The so-called 'lightweight' deep learning (DL) model can be trained with a small number of images, including ones with a high degree of noise, and can be used on mobile devices.
Details of the research have been published in Scientific Reports.
According to Tohoku University, DL model reliant self-monitoring and tele-screening of diseases are becoming more routine but deep learning algorithms are generally task specific, identifying or detecting general objects such as humans, animals, or road signs.
Identifying diseases demands precise measurement of tumours, tissue volume, or other sorts of abnormalities. To do so requires a model to look at separate images and mark boundaries in a process called segmentation. Accurate prediction takes greater computational output, rendering them difficult to deploy on mobile devices.
"There is always a trade-off between accuracy, speed and computational resources when it comes to DL models," said Toru Nakazawa, co-author of the study and professor at Tohoku University's Department of Ophthalmology. "Our developed model has better segmentation accuracy and enhanced model training reproducibility, even with fewer parameters - making it efficient and more lightweight when compared to other commercial softwares."
Professor Nakazawa, Associate Professor Parmanand Sharma, Dr Takahiro Ninomiya, and students from the Department of Ophthalmology worked with Professor Takayuki Okatani from Tohoku University's Graduate School of Information Sciences to produce the model.
Using low resource devices, they obtained measurements of the foveal avascular zone, a region with the fovea centralis at the centre of the retina, to enhance screening for glaucoma.
"Our model is also capable of detecting/segmenting optic discs and haemorrhages in fundus images with high precision," Nakazawa said in a statement.
Looking forward, the group hopes to deploy the lightweight model to screen for other common eye disorders and other diseases.