Adversarial Training with Multiscale Boundary-Prediction DNN for Robust Topologically-Constrained Segmentation in OOD Images.

For image segmentation, typical deep-neural-network (DNN) methods fail to enforce topology constraints/guarantees on the underlying set of objects. Moreover, at the time of deployment, in the presence of out-of-distribution (OOD) images with degradations, typical DNNs perform poorly. We propose a novel DNN framework for segmenting an image comprising multiple objects that exhibit specific topological properties individually as well as jointly. We design our DNN to predict topology-constrained object boundaries at multiple scales, incorporating per-object and inter-object topology constraints. Our DNN architecture and formulation makes the learning more robust to OOD images. To further improve robustness to OOD images, we propose a novel adversarial training formulation for our DNN. Results on two publicly available datasets show our DNNs (with/without adversarial learning) outperform existing methods on OOD images.