Séminaire I&M, par Adrien JULIA

Adrien JULIA

Titre : 3D Segmentation of Mouse Brain Cell Nuclei Imaged Using Light Sheet Microscopy

Résumé :

Medical image segmentation has been a widely studied field in the computer vision community. Several methods have been developed using both conventional and deep learning approaches. Among deep learning methods, U-Net is still one of the most used architectures, based on the use of a convolutional auto-encoder. Several modalities need to be tackled when dealing with biological image processing. We focus our study on two important aspects: the need for annotated datasets to train a neural network, and the need to perform 3D segmentation.

Light Sheet Fluorescence Microscopy provides high-quality images of neural circuitry, enabling the scanning of an entire mouse brain quickly. Analyzing the cell nuclei and the whole neural structure is a challenging task, involving the automatic extraction of both cell nuclei and neurons. Neurons can have a wide variety of shapes and are better captured using 3D imaging. Cell nuclei segmentation in the case of wide-field microscopy is still a challenging task, as the size of nuclei can sometimes be only a couple of pixels. These nuclei can also aggregate, making it even more difficult to differentiate them.