A multi-model deep learning algorithm analyzing standard smartphone images has successfully quantified Botrytis cinerea lesions across 350 distinct strawberry breeding accessions, capturing fine-grained quantitative variation in disease resistance. Developed to evaluate individual fruits arranged systematically on trays during disease trials, the automated phenotyping method addresses longstanding technical and logistical bottlenecks in crop breeding programs.
Fast and accurate screening of Botrytis cinerea lesions on harvested strawberry fruits represents a critical requirement for breeding programs that focus on developing disease-resistant cultivars. In conventional breeding trials, visual evaluation has traditionally depended on manual inspection by human evaluators. However, this established practice is labor-intensive, time-consuming, and prone to human subjectivity, creating substantial operational hurdles during large-scale population screening.
Existing automated screening technologies have also encountered notable limitations in field and laboratory settings. Prior computational methods frequently demand extensive quantities of manually labeled training data to perform reliably, which is difficult to curate at scale. Furthermore, earlier automated systems often struggle to handle the pronounced visual variability in lesion appearance across genetically diverse fruit samples.
To establish a more scalable and accessible screening strategy, researchers collected digital smartphone photographs of individual strawberry fruits arranged on trays during disease trials. This image dataset captured 350 distinct strawberry accessions, providing a diverse genetic testing ground. Standard mobile devices were deliberately utilized to ensure the image acquisition protocol remained practical and reproducible for routine screening operations.
The researchers analyzed the resulting image collection using a multi-model deep learning algorithm that integrates fine-tuned task-specific models with text prompt-based models. In this integrated system, the text prompt-based models improved the generalizability of the overall framework by supporting the pre-annotation of training datasets and offering complementary detection alongside the fine-tuned models. This computational pipeline extracted vital phenotypic parameters that are difficult to quantify by hand, specifically measuring strawberry fruit size, Botrytis lesion size, and the ratio of lesion area to fruit size.
The resulting algorithmic system demonstrated robust segmentation accuracy across the diverse strawberry accessions tested in the study. For strawberry fruit instance segmentation, the multi-model architecture achieved an F1-score of 0.97 alongside a mean intersection over union of 0.92. When evaluating Botrytis lesion segmentation, the framework maintained consistent performance with a mean intersection over union of 0.97. Comparative analysis between subjective visual evaluations and objective computational measurements revealed that the objective data delivered higher discriminatory ability for identifying subtle phenotypic differences.
Marking what the research team identified as the first fruit-level multi-model deep learning approach for automated Botrytis lesion screening, the study provides a practical, consistent, and scalable phenotyping solution. Because the workflow relies exclusively on accessible mobile imaging hardware and moderate computational power, it significantly reduces technical barriers to adopting artificial intelligence in agricultural research. The authors conclude that this high-throughput phenotyping platform may ultimately support the execution of more efficient marker-assisted and genomic selection strategies across future strawberry breeding programs.
Subject of Research: Automated deep learning phenotyping of Botrytis cinerea resistance in strawberry breeding
Article Title: High-throughput fruit-level phenotyping captures quantitative variation in Botrytis cinerea resistance across strawberry breeding populations
Article References: Zerdoner, M., Gabriëls, S., Mensink, M., Arens, P., Visser, R. G. F., & Rustia, D. J. A. (2026). High-throughput fruit-level phenotyping captures quantitative variation in Botrytis cinerea resistance across strawberry breeding populations. Plant Methods. https://doi.org/10.1186/s13007-026-01616-4
Image Credits: AI Generated
DOI: 10.1186/s13007-026-01616-4
Keywords: Botrytis cinerea, strawberry phenotyping, deep learning, instance segmentation, fruit-level phenotyping, disease resistance screening, plant breeding programs, smartphone imaging, lesion segmentation, text prompt-based models, quantitative variation, marker-assisted selection, genomic selection, phenotypic metrics, strawberry accessions, automated disease screening
News Source: Blake Davidson. (November 19, 2026). Deep Learning Phenotyping Accurately Quantifies Botrytis Resistance Across Strawberry Breeding Accessions. Scienmag.



