Machine learning-based detection of label-free cancer stem-like cell fate

Most imaging methods rely on labelling biological samples in order to provide specific and easily detectable features. However, label-free imaging is a non-invasive and non-toxic alternative that requires accurate image analysis algorithm based on cell morphology. Such analysis has to deal with a high image variability while fewer features are extractable, so far, fast analysis of label-free brightfield microscopy (LFBM) images remains a challenging task. With the development of microfabricated devices during the last decades, high throughput image generation makes it possible to use machine learning-based algorithms in order to analyse LFBM images. Fast algorithms are also crucially needed to analyze high throughput experiments. In this paper, we provide a data-driven study in order to assess the complexity of LFBM time-lapses monitoring isolated cancer stem-like cells (CSCs) fate in non-adherent conditions. We combined for the first time individual cell fate and cell state temporality analysis in a unique algorithm. Several image analysis algorithms of increasing processing capacities were tested: a classical computer vision algorithm (CCVA), a shallow learning-based algorithm (SLBA) and a deep learning-based algorithm (DLBA). We show that our optimized DLBA has by far the best accuracy compared to CCVA and SLBA, is at least as accurate as other state-of-the-art DLBAs while being faster. With this study, we demonstrate that optimizing our DLBA accordingly to the image analysis problem can overall provide better results than pretrained models. Such a fast and accurate DLBA is therefore compatible with the generation of high throughput data and opens the route for on-the-fly analysis of CSC fate from LFBM time-lapses.