DeepSTAPLE: Learning to predict multimodal registration quality for unsupervised domain adaptation

Abstract

While deep neural networks often achieve outstanding results on semantic segmentation tasks within a dataset domain, performance can drop significantly when predicting domain-shifted input data. Multi-atlas segmentation utilizes multiple available sample annotations which are deformed and propagated to the target domain via multimodal image registration and fused to a consensus label afterwards but subsequent network training with the registered data may not yield optimal results due to registration errors. In this work, we propose to extend a curriculum learning approach with additional regularization and fixed weighting to train a semantic segmentation model along with data parameters representing the atlas confidence. Using these adjustments we can show that registration quality information can be extracted out of a semantic segmentation model and further be used to create label consensi when using a straightforward weighting scheme. Comparing our results to the STAPLE method, we find that our consensi are not only a better approximation of the oracle-label regarding Dice score but also improve subsequent network training results.

Tasks

Reproductions