Deep Learning Based Cell Segmentation of 3D Confocal Images of Wood for Enzymatic Hydrolysis Applications

Abstract

Plant cell walls are an abundant renewable resource for biobased products, but their recalcitrance limits enzymatic hydrolysis efficiency. Cell and tissue morphological parameters provide reliable biomarkers for predicting enzymatic hydrolysis efficiency. Accurate cell segmentation in 3D microscopy images is essential for quantifying these parameters. Although deep learning has achieved high performance in biological image segmentation, confocal microscopy's limited penetration depth leaves elongated wood cells partially imaged, creating open-ended structures that challenge existing segmentation methods. Based on an artificial longitudinal domain restriction, an adapted framework is presented here for the training of a 3D U-net model for the cell segmentation of wood images. Quantitative evaluation on a dedicated benchmark dataset identified a model configuration that reduced segmentation errors and produced more robust predictions during early enzymatic hydrolysis. These results support the use of the framework to quantify morphological parameters relevant to assessing and predicting biomass conversion efficiency.