Speaker
Description
Automatic segmentation and detection of cells is a fundamental task in many medical fields. The extraction of numbers, densities, types, and patterns of neuronal cells is an important instrument to study the structure of the human brain [1] and is a key task in the analysis of images of biological tissue. For example, a supervised cell instance segmentation method [2] was recently used to produce a dataset revealing detailed properties of cell distributions in different cytoarchitectonic regions [3]. Such supervised deep learning approaches, however, require high-quality training data, which prevents their immediate application to data e.g. from different species or modified staining protocols. In this work we combine Spatially-Adaptive Modulation [4] with Simulated+Unsupervised (S+U) learning [5] to drastically reduce annotation demand for the detection of neuronal cell bodies in 1-micron BigBrain data. A simple algorithmic approach provides basic simulated training data that is refined by a generative model to bridge the gap between simulated and real data, while preserving annotation correctness. We apply a self-regularization term to regulate the difference between simulated and refined images. The proposed method generates images with highly realistic appearance. The workflow significantly accelerates the process of providing cell segmentation models for new data domains, as it allows models to be pretrained on large quantities of generated images without manual annotations and can achieve performance close to that of fully supervised methods. The overall instance information in generated data is reliable in that it allows cell detection models trained on this data to form useful feature descriptions regarding the objectness of certain patterns. We view this method as a valuable pretraining procedure that may be combined with the concept of fine-tuning to yield well generalizing detection and segmentation models.
References:
[1] Amunts et al. (2013), ‘BigBrain: An Ultrahigh-Resolution 3D Human Brain Model’, Science, vol. 340, no. 6139, pp. 1472–1475
[2] Upschulte et al. (2022) ‘Contour Proposal Networks for Biomedical Instance Segmentation’, Medical Image Analysis 77, 102371
[3] Dickscheid et al. (2021), ‘Layer-specific distributions of segmented cells in Area FG1 (FusG) of BigBrain’, EBRAINS, Data set, doi: 10.25493/E1FY-2C4
[4] Park et al. (2019), ‘Semantic Image Synthesis with Spatially-Adaptive Normalization’ (CVPR) 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 2332-2341
[5] Shrivastava et al. (2017), ‘Learning from Simulated and Unsupervised Images through Adversarial Training’, 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2242-2251
Acknowledgements:
HBP SGA3 (grant 945539), DFG SPP 2041, HIBALL (InterLabs-0015), CFREF (Competition 2, 2015-2016), Computing time: JARA-HPC (grant CJINM14)
