Deep learning algorithms have been applied to produce highly accurate systems that can detect various eye conditions from fundus images,1,2 as well as optical coherence tomography (OCT) scans.3,4 Both clinically and in research settings, OCT measurements of the inner layers of the retina (including the ganglion cell layer) are used to quantify glaucomatous damage of the macula ‐ damage that may detectable via OCT in the early stages of the disease.5
In the setting of sufficient data, DLSs are more accurate than more classical techniques
Asaoka and colleagues developed a deep learning system (DLS) to distinguish early-onset glaucoma from normal eyes using macular OCT scans.Their main finding was that a DLS that uses an 8 x 8 grid macular retinal nerve fiber layer (RNFL) thickness and ganglion cell complex (GCC) layer thickness from OCT can achieve an area under receiving-operating characteristic curve (AUC) of 93.0%. They further validated that the DLS performed better than two traditional machine learning techniques, the support vector machine and the random forest. These findings validate a generally-accepted sentiment in the machine learning community6 for medical imaging: in the setting of sufficient data, DLSs are more accurate than more classical techniques.
The major limitations of the study are the size of the data set and the exclusion of difficult cases. A paper describing a DLS for the detection of DR from fundus photos found that the minimum number of images required for development was over 50,000.1 Asaoka and colleagues had a 'pre-training' data set of 4,316 OCTs, while the test set consisted of only 114 patients with early open-angle glaucoma and 82 normal patients. Thus, we postulate that the DLS developed by the authors could potentially be improved further with additional training data. In addition, the authors excluded difficult images, such as ones with tilted discs, from the training and test sets. Because such images are not uncommon in routine clinical workflows, exclusion criteria such as this may limit the ability to extrapolate the model's performance to the general clinical setting.
To conclude, Asaoka et al. have made solid initial steps to applying DLS to OCTs, which contain crucial information for diagnosing and managing glaucoma. Additional work will be needed to further validate their findings on larger datasets of a clinically diverse patient population and breadth of images.