New Deep-Learning Resource Model Can Identify Many Common Eye Diseases

New Deep-Learning Resource Model Can Identify Many Common Eye Diseases

June 28, 2022

The automation of disease diagnosis relies on deep learning models that can accurately and efficiently identify measurements of tumors, tissue volume, or other types of abnormalities. Now, researchers from Tohoku University have unveiled a new lightweight model that can identify many common eye diseases.

A new deep learning (DL) model capable of identifying disease-related features from eye images has been unveiled by a group of researchers from Tohoku University. This “lightweight” DL model can be trained with a small number of images, even those with a high degree of noise, and is resource efficient, meaning it is deployable on mobile devices.

The details were published in the journal Scientific Reports on May 20, 2022.

With many societies aging and medical personnel limited, self-monitoring and remote disease screening based on the DL model are becoming increasingly common. Yet deep learning algorithms are usually task-specific and identify or detect general objects such as humans, animals, or traffic signs.

Disease identification, on the other hand, requires precise measurement of tumors, tissue volume, or other types of abnormalities. To do this, a model must examine separate images and mark boundaries in a process known as segmentation. But accurate prediction requires more computing power, which makes them difficult to deploy on mobile devices.

“There is always a trade-off between accuracy, speed, and computational resources when dealing with DL models. Our developed model has better segmentation accuracy and model training reproducibility, even with fewer parameters, which makes it efficient and lighter compared to other commercial software,” said Toru Nakazawa, study co-author and professor in the department of ophthalmology at Tohoku University.

Professor Nakazawa, Associate Professor Parmanand Sharma, Dr. Takahiro Ninomiya and students from the Department of Ophthalmology worked with Professor Takayuki Okatani of Tohoku University’s Graduate School of Information Sciences to produce the model.

Using low-resource devices, they obtained measurements of the foveal avascular zone, an area with the fovea centralis in the center of the retina, to improve detection of glaucoma.

“Our model is also able to detect/segment optic discs and hemorrhages in fundus images with high accuracy,” Nakazawa added.

In the future, the group hopes to deploy the lightweight model to screen for other common eye disorders and other diseases.

Journal reference:

Sharma, P. et al. (2022) A lightweight deep learning model for automatic ophthalmic image segmentation and analysis. Scientific reports.