A task-dependent active learning method for axon segmentation with CNNs

Convolutional neural networks (CNNs) provide reliable segmentation results on biomedical images. However, they can only develop their full potential with a representative dataset. Unfortunately, a large dataset is hard to create in biomedical research, since labeling images is time consuming and requires expert knowledge. Active learning seeks to determine those images that will yield the best results, which effectively reduces labeling cost. We present an active learning method for the stepwise identification of images that should be labeled next and test this method on an axon segmentation dataset. We outperform a baseline and a state-of-the-art method.