AI platform for microscopy image restoration and virtual staining

人工智能显微镜图像恢复和虚拟染色平台

AI Platform for Microscopy Image Restoration and Virtual Staining Project Summary: Fluorescence microscopy has enabled many major discoveries in biomedical sciences. Despite the rapid advancements in optics, lasers, probes, cameras and novel techniques, major factors such as spatial and temporal resolution, light exposure, signal-to-noise, depth of penetration and probe spectra continue to limit the types of experiments that are possible. Deep learning (DL) algorithms are well suited for image-based problems like SNR/super-resolution restoration and virtual staining, which have great enabling potentials for microscopy experiments. Previously impossible experiments could be realized such as achieving high signal-to-noise and/or spatial-temporal resolution without photobleaching/phototoxicity; simultaneously observing many image channels without interfering with native processes, etc. This could pave the way for a quantum leap forward in microscopy-based discoveries that elucidate biological functions and the mechanisms of disorders, and enable new diagnostics and therapies for human diseases. However, these new methods have not been widely translated to new microscopy experiments. The delay is due to several practical hurdles and challenges such as required expertise, computing and trust. In order to accelerate the adoption of DL in microscopy, novel AI platform tailored for biologists are needed for training, applying and validating DL models and outputs. The present project aims to develop an AI platform for microscopy image restoration and virtual staining called AI for Restoring and Staining (AIRS) platform. With our collaborator, Dr. Hari Shroff (National Institute of Biomedical Imaging and Bioengineering) we have successfully created DL models for SNR restoration, super-resolution restoration and virtual staining for a variety of imaging conditions and organelles in our preliminary studies. The AIRS platform intends to (1)provide a comprehensive suite of validated DL models for microscopy restoration and virtual staining applications including SNR restoration, super-resolution restoration, spatial deconvolution, spectral unmixing, prediction of 3d from 2d images, organelle virtual staining and analysis; (2)provide plug and play for common microscopy experiments; (3)provide semi-automatic update training to tailor DL models to match advanced microscopy experiments; (4)provide user friendly support for new DL model training for pioneering microscopy experiments; (5)provide confidence scores to assess the output results by a DL model, (6) provide DL models that avoid image artifact (hallucination) and allow continuous learning and evolution; (7) and be able to access the required computing infrastructure and database connection.