IHPC Tech Hub

IHPC Tech Hub showcases IHPC's suite of in-house developed applications, tools or technology to help you unlock the possibilities to overcome business challenges. Through providing valuable insights, you can now predict and shape the commercial outcomes, automate processes, and free up resources for repetitive and labour-intensive tasks. 

Discover the power of computational modelling, simulation and AI that brings about positive impact to your business. 

Interpretation of Chest X-rays using AI

Pneumonia is a common lung disease that could be caused by bacterial or viral infection. The Chest X-ray (CXR) is the most widely adopted screening method for the diagnosis of pneumonia. During a pandemic, CXR readings could highlight severe findings for early intervention. For example, affected patients displaying respiratory symptoms might require immediate supplemental oxygen therapy and ventilation administration. As the average time to report an abnormal case is about an hour, it is taxing for a radiologist to review each X-ray image in a time sequence. It could also be challenging to analyse up to 300 X-ray images a day during the peak of a pandemic.

To reduce the workload of radiologists’ workload and efficiently interpret a patient’s CXR image during the pandemic, IHPC developed an ensembled AI model which automatically detects pneumonia from the CXR images. The AI model achieves high accuracy (0.961 AUC) in clinical trials.

Features

  • Analyse each X ray image within 3 seconds (up to an hour for manual process) and highlight abnormal chest X rays with high accuracy (Area under the Curve, AUC = 0.961)
  • Improve the radiologists' diagnostic confidence of pneumonia cases and reduced the turn-around time by 20% than usual
  • Reduce risk to healthcare workers and other patients by segregating high risk individuals
AI Model for Chest X-rays

Credit: CXR images provided by Tan Tock Seng Hospital (TTSH)

The Science Behind

The AI model we build is a deep neural network (DNN) that predicts the probability of pneumonia for each incoming image. The DNN is composed of multiple layers of artificial neurons. Each neuron has multiple inputs and outputs and performs a certain type of math function. The prediction accuracy of the DNN depends on the parameters of the neurons. In the initial step, we set random parameters for the neurons and the train the model with CXR images labelled by radiologists (from both Tan Tock Seng Hospital and open-source data set) to improve the accuracy of the DNN model.

In the training process of the medical personnel using the AI tool, the labels of the CXR images are compared with the predictions made by the DNN and we adjust the weights according to the difference. After extensive training, the accuracy of the predictions made by the AI model proliferates until it meets the satisfying level.

Industry Applications

This AI model can be deployed in hospitals or community healthcare centers to quickly identify patients with pneumonia and help doctors/healthcare workers prioritise the treatment for these patients.

Acknowledgment

The AI model was jointly developed by radiologists from Tan Tock Seng Hospital (TTSH) and researchers from A*STAR’s Institute of High Performance Computing (IHPC) and Institute for Infocomm Research (I2R) for rapid flagging of pneumonia in suspected Covid-19 cases. Read more.

For more info or collaboration opportunities, please write to enquiry@ihpc.a-star.edu.sg.