translational Engineering in Health and Medicine
|September 2, 2022
Summary
An automated hemorrhage detection method using a Generative Adversarial Network (GAN) shows promise for real-time ultrasound guidance. This AI approach can identify bleeding in vascular injuries, aiding novel therapies like High Intensity Focused Ultrasound (HIFU).
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Trauma Care
Background:
- Exsanguination is a leading cause of preventable death in traumatic injuries, particularly non-compressible torso hemorrhage (NCTH).
- High Intensity Focused Ultrasound (HIFU) offers a non-invasive therapeutic approach for deep tissue hemorrhage control.
- Real-time interpretation of ultrasound color Doppler for hemorrhage detection is challenging, requiring expert sonographers.
Purpose of the Study:
- To investigate the feasibility of an automated hemorrhage detection system using Generative Adversarial Network (GAN) for anomaly detection.
- To develop a method for real-time identification of bleeding using ultrasound color Doppler images.
- To support the clinical application of HIFU therapy by improving hemorrhage localization.
Main Methods:
- Trained a GAN-based anomaly detection network on normal blood flow variability from ultrasound color Doppler images in a porcine femoral artery injury model.
- Extracted blood flow velocity information from images for network training and testing.
- Tested the network on normotensive, and post-injury (immediate, 10, and 30 minutes) images.
Main Results:
- The anomaly detection network successfully identified anomalous blood flow patterns indicative of hemorrhage.
- Residual images/reconstructed error maps demonstrated effectiveness in hemorrhage detection.
- Achieved an Area Under the Curve (AUC) of 0.90 (immediate), 0.87 (10 min), 0.62 (30 min), and an overall AUC of 0.83.
Conclusions:
- Automated hemorrhage detection using GANs is feasible for improving ultrasound-guided therapies.
- This AI-driven approach can assist in real-time identification of bleeding, overcoming limitations of expert sonographer dependency.
- The developed method shows potential for enhancing the utility of HIFU and other non-invasive hemorrhage control strategies in diverse settings.
