pix2pix_HUVEC_nuclei_cancer_cells_dataset

This repository contains a Pix2Pix deep learning model to generate synthetic nuclear staining from brightfield images. The model was trained on 258 paired brightfield and fluorescent microscopy images of circulating cancer cells perfused over an endothelial cell monolayer. To improve performance, the dataset was augmented computationally by a factor of 8. The model was trained over 400 epochs using a patch size of 512x512, a batch size of 1, and a vanilla GAN loss function. The final model was selected based on its performance metrics and visual fidelity when compared to ground truth images, achieving an average SSIM score of 0.755 and an LPIPS score of 0.120.

Specifications

Model: Pix2Pix for generating synthetic nuclear staining from brightfield images

Training Dataset:

Cancer Cells: 258 paired brightfield and fluorescent microscopy images

Microscope: Nikon Eclipse Ti2-E, brightfield/fluorescence microscope with a 20x objective

Data Type: Brightfield and fluorescent microscopy images

File Format: TIFF (.tif), 16-bit

Image Size: 1024 x 1022 pixels (Pixel size: 650 nm)

Training Parameters:

Epochs: 400

Patch Size: 512 x 512 pixels

Batch Size: 1

Loss Function: Vanilla GAN loss function

Model Performance:

Circulating Cancer Cells:

SSIM Score: 0.755

LPIPS Score: 0.120

Model Selection: Models were selected based on quality metric scores and visual inspection compared to ground truth images.

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