Attention-based 3D convolutional networks for detection of geographic atrophy from optical coherence tomography scans

Abstract

Geographic atrophy (GA) is the defining lesion of advanced atrophic age-related macular degeneration (AMD). GA can be detected and characterized most accurately using spectral-domain optical coherence tomography (SDOCT), which provides detailed 3D information about changes in multiple retinal layers. Existing methods are limited to 2D convolutional neural networks (CNNs). Therefore, they do not capture the 3D context between adjacent 2D slices of the OCT scan and also require a large inference time. We propose 3D CNNs with 3D attention mechanisms for the automated detection of GA on SDOCT scans using scan-level labels. The best network achieved an accuracy of 88%, and its visualizations suggest the interpretability of its predictions.

Detection of reticular pseudodrusen on optical coherence tomography images

“…For the pooling layer, we used a global average pooling layer. We compared the proposed networks to a baseline shared 2D CNN (Base2D) with average bagging, based on the GA detection work by Elsawy et al 9 We also compared the proposed networks to a baseline 3D network (Base3D), based on the detection work by Elsawy et al, 10 where their network was trained on B-scans of size 128 × 128.…”

“…Making these distinctions are important, since RPD and drusen carry very different risk profiles for progression to late AMD. In this study, motivated by our recent work on detecting GA from volumetric SD-OCT scans using 3D convolutional neural networks (CNN), 9,10 we propose a 3D classification network to detect RPD from volumetric OCT scans using binary labels at the scan level. However, there were many challenges, which included creating ground truth labels and resizing the SD-OCT scan to a size that can preserve the subtle RPD features.…”