Single2Ensemble: An Adapter Network for Segmentation Quality Estimation

Because deep models do not give any signs when they fail, it is necessary to monitor the quality of their outputs with a quality estimation tool before using their results in clinical practice. Recent studies on segmentation quality estimation have extracted quality related information from agreements of ensemble outputs. Despite the simplicity of the approach, it is very costly or impossible to train multiple copies of the same network. To tackle the issue, we present Single2Ensemble, which can transform any segmentation model to an ensemble, with one additional model training, in contrast to many. We propose an implicit ensemble design inspired by Mask2Former, using query ensembles. We validate Single2Ensemble on left atrium segmentation using Stacom 2013. We found it outperforms other ensembles on segmentation performance and the quality of its outputs can be assessed with a quality estimation method, with high accuracy.