Detection and characterization of spike architecture based on deep learning and X-ray computed tomography in barley

Background: The cereal spike is the main harvested plant organ determining the grain yield and quality, and its dissection provides the basis to estimate yield- and quality-related traits, such as grain number per spike and kernel weight. Phenotypic detection of spike architecture has potential for genetic improvement of yield and quality. However, manual collection and analysis of phenotypic data is laborious, time-consuming, low-throughput and destructive. Results We used a barley model to develop a non-invasive, high-throughput approach through combining X-ray computed tomography (CT) and deep learning model (UNet) to phenotype spike architectural traits. We used an optimized 3D image processing methods by point cloud for analyzing internal structure and quantifying morphological traits of barley spikes. The volume and surface area of grains per spike can be determined efficiently, which is hard to be measured manually. The UNet model was trained based on two types of spikes (wheat cultivar D3 and two-row barley variety S17350), and the best model accurately predicted grain characteristics from CT images. The spikes of ten barley varieties were analyzed and classified into three categories, namely wild barley, barley cultivars and barley landraces. The results showed that modern cultivated barley has shorter but thicker grains with larger volume and higher yield compared to wild barley. The X-ray CT reconstruction and phenotype extraction pipeline needed only 5 minutes per spike for imaging and traits extracting. Conclusions The combination of X-ray CT scans and a deep learning model could be a useful tool in breeding for high yield in cereal crops, and optimized 3D image processing methods could be valuable means of phenotypic traits calculation.