Dynamic Deep Networks for Retinal Vessel Segmentation

Abstract

Deep learning has recently yielded impressive gains in retinal vessel segmentation. However, state-of-the-art methods tend to be conservative, favoring precision over recall. Thus, they tend to under-segment faint vessels, underestimate the width of thicker vessels, or even miss entire vessels. To address this limitation, we propose a stochastic training scheme for deep neural networks that robustly balances precision and recall. First, we train our deep networks with dynamic class weights in the loss function that fluctuate during each training iteration. This stochastic approach–which we believe is applicable to many other machine learning problems–forces the network to learn a balanced classification. Second, we decouple the segmentation process into two steps. In the first half of our pipeline, we estimate the likelihood of every pixel and then use these likelihoods to segment pixels that are clearly vessel or background. In the latter part of our pipeline, we use a second network to classify the ambiguous regions in the image. Our proposed method obtained state-of-the-art results on five retinal datasets—DRIVE, STARE, CHASE-DB, AV-WIDE, and VEVIO—by learning a robust balance between false positive and false negative rates. Our novel training paradigm makes a neural network more robust to inter-sample differences in class ratios, which we believe will prove particularly effective for settings with sparse training data, such as medical image analysis. In addition, we are the first to report segmentation results on the AV-WIDE dataset, and we have made the ground-truth annotations for this dataset publicly available. An implementation of this work can be found at https://github.com/sraashis/deepdyn.