Lefebvre, Joël
(2023).
« Exploring CNN-based self-supervised illumination inhomogeneity compensation for Serial Optical Coherence Tomography » (2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), Cartagena de Indias, Colombie, 18-21 avril 2023)
Fichier(s) associé(s) à ce document :
Résumé
Serial blockface histology is a 3D imaging modality that combines a vibratome with a microscope. Whole samples are acquired by sequentially removing small tissue layers with the vibrating blade and by generating a mosaic of several images of the revealed tissue which can be assembled to obtain a 3D representation of the sample at a high resolution. Due to many factors, the acquired mosaic tiles can be affected by complex illumination inhomogeneity that negatively affects the data reconstruction and analysis. Here, we propose a convolutional neural network approach to estimate and compensate the illumination inhomogeneity. The model is trained with simulated vignettes without using illumination ground truth, which is many times harder or even impossible to obtain. Using a small multiresolution dataset consisting in serial OCT images from whole mouse brains, we show that our proposed approach has many advantages compared to an unsupervised a posteriori illumination compensation method.
RECHERCHER
PARCOURIR
LIBRE ACCÈS
