Multidataset Incremental Training for Optic Disc Segmentation

Abstract

When convolutional neural networks are applied to image segmentation results depend greatly on the data sets used to train the networks. Cloud providers support multi GPU and TPU virtual machines making the idea of cloud-based segmentation as service attractive. In this paper we study the problem of building a segmentation service, where images would come from different acquisition instruments, by training a generalized U-Net with images from a single or several datasets. We also study the possibility of training with a single instrument and perform quick retrains when more data is available. As our example we perform segmentation of Optic Disc in fundus images which is useful for glau coma diagnosis. We use two publicly available data sets (RIM-One V3, DRISHTI) for individual, mixed or incremental training. We show that multidataset or incremental training can produce results that are simi lar to those published by researchers who use the same dataset for both training and validation.