CelFDrive: Multimodal deep learning-assisted microscopy for automated detection of rare events

The emergence of automated microscopy has enabled the collection of large datasets of rare biological events, accelerating discovery in biology. However, most existing methods cannot be readily integrated into microscopy set-ups or are limited to a single microscope type. We present CelFDrive, a software package that automatically detects rare events of interest and automates high-resolution 3D imaging of target cells by integrating deep-learning cell classification of auxiliary low-magnification fluorescence images. We show that CelFDrive reduced the mean time required to nominate a mitotic prophase candidate for high-resolution light-sheet imaging by over 30-fold relative to manual expert selection. This approach can be used to define the order of kinetochore assembly upon nuclear envelope breakdown. The trained CelFDrive detector demonstrated transferability by retaining mitotic-cell detection capability in an independently generated dataset using a different cell line, DNA marker and imaging setup. As datasets grow and models improve, CelFDrive offers a clear path toward intelligent imaging systems that not only accelerate data collection but also fundamentally enhance how biological questions are explored.