A Root Foundation Model for Zero-Shot Segmentation

Foundation models pre-trained on massive datasets have demonstrated impressive performance, but in some specialised domains have been found to have lower accuracy. Domain-specific foundation models target a particular domain such as retinal or plant images. These domain-specific models have shown inconsistent results and the benefit to root segmentation is unknown. We train and evaluate the first domainspecific foundation model for root segmentation. Evaluation uses a leave-one-dataset-out design across nine diverse root datasets with two architectures. Applied zero-shot to unseen datasets, the root foundation model achieves 92% of fine-tuned Dice on average (0.636 versus 0.698), with 5 of 9 datasets above 90%. With 10 patches of few-shot fine-tuning, the root foundation model recovers 95% of its full-data Dice on average, versus 69% for a general pre-trained model. At low patch counts the general pre-trained model often failed to converge, with 5 of 9 datasets giving Dice below 0.05 at 3 patches, while the root foundation model produced Dice above 0.47 on every dataset and patch count. With full target-data fine-tuning, the two perform comparably, with mean improvements of +0.011 Dice for MobileSAM and +0.022 for M2F Swin-S, neither significant (Wilcoxon p = 0.150 and 0.064). We release our pre-trained MobileSAM root foundation model for use with RootPainter, enabling fully automatic root segmentation on new datasets with an ordinary laptop or desktop computer, with no need for annotation or training.