StarDist_HUVEC_nuclei_dataset

This repository contains a StarDist deep learning model and its training and validation datasets for segmenting endothelial nuclei while ignoring cancer cells. The cancer cells were perfused over an endothelial cell monolayer. The initial dataset consisted of 17 images, where cancer cell nuclei were manually removed after segmentation with the StarDist Versatile Nuclei model. This dataset was augmented to 68 paired images using computational techniques like rotation and flipping. The model was trained for 200 epochs, achieving an average F1 Score of 0.976, demonstrating high accuracy in segmenting endothelial nuclei while excluding cancer cells.

Specifications

Model: StarDist for segmenting endothelial nuclei while ignoring cancer cells

Training Dataset:

Number of Original Images: 17 paired predictions of nuclei and label images

Augmented Dataset: Expanded to 68 paired images using rotation and flipping

Source Image Generation: Generated using a pix2pix model trained to predict nuclei from brightfield images of cancer cells on top of an endothelium (DOI: 10.5281/zenodo.10617532)

Target Image Generation: Masks obtained via manual segmentation

File Format: TIFF (.tif)

Brightfield Images: 8-bit

Masks: 8-bit

Image Size: 1024 x 1022 pixels (uncalibrated)

Training Parameters:

Epochs: 200

Patch Size: 1024 x 1024 pixels

Batch Size: 2

Performance:

Average F1 Score: 0.976

Average IoU: 0.927

Model Training: Conducted using ZeroCostDL4Mic (https://github.com/HenriquesLab/ZeroCostDL4Mic/wiki)

Reference

Fast label-free live imaging reveals key roles of flow dynamics and CD44-HA interaction in cancer cell arrest on endothelial monolayers

Gautier Follain, Sujan Ghimire, Joanna W. Pylvänäinen, Monika Vaitkevičiūtė, Diana Wurzinger, Camilo Guzmán, James RW Conway, Michal Dibus, Sanna Oikari, Kirsi Rilla, Marko Salmi, Johanna Ivaska, Guillaume Jacquemet

bioRxiv 2024.09.30.615654; doi: https://doi.org/10.1101/2024.09.30.615654