Segmenting Arbuscular Mycorrhizal Fungi (AMF) hyphal networks with an alternative carbon source using U-Net.

The symbiotic relationship between plant roots and mycorrhizal fungi has driven the construction of fungal networks to gather and transport nutri- ents to and from the plant root. To understand the dynamics behind the construction of these fungal networks monitoring them over time is cru- cial. We do so by imaging them and then segmenting the hyphal networks from the images. We started growing hyphal networks under asymbiotic conditions. These asymbiotic networks are especially interesting, since mycorrhizal fungi are characterised by their relationship to plants, mak- ing them an anomaly. Unfortunately, segmenting asymbiotic mycorrhizal fungal networks, where the fungal network grows without a plant rela- tionship, has been challenging. Intrusive shapes in the plane, as well as subsequent out of focus images inhibit previous segmentation tech- niques performance. Machine Learning techniques, especially Convolu- tional Neural Networks (CNNs) have been an incredible tool for many...Show moreThe symbiotic relationship between plant roots and mycorrhizal fungi has driven the construction of fungal networks to gather and transport nutri- ents to and from the plant root. To understand the dynamics behind the construction of these fungal networks monitoring them over time is cru- cial. We do so by imaging them and then segmenting the hyphal networks from the images. We started growing hyphal networks under asymbiotic conditions. These asymbiotic networks are especially interesting, since mycorrhizal fungi are characterised by their relationship to plants, mak- ing them an anomaly. Unfortunately, segmenting asymbiotic mycorrhizal fungal networks, where the fungal network grows without a plant rela- tionship, has been challenging. Intrusive shapes in the plane, as well as subsequent out of focus images inhibit previous segmentation tech- niques performance. Machine Learning techniques, especially Convolu- tional Neural Networks (CNNs) have been an incredible tool for many visual recognition tasks. We propose the use of the U-Net CNN architec- ture in combination with extensive data augmentation, that is modelled after the dominant forms of noise that inhibit segmentation. This model preserves the ability to segment symbiotic hyphal networks from images, even improving on hard to segment networks , while also being robust to noise, making it capable of also segmenting asymbiotic hyphal networks from images.Show less

Leiden University Student Repository

Documents

In Collections

Segmenting Arbuscular Mycorrhizal Fungi (AMF) hyphal networks with an alternative carbon source using U-Net.