Conference paper · Book chapter
How Few Annotations are Needed for Segmentation Using a Multi-planar U-Net?
U-Net architectures are an extremely powerful tool for segmenting 3D volumes, and the recently proposed multi-planar U-Net has reduced the computational requirement for using the U-Net architecture on three-dimensional isotropic data to a subset of two-dimensional planes. While multi-planar sampling considerably reduces the amount of training data needed, providing the required manually annotated data can still be a daunting task.
In this article, we investigate the multi-planar U-Net’s ability to learn three-dimensional structures in isotropic sampled images from sparsely annotated training samples. We extend the multi-planar U-Net with random annotations, and we present our empirical findings on two public domains, fully annotated by an expert.
Surprisingly we find that the multi-planar U-Net on average outperforms the 3D U-Net in most cases in terms of dice, sensitivity, and specificity and that similar performance from the multi-planar unit can be obtained from half the number of annotations by doubling the number of automatically generated training planes.
Thus, sometimes less is more!
Language: | English |
---|---|
Publisher: | Springer |
Year: | 2021 |
Pages: | 209-216 |
Proceedings: | MICCAI Workshop on Deep Generative Models |
Series: | Lecture Notes in Computer Science (including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
ISBN: | 3030882098 , 3030882101 , 9783030882099 and 9783030882105 |
ISSN: | 03029743 and 16113349 |
Types: | Conference paper and Book chapter |
DOI: | 10.1007/978-3-030-88210-5_20 |
ORCIDs: | 0000-0002-0358-4692 , 0000-0003-1261-6702 and Laprade, William Michael |