Fully automated detection of retinal disorders by image-based deep learning

Purpose With the aging population and the global diabetes epidemic, the prevalence of age-related macular degeneration (AMD) and diabetic macular edema (DME) diseases which are the leading causes of blindness is further increasing. Intravitreal injections with anti-vascular endothelial growth factor (anti-VEGF) medications are the standard of care for their indications. Optical coherence tomography (OCT), as a noninvasive imaging modality, plays a major part in guiding the administration of anti-VEGF therapy by providing detailed cross-sectional scans of the retina pathology. Fully automating OCT image detection can significantly decrease the tedious clinician labor and obtain a faithful pre-diagnosis from the analysis of the structural elements of the retina. Thereby, we explore the use of deep transfer learning method based on the visual geometry group 16 (VGG-16) network for classifying AMD and DME in OCT images accurately and automatically. Method A total of 207,130 retinal OCT images between 2013 and 2017 were selected from retrospective cohorts of 5319 adult patients from the Shiley Eye Institute of the University of California San Diego, the California Retinal Research Foundation, Medical Center Ophthalmology Associates, the Shanghai First People’s Hospital, and the Beijing Tongren Eye Center, with 109,312 images (37,456 with choroidal neovascularization, 11,599 with diabetic macular edema, 8867 with drusen, and 51,390 normal) for the experiment. After images preprocessing, 1000 images (250 images from each category) from 633 patients were selected as validation dataset while the rest images from another 4686 patients were used as training dataset. We used deep transfer learning method to fine-tune the VGG-16 network pre-trained on the ImageNet dataset, and evaluated its performance on the validation dataset. Then, prediction accuracy, sensitivity, specificity, and receiver-operating characteristic (ROC) were calculated. Results Experimental results proved that the proposed approach had manifested superior performance in retinal OCT images detection, which achieved a prediction accuracy of 98.6%, with a sensitivity of 97.8%, a specificity of 99.4%, and introduced an area under the ROC curve of 100%. Conclusion Deep transfer learning method based on the VGG-16 network shows significant effectiveness on classification of retinal OCT images with a relatively small dataset, which can provide assistant support for medical decision-making. Moreover, the performance of the proposed approach is comparable to that of human experts with significant clinical experience. Thereby, it will find promising applications in an automatic diagnosis and classification of common retinal diseases.