Abstract
<jats:p>Root anatomical barriers, including the suberized and lignified walls of the endodermis and exodermis, and cortical aerenchyma, regulate water and nutrient transport, gas exchange, and rhizosphere interaction. Their adaptive function places them as an important target for breeding environmentally resilient plant species. Quantifying these structures at high resolution is a manual bottleneck that limits experimental scale. We present RADIX (Root Anatomy Deep-learning Image segmentation across species and platforms), a framework that adapts a large self-supervised vision-transformer foundation encoder (DINOv3), pre-trained on billions of natural images, to root anatomy by fine-tuning its encoder with a dense-prediction-transformer decoder. Transferring these general-purpose vision encoders to a specialized biological domain with a high-quality annotated dataset is what allows RADIX to generalize across species and imaging platforms. We train and evaluate it on the first expert-annotated benchmark of root anatomical structures at scale, comprising 1,695 high-quality fluorescence images spanning 17 monocot and dicot species, six anatomical structures, and three imaging platforms. RADIX segments all six structures at inter-annotator-level accuracy and generalizes to unseen species, genotypes, growth conditions, and an imaging platform from an independent laboratory. A single unified model surpasses monocot- and dicot-specialist models without sacrificing in-group accuracy. Predicted masks yield aerenchyma and suberin/lignin measurements matching expert annotation at ~1.2 s per image with a single GPU, reducing weeks of manual analysis to minutes. Applying RADIX across genotypes, microbial treatments, and growth systems, we show that these cell type features form a coordinated, multidimensional, and context-dependent system shaped by genetic and environmental factors.</jats:p>