Quantifying Topology In Pancreatic Tubular Networks From Live Imaging 3D Microscopy

• DOI: 10.48550/arxiv.2105.09737
• Publication Date: 2021-01-01
• Authors: Kasra Arnavaz; Oswin Krause; Kilian Zepf; Jelena M. Krivokapic; Silja Heilmann; Jakob Andreas Bær...; Pia Nyeng; Aasa Feragen
• Keywords: Ground truth

Motivated by the challenging segmentation task of pancreatic tubular networks, this paper tackles two commonly encountered problems in biomedical imaging: Topological consistency segmentation, and expensive or difficult annotation. Our contributions are following: a) We propose a topological score which measures both geometric between predicted ground truth segmentations, applied to model selection validation. b) provide full deep-learning methodology for noisy on time-series image data. In our method, we first use semisupervised U-net architecture, applicable generic tasks, jointly trains an autoencoder network. then tracking loops over time further improve topology. This semi-supervised approach allows us utilize unannotated data learn feature representations that generalize test with high variability, spite annotated training having very limited variation. validated task, locating structures fetal pancreas from live imaging confocal microscopy. show outperforms not only fully supervised pre-trained models but also takes into account during training. Further, achieves mean loop 0.808 detecting pancreas, compared trained clDice 0.762.