ISOODL: Instance segmentation of overlapping biological objects using deep learning

Image segmentation is an important first step for the quantitative analysis of biomedical images. We present a method to simultaneously segment and classify translucent overlapping objects in 2D images. For this we propose an approach using a fully-convolutional neural network simultaneously solving two tasks: object detection and instance segmentation. Object detection predicts reference points, object class labels and sizes. To solve the problem of multiple labels per location, we lift our label-space from 2D to 3D, resulting in a non-overlapping representation of the instance masks. To our knowledge it is the first method that handles overlapping biological objects using deep learning making it easily applicable to a large variety of challenging datasets.