Posters
Presenting Author Academic/Professional Position
Medical Student
Academic Level (Author 1)
Medical Student
Discipline/Specialty (Author 1)
Internal Medicine
Academic Level (Author 2)
Medical Student
Academic Level (Author 3)
Graduate Student
Academic Level (Author 4)
Graduate Student
Academic Level (Author 5)
Medical Student
Presentation Type
Poster
Discipline Track
Translational Science
Abstract Type
Research/Clinical
Abstract
Background: Visualizing the common carotid artery is a crucial step during ECMO (Extracorporeal Membrane Oxygenation) and ECPR (Extracorporeal Cardiopulmonary Resuscitation) procedures where timely and accurate cannulation is essential to sustain a patient’s vital functions. These procedures often involve critical time-sensitive scenarios where every second can make a significant difference in survival outcomes. Current cannulation techniques, such as the open cannulation method and the semi-Seldinger technique, are not only time-intensive but also require a high degree of precision to ensure proper arterial access. Delays or inaccuracies in identifying the common carotid artery can compromise the efficacy of these life-saving interventions. Recognizing these challenges, our project aimed to develop a cutting-edge neural network model leveraging ultrasound imaging. The goal of this innovative approach is to significantly enhance the speed and accuracy of common carotid artery detection, thereby optimizing the initiation of ECPR and improving overall patient outcomes during emergency situations.
Methods: To develop our neural network model, our team utilized a combination of publicly available data and hospital-acquired ultrasound DICOM files from the left and right neck regions. This process yielded over 1,000 ultrasound images, which were then converted into PNG format for processing. To further enhance data usability, we developed specialized software to segment live ultrasound videos into individual image slices. The dataset was then divided into training (80%), testing (10%), and validation (10%) sets to optimize model preparation. Leveraging TensorFlow, we employed a U-Net architecture featuring ReLU activation functions, convolution layers, and max pooling layers for deep learning-based image segmentation. The model was developed in Python 3.7 and utilized a high-performance GPU with 16 GB of RAM, ensuring high computational efficiency.
Results: The model exhibited consistently high performance across both publicly available and hospital-acquired ultrasound datasets, with an average processing time of just 2.5 seconds per image. The model underwent training for 50 epochs, optimizing with a dice loss function that prioritized pixel-wise accuracy in segmentation by accounting for overlap between predicted and actual regions. The resulting dice coefficient of 0.95 indicates the model’s high precision, indicating near-perfect similarity between predictions and ground truth segmentations.
Conclusion: The developed model demonstrates significant potential for expediting the initiation of ECPR by reducing setup times and, consequently, improving patient survival rates in critical situations. Future research could focus on real-time deployment in clinical settings where integration into portable ultrasound devices could enhance its practical utility. Additionally, expanding the model’s application to diverse clinical environments would address patient variability and ensure robustness across different populations. While our results are promising, further validation is necessary to evaluate performance under varying image quality and complex scenarios.
Recommended Citation
Bellamkonda, Arjun; Ede, Nneka; Abdallah, Marc; Gupta, Alaukik; Ortega, Gilmar; and Gadad, Bharathi S., "AI Driven Detection of Common Carotid Artery for Emergent Cardiac Procedures" (2025). Research Symposium. 24.
https://scholarworks.utrgv.edu/somrs/2025/posters/24
Included in
AI Driven Detection of Common Carotid Artery for Emergent Cardiac Procedures
Background: Visualizing the common carotid artery is a crucial step during ECMO (Extracorporeal Membrane Oxygenation) and ECPR (Extracorporeal Cardiopulmonary Resuscitation) procedures where timely and accurate cannulation is essential to sustain a patient’s vital functions. These procedures often involve critical time-sensitive scenarios where every second can make a significant difference in survival outcomes. Current cannulation techniques, such as the open cannulation method and the semi-Seldinger technique, are not only time-intensive but also require a high degree of precision to ensure proper arterial access. Delays or inaccuracies in identifying the common carotid artery can compromise the efficacy of these life-saving interventions. Recognizing these challenges, our project aimed to develop a cutting-edge neural network model leveraging ultrasound imaging. The goal of this innovative approach is to significantly enhance the speed and accuracy of common carotid artery detection, thereby optimizing the initiation of ECPR and improving overall patient outcomes during emergency situations.
Methods: To develop our neural network model, our team utilized a combination of publicly available data and hospital-acquired ultrasound DICOM files from the left and right neck regions. This process yielded over 1,000 ultrasound images, which were then converted into PNG format for processing. To further enhance data usability, we developed specialized software to segment live ultrasound videos into individual image slices. The dataset was then divided into training (80%), testing (10%), and validation (10%) sets to optimize model preparation. Leveraging TensorFlow, we employed a U-Net architecture featuring ReLU activation functions, convolution layers, and max pooling layers for deep learning-based image segmentation. The model was developed in Python 3.7 and utilized a high-performance GPU with 16 GB of RAM, ensuring high computational efficiency.
Results: The model exhibited consistently high performance across both publicly available and hospital-acquired ultrasound datasets, with an average processing time of just 2.5 seconds per image. The model underwent training for 50 epochs, optimizing with a dice loss function that prioritized pixel-wise accuracy in segmentation by accounting for overlap between predicted and actual regions. The resulting dice coefficient of 0.95 indicates the model’s high precision, indicating near-perfect similarity between predictions and ground truth segmentations.
Conclusion: The developed model demonstrates significant potential for expediting the initiation of ECPR by reducing setup times and, consequently, improving patient survival rates in critical situations. Future research could focus on real-time deployment in clinical settings where integration into portable ultrasound devices could enhance its practical utility. Additionally, expanding the model’s application to diverse clinical environments would address patient variability and ensure robustness across different populations. While our results are promising, further validation is necessary to evaluate performance under varying image quality and complex scenarios.
