Files
Abstract
Arbuscular mycorrhizal fungi (AMF) form one of the most ancient and widespread symbioses, enhancing plant nutrient acquisition in exchange for photosynthetically derived carbon. Accurately quantifying AMF colonization at scale, however, remains a major bottleneck. This dissertation presents a three-stage roadmap for developing robust, deep learning–based tools that enable high-throughput, automated analysis of AMF in sorghum (Sorghum bicolor) roots.First, a pilot study combined Mask R-CNN with mixed linear models to segment individual fungal structures in a recombinant inbred sorghum population and to relate colonization levels to root niche and fungal structure allocation. The study demonstrated that deep learning can capture biologically meaningful AMF phenotypes.
Second, to overcome data scarcity in training deep learning models, I assembled MycorrhiSEE, a 15 TB collection of ~137,500 whole-slide images (WSIs) from 5,500 sorghum plants spanning 337 genotypes and diverse field treatments. A spline-guided tiling algorithm transformed gigapixel WSIs into uniform patches. Eight bootstrap evaluations confirmed consistent spline interpolation across expert-rated image quality classes.
Third, building on MycorrhiSEE, an integrated pipeline was developed featuring (i) an enhanced spline-guided tiling algorithm with quantitative tiling quality metrics, and a two-step CNN-based classification that (ii) first removes background tiles with 99.7 % accuracy and then (iii) distinguishes AMF colonized from non-colonized image tiles. A wide selection of ImageNet-pretrained architectures was benchmarked to identify the optimal classifiers. DenseNet and ResNet50-based classification classifier achieved >98% accuracy and superior generalization on both MycorrhiSEE and the external AMFinder dataset.
Collectively, these contributions—from computer vision modeling to large-scale dataset curation and pipeline optimization—provide a practical framework for rapid, unbiased AMF phenotyping. The resulting tools enable scalable integration of imaging, genomic, and environmental data, advancing precision agriculture and ecological research on AMF to improve sorghum performance under diverse field conditions.