New applications of AI in the form of deep learning are helping healthcare practitioners detect diabetic retinopathy (DR) earlier and more accurately. DR is a serious eye disease associated with long-standing diabetes that results in progressive damage to the retina, eventually leading to blindness. However, if DR is detected early on, vision impairment can be prevented with laser treatments.
Some diabetes and DR facts:
Worldwide, 425 million adults (1 in 11) have diabetes; half the cases are undiagnosed. This number is expected to reach 642 million by 2040.
Of U.S. citizens, 30.3 million, or nearly one in ten, have diabetes and 84.1 million adults, approximately one in three, have prediabetes.
Of all visual impairment caused by DR, 80% can be cured with regular screening and early detection. Major vision loss due to DR is preventable with timely remedial intervention like eye exams (including visual acuity testing, tonometry and pupil dilation) and regular screening at the earlier stages by using a DR diagnostic and grading screening tool.
Need for Automated DR Screening
DR screening is handicapped by a lack of trained clinicians, as well as challenges posed by data set availability, fundus image analysis and multiple camera imagery. Moreover, the screening process is time-consuming. The delay in delivering results can lead to lost follow-ups, miscommunication, and missed or postponed treatments — all of which may increase the probability of vision loss.
There is a need for an automated DR detection system that can input retinal images from color fundus photography, provide a quick DR classification with confidence and refer the patient to specialists if needed. This will enable doctors addressing DR cases to utilize their time effectively and thereby treat more DR patients in a timely fashion.
We developed a DR automated detection system solution that makes use of machine learning techniques such as deep convolutional neural networks (CNNs) — neural networks that are used to analyze and classify visual imagery. The CNN extracts diagnostic features using a deep learning algorithm trained to classify images across labels to determine whether or not the patient has DR.
Our AI-based DR diagnostic tool helps doctors detect and grade the level of DR disease based on the fundus images. This DR tool will enable doctors to view variations from multiple fundus camera images with the help of image preprocessing techniques. The tool makes use of emerging machine learning technology to process fundus images quickly — and as accurately — as manual screening. Most important, it reduces the time taken for the whole process to less than a minute, from a minimum of 15 minutes manually. We are confident this speed will improve over time.
In the future, clinicians will be able to use our DR detection and grading solution as an app with a mobile-attached, hand-held fundus camera to diagnose DR immediately and guide patients toward further treatment.
DR Prediction & Grading
DR prediction is the process of identifying whether the patient is affected by DR, given the set of the patient’s input fundus images. DR grading is the process of identifying the stage of DR with the input of fundus images. This process makes use of a huge corpus of fundus images with labels varied from 0 to 4. (See the accompanying infographic for a depiction and description of the progressive stages of DR.)
A deep learning model has been trained with a corpus of fundus images that have undergone a series of image preprocessing operations. The images are used to extract features using CNN, which in turn passes the features on to a classification model to predict whether the given image is affected by DR or not, and predict the disease grading level.
Key Benefits of Our Approach
Early detection of DR. This will help people to retain their sight and enable specialists to focus on treatments.
Automatic feature extraction. We use deep learning techniques and iterative learning to continuously improve outcomes.
Overall solution can incorporate customer feedback.
Advanced image processing capabilities. These allow clinicians to work even with low-resolution fundus images.
Lower cost than manual methods.
Accuracy of 90%, which will improve as additional data sets are gathered.
Highly scalable process with quick response time.
Solution is extensible (e.g., for detection of glaucoma, retinopathy of prematurity, etc.).
Advances in mobile hardware and OS support for machine and deep learning are enabling both iPhone and Android smartphones to run stronger forms of AI for offline medical imaging solutions. Many new-age handheld fundus cameras can be attached directly to mobile phones. The cost and size of mobile-ready handheld fundus imaging equipment is gradually declining while fundus image quality is improving.
These developments are opening up new possibilities for running DR detection and grading algorithms. These algorithms can either run directly on the fundus camera itself (DR on device) or run on mobile device/desktop (DR on mobile) or on the cloud (DR on the cloud) to provide an immediate outcome.
The approach described above could overcome the barriers to reaching more diabetic patients and provide regular DR screening checkups worldwide.