2D Label-free Prediction of Multiple Organelles Across Different Transmitted-light Microscopy Images with Bag-of-Experts

Abstract

Label-free prediction has emerged as a significant application of artificial intelligence (AI) in the field of bioimaging, which aims to predict the localization of specific organelles directly from readily-accessible transmitted-light images, thereby alleviating the need for acquiring fluorescent images. Despite the existence of numerous research, in practice, the high variability in imaging conditions, modalities, and resolutions poses a challenge to the final prediction. In this study, we propose a "Bag-of-Experts" strategy, targeting at different organelles, with self-supervised pre-training. The comprehensive experimentation showcases that our model is agnostic to the transmitted-light image modalities and the imaging conditions, to certain extent, indicating considerable generalizability. The code is released at: https://github.com/MMV-Lab/LightMyCells