Weakly Supervised Fluid Filled Region Localization In Retinal Oct Scans

Retinal Optical Coherence Tomography (OCT) scans are an important diagnostic tool for ophthalmologists. These scans provide a cross-sectional view of the retina for ophthalmologists to detect abnormalities. A common type of abnormality found in these scans is a Fluid Filled Region (FFR). In this paper, we present a method to simultaneously classify and localize FFRs within retinal OCT scans using a specialized Convolutional Neural Network (CNN). The training data is weakly labeled, with only an indication of whether a scan contains FFRs or not. We compare different architectures to see which ones give us the best localization and classification metrics. We have found that architectures using Dense Blocks and Scaled Exponential Unit (SeLU) activations give us the best localizations with a Mean Average Precision (mAP) of 0.75 on true positive images and a classification accuracy of 94.8%.