Biomarkers in retinal disease
Automated system outperforms human experts for biomarker identification
Initial trials of an automatic biomarker classifier show that it is capable of outperforming human experts in the task of biomarker detection in optical coherence tomography (OCT) scans, paving the way for more efficient and economical ocular disease screening and detection, according to a study presented at the 19th EURETINA Congress in Paris.
“The reality is that humans have a hard time with biomarker detection – at least, compared to a machine that has been specifically trained for the task,” said Thomas Kurmann, a PhD student at the ARTORG Center for Biomedical Engineering Research at the University of Bern.
OCT is the gold standard in the diagnosis of retinal diseases such as age-related macular degeneration (AMD) or diabetic macular oedema (DME), with trained experts diagnosing patient scans by searching for biological markers, said Mr Kurmann.
“It is a very difficult task for human experts. It is time consuming, tedious, error prone and expensive when one considers that it takes about five minutes per scan for an expert grader,” he said.
From this starting point, the team from the ARTORG Center jointly with the Department of Ophthalmology at the Bern University Hospital set out to test its hypothesis that an automatic biomarker classifier could perform as well or better than human grading experts in the task of biomarker detection in OCT B-scans.
The biomarker definition was mainly focused on AMD and DME patients in the outpatient clinic at Inselspital, Bern University Hospital. In total, 470 volume scans of 327 patients were included, making a total of 23,030 OCT B-scans, with an additional test set of 1000 Scans from 21 AMD patients using the same device type and scanning protocol .
Morphological biomarkers included subretinal fluid, intraretinal fluid, intraretinal cysts, hyperreflective foci, drusen, reticular pseudodrusen, epiretinal membrane (ERM), geographic atrophy, outer retinal atrophy and fibrovascular pigment epithelial detachment. A healthy biomarker was also included to denote the lack of any previously mentioned biomarker.
In terms of results, the automatic classifier achieved higher Kappa values for eight biomarkers compared to the median human grader, with highest values recorded for ERM and the lowest for reticular pseudodrusen.
“The biomarker detector is accurate and 100 times faster than a human grader. With a single server at our lab we can compute every single volumetric scan that is taken worldwide in real-time. Automated analysis of scans using machine learning algorithms provide a cost-effective and reliable alternative to assist ophthalmologists in clinical routine and research,” he concluded.
For more information about the study, please see https://www.nature.com/articles/s41598-019-49740-7