Machine learning quantitative cellular-level wood characterization with volumetric micro- and submicro-CT

This study investigates the integration of volumetric micro- and sub-micro-CT (computed tomography) imaging with quantitative cellular-level wood characterization using transformer and U-Net deep learning models within the scope of a proof of concept study. The custom dataset, captured using two imaging setups targeting micro- and sub-micrometer resolutions, includes sycamore maple and Scots pine samples. Leveraging pretrained transformer-based foundation models and U-Net architectures, we developed a fine-tuned processing pipeline capable of extracting discriminative anatomical features at species and intra-species levels in under 15 seconds per sample.