USWBSI

USWBSI Abstract Viewer

2026 National Fusarium Head Blight Forum


Variety Development and Host Resistance (VDHR)

Poster # 523

Automated Spikelet-Level Phenotyping of Fusarium Head Blight Severity in Wheat Using Vision Transformer Models

Authors & Affiliations:

Jinglie Zhou 1,2, Anil Karmacharya 3, Dhondup Lhamo 1, Steven S. Xu 1, Yong-Qiang Gu 1, and Taner Z. Sen 1,4
1. USDA-ARS, Crop Improvement and Genetics Research Unit, Western Regional Research Center, Albany, CA 94710, USA
2. Oak Ridge Institute for Science and Education (ORISE), Oak Ridge, TN 37830, USA
3. Department of Plant Sciences, University of California, Davis, CA 95616, USA
4. Department of Bioengineering, University of California, Berkeley, CA, 94720, USA
Corresponding Authors: Jinglie Zhou, Jinglie.Zhou@usda.gov and Taner Z. Sen, taner.sen@usda.gov

Presenting Author:

Zhou, Jinglie
jingliezhou@gmail.com

Abstract:

Fusarium head blight (FHB), primarily caused by Fusarium graminearum, is a devastating fungal disease that causes severe yield losses and toxic deoxynivalenol (DON) contamination, threatening global food security. Breeding resistant wheat cultivars is the primary strategy to mitigate FHB hazards; however, traditional visual phenotyping of FHB severity is labor-intensive, subjective, and non-auditable, creating a bottleneck in large-scale germplasm screening. While image-based deep learning offers a non-destructive alternative, accurate spikelet-level quantification requires robust generalization across diverse imaging environments and plant morphologies. Here, we developed an automated, end-to-end computer vision pipeline centered on our flagship Vision Transformer model (RF-DETR Large) that directly localizes and counts individual healthy and Fusarium-infected spikelets in wheat spike images to compute the infected-to-total spikelet ratio. To evaluate cross-domain generalization, we tested our pipeline across independent datasets from two distinct geographic regions and laboratories, spanning diverse wheat genotypes and spike morphologies, at multiple infection stages at 7, 10, and 14 days after inoculation (DAI), and in contrasting backgrounds (natural greenhouse and black backdrops). In the internal USDA test cohort, the RF-DETR Large transformer model achieved strong predictive accuracy for the infected-to-total spikelet ratio, yielding a Pearson correlation of r = 0.97 (RMSE = 7.9%) at the single-spike level under natural greenhouse backgrounds. Furthermore, zero-shot evaluation on an external benchmark of 2,117 images from an independent laboratory demonstrated that RF-DETR Large maintained high accuracy across all temporal infection stages (r = 0.89–0.92, RMSE = 12.3%–15.7%; overall r = 0.91, RMSE = 13.7%), exhibiting superior predictive performance and generalization compared to mainstream YOLOv8–v11 series models and existing threshold-based methods, particularly in early-stage (7 DAI) lesion detection. Our computational pipeline provides a consistent and reliable way to capture, measure and store wheat spike images. In this presentation, we will discuss our pipeline architecture, highlight its cross-domain generalization, and demonstrate its web-powered mobile prototype within the GrainGenes ecosystem.


© Copyright 2026 by individual authors. All rights reserved. No part of this abstract or paper publication may be reproduced without prior permission from the applicable author(s).