MycorrhizaFinder: a user-friendly machine-learning tool to quantify endomycorrhizal colonisation of nursery and field...

Root colonisation by endomycorrhizal fungi has long been a widely used metric in mycorrhizal studies. However, due to the significant time required to assess colonisation using traditional microscopy techniques, studies of colonisation at large scales are impractical. Machine-learning-based approaches may increase output and facilitate ecosystem assessments. We trained a machine-learning-based tool (MycorrhizaFinder) on 42 field roots (132 unique 1 cm sections, with > 100,000 tiles) and validated on a further 37 sections from grasslands and heathlands hosting common Northern European plants with arbuscular (AM) and ericoid mycorrhizal (ErM) fungal structures, and dark septate endophytes (DSE), also common in field-sourced roots. We incorporated a user-customized confidence threshold to encourage engagement with morphological ambiguities, in e.g. ErM and DSE. A Macro F₁ statistic was used to assess the tool's development performance. We provide a sample workflow from root processing and microscope slide scanning to semi-automated model training and performance evaluation. Our model performance combining all classes, the Macro F₁, is conservatively 66% for arbuscular and 57% for ericoid mycorrhizal colonisation assessment. MycorrhizaFinder is user-friendly, requires no programming skills and offers flexibility for advanced agronomists or ecologists to train the tool using their own labelled mycorrhizal root datasets, including images acquired from different instruments or staining protocols. This adaptability allows users to customize the model for specific ecosystems or experimental designs. Leveraged with molecular identification and/or functional traits of fungi (e.g. arbuscules), MycorrhizaFinder supports scalable and repeatable nursery experiments and monitoring across ecosystems to assess mycorrhizal status and track land-use changes over time.