Probabilistic Image Diversification to Improve Segmentation in 3D Microscopy Image Data

Dennis Eschweiler, Justus Schock, Johannes Stegmaier

Abstract: The lack of fully-annotated data sets is one of the major limiting factors in the application of learning-based segmentation approaches for microscopy image data. Especially for 3D image data, generation of such annotations remains a challenge, increasing the demand for approaches making most out of existing annotations. We propose a probabilistic approach to increase image data diversity in small annotated data sets without further cost, to improve and evaluate segmentation approaches and ultimately contribute to an increased efficacy of available annotations. Different experiments show utilization for benchmarking, image data augmentation and test-time augmentation on the example of a deep learning-based 3D segmentation approach. Code is publicly available at https://doi.org/https://github.com/stegmaierj/ImageDiversification.