Fig. 1.

Overview of proposed method. Starting from a stack of 2D images with a minimum subset of heterogeneously manually labelled structures, we first extract statistics about the annotation effort and we train a segmentation network. We then predict the segmentation accuracy and combine it with the annotation effort prediction to formulate a query. Once a new structure is delineated, the statistics about the annotation effort is updated and the segmentation network is fine-tuned. The algorithm iterates until a target accuracy is reached.