Diffusion-Based Synthetic Pretraining for Root Segmentation in Rhizotron Images

Rhizotrons provide a non-destructive method to examine the root system architecture within soil. However, the automated analysis of rhizotron images proves to be challenging due to the low contrast of roots against the soil background, the similarity of soil particles to fine roots, and the requirement of extensive manually annotated images for deep learning model training.

Synthetic images could potentially provide abundant labeled data, yet rhizotron images are scarce in the datasets utilized for training generative models. To address this issue, researchers developed a pipeline incorporating Low-Rank Adaptation (LoRA), ControlNet, and inpainting to create synthetic rhizotron images with associated root masks.

This pipeline was then evaluated against training on synthetic data alone, real data alone, and a combination of synthetic pretraining followed by fine-tuning on real images, employing three segmentation models: nnU-Net v2, YOLO26, and Mask2Former.

The synthetic data-only training attained a Dice score of up to 0.60. When synthetic pretraining was followed by fine-tuning on the complete set of real images, consisting of just two annotated rhizotron images, the Dice score increased to 0.80. Notably, both nnU-Net v2 and Mask2Former achieved a Dice score of 0.62 when trained on just two annotated image patches containing visible roots.

Furthermore, when the number of real annotations was limited, pretraining proved beneficial in improving the segmentation of unseen crops, such as lettuce, barley, wheat, maize, tomato, cucumber, French bean, and field bean. To account for potential annotator disagreement, the researchers introduced a Soft Dice score that compares predictions with the probabilistic consensus maps from seven independent annotators.

While the models managed to align their performance with the average human agreement on low-ambiguity patches, they still fell short on high-ambiguity patches.

In summary, the utilization of diffusion-generated images can effectively complement real annotations and facilitate the scaling of root segmentation. However, the impact of these synthetic images on downstream root trait measurements necessitates further validation.