Post-processing of light sheet fluorescence microscope images using auto-encoders and Richardson-Lucy deconvolution
Résumé
Light Sheet Fluorescence Microscopy (LSFM) is a powerful tool for neurobiologists, allowing high-quality fast volumetric imaging. However, the used imaging technique induces some artifacts and distortions on the reconstructed 3D volumes. Thus enhancing LSFM images before 3D reconstructions is a crucial step for high-quality reconstructions. In this study, we proposed a pipeline for enhancing the image quality slice by slice. To achieve this, we proposed a novel approach using a pipeline of three steps. We started by implementing a deblurring algorithm based on the work of [1] [2], followed by an automatic contrast enhancement. Then, in order to remove the noise accentuated by this contrast enhancement, we developed a convolutional denoising auto-encoder using skip-connections, providing outstanding results on mixed Poisson-Gaussian noise. Finaly, we addressed the issue of axial distortion occuring on LSFM. We proposed a novel approach based on an auto-encoder trained on bead calibration images. Our proposed pipeline provides a comprehensive solution with promising results surpassing existing methods for improving the quality of LSFM images, which can boost the interpretation of biological data.