Data Set for 'Self-Supervised Machine Learning for Live Cell Imagery Segmentation'

Self-supervised machine learning code and data for segmenting live cell imagery (Matlab) Running the Code SSL_Demo_2.m : main program for self-supervised machine learning segmentation SSL_Declumping_2.m : main program for declumping application (applied to output of SSL_Demo_2.m) This Matlab code is designed to be used with time-resolved live cell microscopy images (tiffs) for the automated segmentation of cells from background. It is recommended you first run this code with its accompanying demo data (included in this package), keeping the current directory structure. Simply open SSL_Demo_2.m or SSL_Declumping_2.m in Matlab and hit Run. Code Methodology The principle of self-supervised machine learning is that you simply load your images and Run - no parameter tuning needed, no training imagery required. Run from start to finish, the SSL_Demo_2.m code uses consecutive pairs of images to generate training data of 'cells' and 'background' via dynamic feature vectors based on optical flow (unsupervised). These self-labeled pixels are then used to generate static feature vectors (entropy, gradient), which in turn are used to train a classifier model. The training data is updated every image in order to automatically adapt to temporal changes in cell morphologies or background illumination. The code was tested for high fidelity segmentation using five different modes of light microscopy: transmitted light, DIC, phase contrast, fluorescence and interference reflection microscopy. Six different cell lines were imaged to cover a range of morphologies and phenotypic dynamics using three cameras of differing resolutions. The associated manuscript for this work can be found here (although the latest version is under peer review as of this writing): https://www.biorxiv.org/content/10.1101/2021.01.07.425773v1 This code was tested on Matlab v2020a and v2021a using commercially available laptop computers running the Windows 10 operating system.