Diffusion-Based Synthetic Pretraining for Root Segmentation in Rhizotron Images

Abstract

Rhizotrons allow root system architecture to be studied non-destructively in soil. However, automated analysis of rhizotron images remains difficult: roots have low contrast against soil, soil particles resemble fine roots, and deep learning models need many manually annotated images that are slow and subjective to produce. Synthetic images could provide paired images and annotations at scale, but rhizotron images are largely absent from the data used to train generative models. We built a pipeline combining Low-Rank Adaptation (LoRA), ControlNet and inpainting to generate synthetic rhizotron images with paired root masks. We compared training on synthetic data only, real data only, and synthetic pretraining followed by fine-tuning on real images, using three segmentation models: nnU-Net v2, YOLO26 and Mask2Former. Synthetic-only training reached a Dice score of up to 0.60. Synthetic pretraining followed by fine-tuning on the full real training set (two annotated rhizotron images) reached 0.80, and nnU-Net v2 and Mask2Former reached 0.62 with only two annotated image patches containing visible roots. When real annotations were limited, pretraining also improved segmentation of unseen crops, including lettuce, barley, wheat, maize, tomato, cucumber, French bean and field bean. We also introduced a Soft Dice score that compares predictions with probabilistic consensus maps from seven independent annotators, accounting for annotator disagreement. Models matched average human agreement on low-ambiguity patches but remained below it on high-ambiguity patches. In conclusion, diffusion-generated images can complement real annotations and help scale root segmentation, although their effect on downstream root trait measurements still needs to be validated.