NLM IRP Seminar Schedule

Abstract:

Age-related macular degeneration (AMD) is one of the leading causes of vision loss in the elderly. Late-stage AMD can develop into atrophic or neovascular AMD where atrophic AMD is the most common form. Geographic atrophy (GA) is the primary lesion in late atrophic AMD and is usually accompanied by very poor central vision. GA is predicted to affect more than five million people worldwide. Fast and accurate identification of eyes with GA could lead to improved management of the disease. GA can be detected on 2D imaging modalities, e.g., fundus autofluorescence (FAF) images and color fundus photographs (CFP). However, these modalities provide no details about the underlying layer structures and how they change. In this work, we use optical coherence tomography (OCT) volumetric scans for the detection task. OCT scans have the advantages of being easily available and providing volumetric context. For this purpose, we developed 3D convolutional neural networks (CNNs) as well as 2D CNNs. A dataset of 1,284 SD OCT scans from 311 participants was used to train networks, where cross-validation was used for evaluation with each testing set, containing no participant from the corresponding training set. En-face heatmaps and important regions at the B-scan level were used to visualize the outputs of networks, and three ophthalmologists graded the presence or absence of GA in them to assess the explainability of its detections. Compared to other networks, 3D CNNs achieved the best metrics, with the best accuracy of 0.93, AUC of 0.94, and APR of 0.91, and received the best gradings, of 0.98 and 0.68 on the en face heatmap and B-scan grading tasks, respectively.