During embryonic development, the tubular network in the pancreas remodels from a web-like structure to a tree-like structure. Biologists have hypothesized that this morphological change might have an influence on the appearance of nearby beta-cells which produce insulin.

For a viable test of the biological hypothesis, we would need reliable and efficient segmentations. While manual segmentations of human experts are in general reliable, they are not efficient. This makes the manual approach unattainable for a scalable study of several movies. On the other hand, deep learning segmentation models are efficient, taking a short time to make new predictions. However, they pose three challenges for our application:

1. Their high performance is predicated on having large amounts of labeled data.

2. Their predictions are not always reliable.

3. They are usually trained for pixel-wise performance.

We investigate two solutions for costly annotations: 1. active learning which makes efficient use of annotation budget and 2. semisupervised segmentation which uses unannotated data to improve segmentation performance.

Endowing segmentation models with uncertainty estimates can be an effective way to bring a sense of reliability into the predictions. We study established probabilistic segmentation models to find whether their uncertainty estimates correlate with segmentation error and whether these uncertainties are of use for active learning.

For our biological application, extracting topologically accurate structures is more important than pixel-wise accuracy. To that end, we derive a highly interpretable topological score function which can be used to encourage topological accuracy through model selection.