Academic Leader for Research and Higher Degrees in the School of Life Sciences Professor Evariste Bosco Gueguim Kana, has collaborated with researchers in South Africa, Nigeria, Cameroon and the UNESCO regional office for Southern Africa to develop a web-based diagnostic tool for COVID-19 that employs machine learning to detect evidence of COVID-19 on chest radiographs (CXR).
The method could provide much-needed decision support for radiologists and clinicians tackling the pandemic.
Gueguim Kana explained that the medical interpretation of CXR to draw information about a possible disease takes time and requires substantial medical expertise, leading to delays in obtaining an outcome. In the preliminary report, published on theMedRxiv server, Gueguim Kana and colleagues used machine learning to develop a model that has the capacity to recognise pixels of glass-patterned areas on CXR as distinguishing features of COVID-19 and to differentiate these from other viral- and pneumonia-infected lungs or healthy CXR images.
The model comprises a web interface where medical practitioners can log on and upload chest X-ray images within stipulated specifications. The system analyses the image and generates an outcome, specifying its probability of certainty, within seconds.
The system, freely accessible online, is intended to help where there is a lack of available medical expertise and high demand for swift results to determine a diagnosis. The exponential growth of the COVID-19 pandemic has created an equal upsurge in the demand for chest radiographs. Considering that expert interpretation of these images may not be affordable or accessible in all clinical settings, and that screening goals recommended by the World Health Organization to flatten infection curves could generate hundreds of millions of images for analysis, the researchers developed a computer-assisted decision support to relieve the burden on human experts and help prioritise cases.
Leveraging the advances in Artificial Intelligence by using a machine-learning algorithm, the team trained a model architecture on over 9 000 chest X-ray images, of which 2 000 were for COVID-19 patients, then validated it on other images not used for training. They achieved an accuracy rate of more than 90%.
‘Further collaboration with the medical fraternity in the validation of locally available images, fine-tuning with more local datasets, or the extension of the model’s capacity to accurately discern other medical conditions, will create a robust free system that could save time, resources, and lives,’ said Gueguim Kana.
The first phase of the project involved the establishment of the online system and the pre-print publication on MedRxiv – the report was downloaded 100 times within 48 hours. Gueguim Kana and colleagues say that, with relatively little funding and commitment of time and resources, the project could move into a second phase of approval and wider implementation.
Gueguim Kana, whose background is in applied biology, microbiology and biotechnology, has conducted research in the application of artificial intelligence and machine learning to microbial processing and to bioproduction systems for commodities like biofuels.
He is an innovative teacher, whose use of information technology in his instruction earned him a Distinguished Teacher Award from UKZN for 2017. One of UKZN’s Top Published researchers, Gueguim Kana holds a C-rating from the National Research Foundation.
Words: Christine Cuenod