- Reviews Visibility: The authors agree that reviews are made publicly visible, if the submission is accepted.
- Abstract: Recent microscopy imaging techniques allow to precisely analyze cell morphology in 3D image data. The arising vast amount of image data and the varying image quality render manual investigations a challenging task, to which end manual annotations of such data is cumbersome or even infeasible. The resulting lack of annotated data not only impedes the successful training procedure of machine learning approaches, but also complicates the comparability of publicly available methods. This can be compensated by image synthesis approaches, which require an exact modeling of the underlying cellular structures. The spherical structure of cells in the 3D space can be realistically modeled by the application of spherical harmonics, which offer a way to directly constrain segmentations to represent seamless spherical shapes. This work proposes how this representation of spherical objects can be utilized to model and to predict cellular structures in 3D microscopy image data. We incorporate those descriptors into an image synthesis pipeline and, thereby, demonstrate a way to generate synthetic 3D image data that can be utilized to replace manually obtained ground truth annotations for training of segmentation approaches. Moreover, we propose a network architecture, which is designed to predict spherical coefficients to obtain immanently shape-constrained segmentations and compare the results to the results obtained by established methods used in the field.
- Supplementary Material: zip
- Keywords: 3D Cell Analysis, Spherical Harmonics, Annotation-Free, Shape-Constrained