Diffusion-Based Synthetic Pretraining for Root Segmentation in Rhizotron Images

Diffusion-generated rhizotron images with paired root masks enabled nnU-Net v2 and Mask2Former to reach 0.80 Dice after synthetic pretraining plus fine-tuning on just two annotated real images.

• Why it matters: Automated root segmentation in rhizotron imagery is hindered by low root–soil contrast and scarce, subjective annotations, while generative models lack rhizotron training data.
• What they did: A pipeline using LoRA, ControlNet, and inpainting generated synthetic rhizotron images and masks; segmentation models (nnU-Net v2, YOLO26, Mask2Former) were trained on synthetic-only, real-only, or synthetic-pretrained-then-fine-tuned data.
• The result: Synthetic-only training achieved up to 0.60 Dice, while synthetic pretraining plus fine-tuning on the full real training set reached 0.80 and improved unseen crops; Soft Dice matched human agreement on low-ambiguity patches but not on high-ambiguity ones.

AI summary of the abstract · may contain errors · not medical advice · read the paper.