Spotting Diabetic Retinopathy by analyzing medical images pixel by pixel

Share this post:

Medical images are a rich source of data for clinicians in their diagnosis and treatment of diseases. In fact, specialized fundus photography can help pinpoint tiny pathologies in the eyes of diabetics, revealing signs of diabetic retinopathy (DR), one of the world’s leading causes of blindness.

In the vast majority of these cases, early detection is the key to a patient’s survival and treatment outcome. Yet it is estimated that half of Australians with diabetes do not undergo the recommended frequency of screening. even though early intervention can reduce the risk of blindness by 95 percent (CERA). And this is not only a challenge in Australia. Eighty percent of blindness worldwide is preventable if detected and treated early (WHO).

A new vision for preventable blindness

While education is a major part of encouraging regular screening, ensuring easy access for all Australians is also a key factor. Picking up subtle signs of DR in images is often a manual and subjective process for clinicians. Accuracy of the diagnosis heavily depends on their level of expertise, which can be hard to come by, particularly in rural or remote communities. We believe we can help with this challenge. Scientists at IBM Research-Australia are studying how new cognitive technologies could support clinicians in streamlining their analysis of images, as well as enable greater access to health services for everyone, regardless of their location.

Last week, my team and I presented new research findings at the 2017 IEEE International Symposium on Biomedical Imaging (ISBI), using deep learning and visual analytics technology for the early detection of DR.

Our new method uses deep learning algorithms combined with pathology insights to analyze real-world images of both the left and right eyes of patients from datasets collected by EyePACS. The key to identifying DR is looking for tiny signs of hemorrhages and micro-aneurysms, which signal to a doctor the presence and severity of the disease. The algorithm analyzes images pixel by pixel and patch by patch and, based on the visual characteristics, learns patterns associated with a particular pathology and disease. It takes a lot of training and feedback for the algorithms to identify, for example, mild diabetic retinopathy vs. moderate, but they get better over time as they’re trained on more samples.

The number of pathologies and where they are distributed in the eye informs the clinician of the severity level of the disease. At ISBI, we presented a novel method for classifying severity across the five levels of the international scale for DR: none, mild, moderate, severe and proliferative.

Our technology combines two various analytics approaches into one hybrid method, wherein a convolutional neural networks (CNN)-based method for DR classification is integrated with a dictionary-based learning that incorporates DR-specific pathologies. This hybrid analysis resulted in a great improvement in classification accuracy1. Our method takes approximately twenty seconds to analyze the image and achieves an accuracy score of 86 percent in classifying the disease across the five severity levels.

An eye scan with Diabetic Retinopathy haemorrhages highlighted

An eye scan with Diabetic Retinopathy hemorrhages highlighted

We also showed a new method for the accurate segmentation of the fovea in retinal color fundus images. Fovea is responsible for sharp central vision. Location of retina lesions and pathologies with respect to the fovea impacts their clinical relevance. Our technology allows for a pixel-wise segmentation of the fovea. It does not require prior knowledge of the location of other retinal structures such as optic disc or retina vasculature. This is an advantage over other published methods, which can either localize only the center of the fovea or need a priori information about major structures in the image2.

These same methods can also be applied to other eye diseases, so we will continue to extend our research to support the early detection of all preventable causes of blindness worldwide.

Our ongoing challenge is improving accuracy and providing deeper insight into severity levels for clinicians. In addition, it is also very important to us to validate our technologies with new data sources. We want to make sure that our methods do not only work with one source of data, but many different types of databases.



[1] P. Roy, R Tennakoon, K. Cao, S. Sedai, D. Mahapatra, S Maetschke, R. Garnavi, “A novel hybrid approach for severity assessment of diabetic retinopathy in colour fundus images”, to appear in the proceeding of The 2017 IEEE International Symposium on Biomedical Imaging.

[2] S. Sedai, R Tennakoon, P. Roy, K. Cao, R. Garnavi, “Multi-stage segmentation of the fovea in retinal fundus images using fully convolutional neural networks”, to appear in the proceeding of The 2017 IEEE International Symposium on Biomedical Imaging.



Research Manager, Research Staff Member, and Master Inventor, IBM Research

More Healthcare stories

Largest Dataset for Document Layout Analysis Used to Ingest COVID-19 Data

Documents in Portable Document Format (PDF) are ubiquitous with over 2.5 trillion available from insurance documents to medical files to peer-review scientific articles. It represents one of the main sources of knowledge both online and offline. For example, just recently The White House made available the COVID-19 Open Research Dataset, which included links to 45,826 papers […]

Continue reading

Meet Ari, the Smart Bike that Helps you Catch Green Lights

As more bikes take to the city streets than ever, researchers at RMIT have designed Ari, a smart e-bike that could help riders cruise the ‘green wave’.

Continue reading

Bank Guarantees Go Digital with Blockchain

IBM has developed Lygon, a first-of-a-kind platform to digitise and transform the bank guarantee process.

Continue reading