Abstract
Vascular discoloration is used to assess Fusarium wilt symptoms in cotton, but manual image an-notation limits the scale of quantitative phenotyping. We developed an nnU-Net workflow to seg-ment stem cross-sections and estimate the percentage of included stem cross-sectional area showing visible vascular discoloration. The dataset comprised 274 images from an F8 Gossypium hirsutum population grown in a Fusarium oxysporum f. sp. vasinfectum race 4 screening nursery in Clint, Texas. Manual annotations delineated stem tissue and discolored regions, and model development used five-fold cross-validation. Model-derived percentages had a mean absolute error of 1.65 percentage points and a median absolute error of 0.70 percentage points relative to the reference annotations; 255 of 274 estimates (93.1%) were within 5 percentage points. Preprocessing, inference, and post-processing required approximately 50 minutes for the complete dataset, excluding image acquisition and model training. Disagreements included faint and marginal discoloration and differences in annotation boundaries. The workflow supports automated quantification of stem discoloration in this dataset; independent validation is needed to establish performance across populations and imaging conditions.