A new deep-learning program can identify eyes with AMD with 94% accuracy.
Deep learning techniques can differentiate between optical coherence tomography (OCT) images that are normal and those that contain evidence of age-related macular degeneration (AMD) with 93.45% accuracy, per a new study.
The use of OCT scans has increased 70-fold between 2002 and 2009, per study data. The increased use of OTC scans has created a diagnostic burden on providers, who must identify pathologic OCT images accurately and quickly based on AMD characteristics, according to study author Aaron Lee (pictured), MD, MSCI, and colleagues at the Department of Ophthalmology at the University of Washington School of Medicine in Seattle.
Lee and colleagues suggest that computer-aided diagnoses (CAD) have the "potential for allowing more efficient identification of pathologic OCT images" by directing clinicians to areas of interest on images. They also theorized that training deep neural (or deep learning) networks to utilize images in large electronic databases for comparison could potentially relieve the burden of clinicians performing AMD diagnosis.
To prove their theory, which states that new developments in deep learning can assist with classification of images and can accurately distinguish AMD from normal OCT images, Lee and colleagues extracted macular OCT scans spanning 10 years (2006-2016) from the large Heidelberg Spectralis imaging database, and trained a deep learning module to identify atypical retinal images.
43,328 macular OCT scans (a total of 2.6 million images) from 9,285 patients — 4,302 normal OCT scans and 4,702 AMD OCT scans – in the Heidelberg Spectralis database were extracted along with associated electronic medical records (EMRs). EMRs were stripped of protected health data identifying patients but included clinical diagnosis and dates of clinical visits, as well as all laser procedures or intravitreal injections. Lee and colleagues used these images and EMRs to train an Oxford, or VG116, convolutional neural network to focus on key differences between normal and AMD OCT scans to determine the probability of AMD.
After training, the deep learning model could evaluate individual images for the probability of AMD in 4.97 milliseconds at an accuracy rate of 87.63%, a sensitivity of 84.63%, and a specificity of 91.54%. When averaging the result of multiple images the accuracy of diagnosis increased to 88.98%, and when averaging possibilities of AMD in all images from a single patient, a 93.45% accuracy was achieved. The increased accuracy of patient level diagnoses is due to aggregation of probabilities, according to Lee and colleagues.
"One of the most remarkable properties of deep learning is its ability to continue to improve and learn as more data is given for training. I do think that the accuracy will improve as more data is given. In addition, having data from multiple centers will be critical for improving the external generalizability of the model," Lee said in an interview with MD Magazine.
The study demonstrates that using deep learning software to assist with identification of AMD via OCT images is an effective and accurate means of reducing time to diagnoses. Lee stated that he believes "deep learning CAD for OCT images for delivery of care for AMD will be transformative and disruptive. In addition, it may reduce error rates by directing the attention of physicians to subtle areas of pathology."
The deep learning algorithm developed by Lee and colleagues as a novel application for diagnosis of AMD could lead to the development of other deep learning modules with wide reaching applications.
Lee told MD Magazine that "multi-class diagnosis would be an important next step in this area. Unlike many prior deep learning classification models, such as Google’s skin cancer classifier and the diabetic retinopathy classifier, where the classes are mutually exclusive, the retinal OCTs may have many concomitant features such as an epiretinal membrane with a full thickness macular hole in addition to age-related macular degeneration changes. It will be critical to evaluate how such a classifier performs when presented with such real-world complex cases."
More Like This >>>