Fusarium head blight (FHB) is an economically impactful fungal pathogen of wheat. Although there are numerous strategies to limit the effects of this disease, one important approach is to develop resistant varieties. Currently, phenotyping is a limiting step in the development of resistant varieties as it requires a lot of labor, time, is not standardized across the community, and is subjective. We aimed to reduce these phenotyping limitations by using a novel, high-throughput phenotyping rover to capture in-field RGB images of inoculated wheat spikes. The images acquired from the rover were manually annotated for spikes and regions of pathogen symptoms on the spikes and used in deep neural networks to locate and quantify FHB on each spike. The inference models were validated at the spike and plot level with manual disease scores from five raters on images and in the field. The combination of the rover and image models exceeded conventional rating methods and can reduce many of the limitations of conventional FHB phenotyping. The plot level scores from in-field raters and from image-based inferences were strongly correlated. Likewise, a high correlation was found in comparisons of image-based individual spike FHB disease scores from the models and raters. The developed image-based FHB disease detection methods continued to perform well across environments, image types, and levels of disease. These results show that current field FHB phenotyping methods can be improved through precise and efficient quantification of disease symptoms on a plot and individual spike throughout the growing season.