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

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Increasing prevalence of age-related macular degeneration (AMD) and diabetic macular edema (DME) due to aging populations and diabetes epidemics.
• Intravitreal anti-vascular endothelial growth factor (anti-VEGF) injections are standard treatment.
• Optical coherence tomography (OCT) is crucial for guiding anti-VEGF therapy by visualizing retinal pathology.

Purpose of the Study:

• To explore the use of a deep transfer learning method (VGG-16) for accurate and automated classification of AMD and DME in OCT images.
• To reduce clinician labor and provide pre-diagnosis support through automated OCT image analysis.

Main Methods:

• Utilized a dataset of 207,130 retinal OCT images from multiple centers.
• Employed a deep transfer learning approach, fine-tuning the VGG-16 network pre-trained on ImageNet.
• Evaluated performance using a validation dataset of 1000 images, calculating prediction accuracy, sensitivity, specificity, and ROC.

Main Results:

• The VGG-16 based deep transfer learning model achieved superior performance in OCT image detection.
• Achieved a prediction accuracy of 98.6%, sensitivity of 97.8%, and specificity of 99.4%.
• Demonstrated an area under the receiver-operating characteristic (ROC) curve of 100%.

Conclusions:

• The deep transfer learning method based on VGG-16 is highly effective for classifying retinal OCT images, even with a relatively small dataset.
• The approach offers valuable assistant support for medical decision-making in diagnosing common retinal diseases.
• Performance is comparable to experienced human experts, indicating promising applications in automatic diagnosis.