Studies of Plausibility Using Deep Learning Techniques to Automatically Detect Retinal Neovascularization in Enface Optical Coherence Tomography Angiography Images

Retinal neovascularization (RNV) is a pathological condition characterized by the abnormal growth of new blood vessels within the retina, commonly found in patients with diabetic retinopathy (DR). Early detection of RNV is essential for preventing severe vision loss in DR patients. RNV can be detected using optical coherence tomography angiography (OCTA), an advanced imaging technology that visualizes blood flow within retinal and choroidal vessels without penetrating tissue layers. Automated RNV detection in OCTA is crucial for accelerating the prescreening process in DR management. However, detecting RNV is challenging due to the complexity of OCTA vessels, the presence of noise in OCTA images, and the significant variability in RNV patterns. Additionally, the limited availability of OCTA images further complicates the detection process. In this study, we investigated the feasibility of utilizing advanced deep learning methods to automatically classify OCTA images into two categories: images with RNV and images without RNV. We experimented with five modern deep learning architecture models: DenseNet121, Inception V3, EfficientNet B2, and MobileNet V3 and VNet adapted for detection, using a collection of 1,183 images, which included only 73 RNV images, reflecting the real-world rarity of RNV compared to other conditions. Among the models, Inception V3 achieved the best performance, with an accuracy of 97.79% and the lowest false negative rate (FNR) and false positive rate (FPR) of 6.67% and 2.14%, respectively. Our findings demonstrate that, by using appropriate deep learning architectures and parameters, we can effectively identify OCTA images with retinal neovascularization (RNV) from those without RNV with high accuracy, even when working with a small number of RNV images.