Dr Yulia Arzhaeva1, Dr Dadong Wang1, Mr Saeed Amirgholipour Kasmani1, Dr Rhiannon McBean3, Dr Katrina Newbigin3, A/Prof Deborah Yates2
1CSIRO Data61, Marsfield, Australia,
2St Vincent’s Hospital , Sydney, Australia,
3Wesley Medical Imaging, Brisbane, Australia
Mixed dust pneumoconiosis (MDP) is caused by long-term inhalation of respirable dust, such as coal, asbestos, and silica, and it is more commonly known as black lung. About 25,000 people died of MDP globally in 2013. In 2017 there were more than 50 confirmed cases of mine dust lung diseases among current and former Queensland and NSW mine workers. In mining industry, periodic chest X-ray checking is employed for respiratory health surveillance of miners. However there are only 13 certified radiologists in Australia to identify MDP in chest radiographs. We address this problem by developing automated and quantitative tools for objective and systematic detection and monitoring of MDP on chest imaging.
We have collected more than 100K of chest radiographs from public and proprietary sources encompassing normal radiographs, radiographs with various stages of MDP, as well as radiographs with other abnormalities. Using these data we investigated several machine learning approaches to detect abnormal chest X-rays and diagnose MDP, including traditional machine learning (ML) methods that use custom image features, as well as deep neural networks (DNNs). The best performing ML method achieved accuracy of 77% in detecting pneumoconiosis. DNNs were trained to differentiate between normal and abnormal radiographs. The best classification accuracy achieved was 68%. In this presentation we describe our approaches and discuss possible improvements to DNNs that could lead to a better classification performance.
The project has been financially supported by Coal Services NSW Health and Safety Trust since May 2017.
Biography:
Dr Yulia Arzhaeva obtained her PhD in Medical Imaging Analysis from Utrecht University, the Netherlands, in 2009. Since then she has been with CSIRO Quantitative Imaging working on biomedical applications and algorithms.