Detection of Neuronal Soma from 3D Large-Scale Light-Sheet Microscopy Data

Learning spatial distribution of neurons in specific brain areas is crucial for advancing the knowledge of brain structures and functions. Thanks to the advancements brought by light-sheet fluorescence microscopy, large-scale brain tissues are nowadays available at sub-cellular resolution. This comes with the challenge of processing Tera-bytes of data in a reasonable time and with strong performances. The work described in this thesis presents a reliable and accurate two-step approach for cell detection from vast and highly variable 3D image datasets. Multiple convolutional neural network variants are implemented in order to extract truthful probability maps from raw images and to facilitate the cell localization performed by a subsequent blob detector. The efficacy and scalability to huge data of the proposed technique is demonstrated through extensive validation and application on a cohort study of whole mouse brains and the entire Broca's area of a human. The automatic detection of neuronal soma from whole mouse brains allowed for brain-wide quantitative analyses, confirming biologically accepted theories of brain activations and connectivity on fear memory experiments, but also suggesting novel neuroscientific research directions. An extensive comparison with other state-of-the-art algorithms and with stereology, the gold-standard for large-scale neuronal counts, is presented on an entire human Broca's area. Results therefore prove the adaptability and effectiveness of deep-learning approaches in a high variety of contexts, hoping to provide many life-science laboratories worldwide with a tool to advance their researches.