A computer vision-based approach for high-throughput automated analysis of Arabidopsis seedling phenotypes

Phenotype observations are common methodologies in plant biology studies, ranging from recording growth parameters to flowering dates. Identifying mutants or varieties with different phenotypes greatly advances our understanding of regulatory mechanisms in plant growth and development. Over the past 2 decades, naked-eye-based observations and manual measurements using ImageJ software have been leading approaches for recording phenotypes. However, these low-efficiency and error-prone methods have met difficulties in large-scale pipelines. Although some high-throughput imaging platforms have been commercialized, it remains challenging to efficiently, conveniently, accurately, and automatically analyze data generated by these platforms. To address this issue, we designed an automatic phenotype analysis tool. We trained a YOLOv11 (You Only Look Once version 11) model to locate Arabidopsis thaliana seedlings grown on petri dishes and developed a high-accuracy semantic segmentation model based on Swin Transformer and kernel update head, achieving a segmentation accuracy of 83.56% mIoU. By postprocessing the segmentation masks, we automated the analysis of 5 representative seedling phenotypes: hypocotyl length, root length, root gravitropic bending angle, petiole length, and cotyledon opening rate. Compared with manual recording, our tool demonstrated high accuracy across all 5 phenotypes, offering a reliable and efficient solution for phenotypic analysis in plant research. Our automatic tool enables high-throughput phenotyping and will shift the traditional paradigm of phenotype recording.