Computational framework to interpret chest X-rays and diagnose pneumonia
This project is based at Child Health Research Foundation in Bangladesh
- Project title: Construction of a computational framework to automatically interpret chest X-rays and diagnose pneumonia
- Acute or chronic: Acute
- Based at: Child Health Research Foundation (CHRF)
- Start date: 01 July 2018
- End date: 28 February 2020
- Principal investigator: Professor Samir K Saha
- Project team: Professor Samir K Saha, Dr Senjuti Saha, Dr Mark Sun
It is estimated that 95% of the two million deaths due to pneumonia occur in developing countries. In Bangladesh alone, six million cases of pneumonia are diagnosed every year. Unfortunately, diagnostic methods to date lack sensitivity or are difficult to fully standardized. The lack of a reliable diagnostic hampers the execution of evidence based interventions, impacting the monitoring of interventions, like vaccines.
The “gold standard” for defining pneumonia are chest X-rays. However, the interpretations are subjective, sometimes requiring multiple radiologists/clinicians to reach a conclusive diagnosis. As there are few well-trained radiologists/clinicians in resource-poor settings, having a tool to aid in the diagnosis of pneumonia would be invaluable in the impact monitoring of interventions.
Aim and impact
The aim of the project is to construct a computational framework to automatically and systematically interpret paediatric chest X-rays to diagnose pneumonia.
The automated diagnostic system will enable members of the broader community, such as health care workers in areas without experts, to efficiently diagnose pneumonia. If successful, it will eventually incentivise the use of chest x-ray as the go-to diagnostic for evidence-based interventions.
Additionally, this project will aid in clinical studies aiming to monitor the impact of interventions like vaccines, support patient management by providing real-time interpretations, and to facilitate antibiotic stewardship.
- Readers have begun assessing a bank of paediatric x-ray images to create a training set. To date, 7000 x-ray images have been labelled by one reader of which 3500 x-ray images have been rescreened by a second reader. Images with a discordant label are being read by a third reader. In this manner the training set used to build the computation model is of a high quality.
- Performance assessment of the computational prediction model is underway.
- A protocol paper is being written.