Abstract We present the Light My Cells Database, a large-scale open-access collection comprising 2,574 acquisition sets and 56,984 microscopy 2D images designed to support the development of machine learning models for fluorescence prediction from transmitted light images. The dataset aggregates data from 30 independent studies conducted across 8 national imaging centers and captures a wide diversity of biological samples, imaging modalities, and acquisition systems. Each transmitted light image - recorded in bright-field, phase contrast, or differential interference contrast -is paired with at least one fluorescence image labeling key subcellular structures: nucleus, mitochondria, tubulin, or actin. All images are standardized in OME-TIFF format and annotated with rich metadata following REMBI guidelines. A dedicated preprocessing pipeline ensures dimensional harmonization, best-focus plane selection, and consistent file naming. The database reflects the variability encountered in real-life microscopy experiments, making it suited for training and benchmarking generalizable deep learning models. It is accessible via the BioImage Archive and supports a range of downstream applications, including in silico labeling, segmentation, and cell profiling from label-free imaging.
Kauffmann et al. (Tue,) studied this question.