Novel artificial intelligence-based image recognition technology for point-of-care (POC) ultrasound confirmati
Technology
Ng Teng Fong Healthcare Innovation Programme
National Healthcare Group
31 December 2024
To develop an AI-based image recognition technology that can demonstrate high reproducibility in recognising the NGT within. This Proof-of-Concept shows that AI can be used to improve detection of NG tube on US with high accuracy.
Year Submitted: 2024
Published Date: 31 December 2024
Tags: Technology, Proof Of Concept, Artificial Intelligence, Machine Learning
About this Content
Aims
To develop an AI-based image recognition technology that can demonstrate high reproducibility in recognising the NGT within the stomach on US and high accuracy in categorising presence or absence of NGT within the stomach on US.
Background
Nasogastric tubes (NGT) are widely used in clinical practice for feeding, inserted via the nose into the stomach. Confirmation of placement is important prior to feeding. Conventional bedside techniques have limitations and do not provide direct tube visualisation. Ultrasound (US) is an alternative modality but is dependent on the skill of the US operator and suffers from variable image quality. This project aims to develop an AI-based image recognition technology to assist the US operator in recognising the NGT within the stomach, reducing the impact of variable image quality and operator-dependence.
Methods
The project consisted of three parts: 1. Developing the dataset of ultrasound images of the stomach with and without NG tube in situ. 2. Development of the AI software (Data-Efficient Medical Segmenter) trained and validated on the dataset. 3. Testing the AI software for performance of segmentation and classification of images into presence or absence of NG tube.
Results
The dataset consisted of 532 unique stomach ultrasound images with NG tube in situ and 301 unique stomach US images without NG tube. The AI model DEMS achieved higher evaluation metrics compared to other available models. The CNN Transformer Ensemble Classifier achieved higher accuracy and positive predictive value compared to all other available models. Sensitivity and specificity were comparable or better than conventional techniques and chest x-ray.
Conclusion
This Proof-of-Concept shows that AI can be used to improve detection of NG tube on US with high accuracy. It is comparable or better than conventional techniques and shows comparable or higher sensitivity than chest x-ray. When integrated into US systems, this will potentially allow time-savings and cost-savings.
Lessons Learnt
1. Collaboration between medical and non-medical personnel requires clear communication of aims and findings. 2. It is important to have a clear idea of the potential scope of the project and the available resources to ensure aims and KPIs are achievable.
Additional Information
Potential industry partners can be explored to integrate the AI software into commercial US systems.
Keywords
proof of concept, artificial intelligence, machine learning
Innovators' Details
Innovators' Details
Healthcare Cluster(s) | National Healthcare Group |
Organization(s) Involved | National University of Singapore, Tan Tock Seng Hospital |
Platform(s) | Ng Teng Fong Healthcare Innovation Programme |
Healthcare Professional Group(s) | Medical, Nursing, Allied Health |
Applicable Specialty or Discipline | Radiology, Geriatric Medicine, Nursing |
Project Lead(s) | Lee Chau Hung |
Project Member(s) | Shen Lei |
Connect with this contributor!
Lee Chau Hung - chau_hung_lee@ttsh.com.sg