StarDist model and data for the segmentation of Yersinia enterocolitica cells in widefield images
Creators
- 1. Max Planck Institute for Terrestrial Microbiology
Description
Dataset and StarDist model for the segmentation of Yersinia enterocolitica cells
This dataset and StarDist model are part of the publication "Active downregulation of the type III secretion system at higher local cell densities promotes Yersinia replication and dissemination".
It contains the dataset that was used for training the provided StarDist model using ZeroCostDL4Mic.
Data:
Yersinia enterocolitica cells were spotted on an agarose pad (1.5% low melting agarose (Sigma-Aldrich) in minimal medium, 1% Casamino acids, 5 mM EGTA, glass depression slides (Marienfeld)). For imaging, a Deltavision Elite Optical Sectioning Microscope equipped with a UPlanSApo 100×/1.40 oil objective (Olympus) and an EDGE sCMOS_5.5 camera (Photometrics) was used. Z-stacks with 9 slices (∆z = 0.15 µm) per fluorescence channel were acquired and 5 slices were selected for network training. Images were annotated in Fiji using the Freehand selection tool, and brightlight and mask images were quartered to obtain the final dataset of 300 paired images. 260 images were used for training, while 40 images were used to test model performance.
Model:
The StarDist 2D model was trained from scratch for 100 epochs on 300 paired image patches (image dimensions: (480 x 480 px²), patch size: (480 x 480 px²)) with a batch size of 4 and a mae loss function, using the StarDist 2D ZeroCostDL4Mic notebook (v 1) (von Chamier & Laine et al., 2020). Grid parameter was set to 2 and the number of rays to 120. The model was trained with an initial learning rate of 0.0003 using a 80/20 train/test split. The dataset was augmented 4-fold by flipping and rotation.
Key python packages used include tensorflow (v 0.1.12), Keras (v2.3.1), csbdeep (v 0.7.2), numpy (v 1.21.6), cuda (v 11.1.105Build cuda_11.1.TC455_06.29190527_0). The training was accelerated using a Tesla T4 GPU.
