Semi-automatic Annotation of OCT Images for CNN Training

摘要

Annotating image data is one of the most time-consuming parts of the training of machine learning algorithms. With this contribution, we are looking for a solution that decreases the time needed for annotating images of the human retina created by Optical coherence tomography (OCT). As a first step, we use a simple annotation tool to test whether the sorting of images by their predicted amount of parts that contain anomalies decreases the time needed for annotation without increasing the number of annotation mistakes. The predictions are made by a convolutional neural network (CNN) that was trained on a previously annotated image set. We investigated the annotation behaviour in two groups of five subjects each. The first group received the (OCT) images in the order of recording, the second group sorted by the number of predicted anomalies. We observed a significant increase in annotation speed in the subjects of the second group while the quality of annotation remained at least stable.