Accurate quantification of root nodules is essential for understanding legume–rhizobia symbiosis and improving biological nitrogen fixation. Fluorescently labeled rhizobial strains enable clear visualization of nodules; however, automated segmentation and counting remain challenging under data-limited conditions. In this study, we present a systematic benchmarking framework to evaluate three complementary approaches for fluorescent root nodule segmentation and quantification: a rule-based computer vision pipeline with optimized color-space thresholding, supervised deep learning using YOLOv12-seg transfer learning, and the training-free Segment Anything Model (SAM). Experiments were conducted on a pilot-scale dataset comprising 16 fluorescent images of Pisum sativum roots with manually annotated blue and yellow nodules. To address instability associated with small test sets, a 4-fold cross-validation strategy was employed, and performance was reported as mean ± standard deviation across folds. Model performance was evaluated using pixel-level overlap metrics (IoU, Dice), instance-level precision, recall, and F1-score, and total nodule counting error (MAE and ). The rule-based approach achieved strong segmentation and counting accuracy when fluorescence provided clear chromatic separation, demonstrating near-zero bias in total counts. Among deep learning models, YOLOv12-m provided the most balanced performance, achieving high segmentation accuracy while minimizing counting error and inter-fold variability. Larger YOLO variants did not consistently improve quantitative outcomes, suggesting overfitting under data-scarce conditions. SAM produced stable segmentation masks without training, but systematically underestimated nodule counts and lacked intrinsic class discrimination. Overall, the results highlight that segmentation fidelity alone is insufficient for reliable biological phenotyping and that accurate nodule counting must be explicitly considered. While limited in scope, this study establishes a reproducible benchmarking framework for fluorescent nodule phenotyping and provides practical guidance on method selection under constrained data and computational resources. The findings are intended as a proof-of-concept and motivate future work on larger, publicly available fluorescent datasets and hybrid segmentation strategies. • A comparative framework is presented for fluorescent root nodule segmentation under limited data conditions. • Rule-based computer vision, YOLOv12-seg transfer learning, and SAM zero-shot segmentation are systematically benchmarked. • Fluorescent imaging enables clear nodule–background separation, facilitating proof-of-concept automated phenotyping. • Rule-based segmentation provides a strong, interpretable baseline, while deep learning models offer improved scalability at higher computational cost.
Salem et al. (Wed,) studied this question.