Abstract: Deep learning methods outperform classical approaches, but require labelled data to be trained. One task where acquiring labeled data is expensive is fiber segmentation in 3D CT scans. Existing methods show the feasibility of deep learning in this area [1], but are usually trained with simulated data. We have designed a method to iteratively train a neural by using user input during the training procedure [2]. The amount of user input is minimized by using weighted loss functions in combination with a U-Net architecture [3][4], to force the network to focus on user corrections and areas that are assumed to be segmented correctly. Since no ground truth data is available, the results of the segmentation are evaluated with problem specific shape metrics.

[1] Instance Segmentation of Fibers: https://arxiv.org/abs/1901.01034
[2] Interactive Image Segmentation: https://arxiv.org/abs/1710.04043
[3] 3D U-Net with Sparse Annotations: https://arxiv.org/abs/1606.06650
[4] UNet Architecture: https://arxiv.org/abs/1505.04597