UNet-ECA-Bio: a biologically informed deep learning model for high-throughput micro-phenotyping of rice stem vascular...

Rice stem internal structure is a critical micro-phenotype influencing lodging resistance and yield; however, its analysis remains constrained by labor-intensive manual methods. Here, we present a publicly available dataset of 686 rice stem cross-sections, with 21,027 large vascular bundles (LVBs) and 19,342 small vascular bundles (SVBs) manually annotated. Five deep learning architectures were systematically evaluated, among which UNet-VGG16 achieved the best performance with a mean intersection over union (mIoU) of 87.4% (82.49% for LVBs and 74.41% for SVBs). An improved model, UNet-ECA-Bio, further raised mIoU to 89.32% and SVB IoU to 78.97% by integrating Efficient Channel Attention (ECA) and biologically informed class weighting using an image-level dataset. Leveraging these high-accuracy phenotypic predictions, genome-wide association studies (GWAS) indicated concordance between annotated and predicted traits, with SNP overlap rates of 96% (LVB count: 1,217/1,262), 43% (SVB count: 6/14), 98% (stem area: 122/124), and 100% (cavity area: 3/3) at -log10(p) ≥ 6. Meanwhile, compared with manual annotation (estimated 10-30 minutes per image), the proposed approach processed all 686 images within 10 minutes, representing a >600-fold increase in throughput. We further developed a user-friendly software tool, "Rice_Stem_Pre_V1.1.exe," for automated phenotyping of 14 stem traits, providing a cost-effective platform for genetic studies of lodging resistance and yield improvement.