The Automatic Mycorrhiza Finder (AMFinder) allows for high-throughput computer vision-based identification and quantification of AM fungal colonisation and intraradical hyphal structures using convolutional neural networks.

The current version of AMFinder is v2.0.

If you use AMFinder in your manuscript, please cite: Evangelisti et al., 2021, Deep learning-based quantification of arbuscular mycorrhizal fungi in plant roots, New Phytologist 232(5): 2207-2219.

Summary

1. Installation
2. Batch processing (amf)
3. Annotation browser (amfbrowser)
4. A typical annotation pipeline
5. How to batch stain plant roots?

Installation

Detailed installation instructions for Linux, Mac and Windows can be found here. If you are not familiar with Python virtual environments, you may want to read this page first.

Batch processing (amf)

The AMFinder tool amf is a batch mode program that predicts fungal structures, converts predictions into annotations, and train neural networks. It is used as follows:

$ source amfenv/bin/activate
(amfenv) $ amf <action> <parameters> <images> 

where <action> is either:

• predict: prediction of fungal colonisation (CNN1) and intraradical hyphal structures (CNN2),
• convert: automatic conversion of predictions to annotations, or
• train: neural network training.

<images> are the paths to the JPEG or TIFF images to analyse.
Details about <parameters> are given in the following sections.

Prediction mode

This mode is used to predict fungal colonisation (CNN1) and intraradical hyphal structures (CNN2).

Short	Long	Description	Default value
-net CNN	--network CNN	Mandatory. Use network CNN (see list below).
-t N	--tile_size N	Optional. Use N pixels as tile size.	N = 126

Pre-trained networks to be used with the parameter -net are available in the folder trained_networks. AMFinder is looking for trained networks in this folder only. Below is a list of publicly available networks. The image datasets used to generate them are available on Zenodo.

File name	Annotation level	Description
CNN1v1.h5	CNN1	Ink-stained, ClearSee-treated root pictures (flatbed scanner/microscope).
CNN1v2.h5	CNN1	Same, but trained with data augmentation.
CNN2v1.h5	CNN2	Ink-stained, ClearSee-treated microscope root pictures.
CNN2v2.h5	CNN2	Same, but trained with data augmentation.

Are you working with a system that appears challenging for AMFinder? Please get in touch! We would be happy to help generate specialised CNN1/2 networks and make them widely available to the research community.

Conversion mode

This mode is used to convert amf predict predictions (i.e. probabilities) to annotations.

Short	Long	Description	Default value
-1	--CNN1	Optional. Convert CNN1 predictions.	yes
-2	--CNN2	Optional. Convert CNN2 predictions.	no
-th X	--threshold X	Optional. Use X as threshold for CNN2 conversions.	X = 0.5

Training mode

This mode is used to train AMFinder neural networks on different images. All parameters listed below are optional.

Short	Long	Description	Default value
-net N	--network N	Use network N.	none
-b N	--batch_size N	Use a batch size of N tiles.	N = 32
-k	--keep_background	Do not skip any background tile.	False
-a	--data_augmentation	Activate data augmentation.	False
-s	--summary	Save CNN architecture and graph.	False
-o DIR	--outdir DIR	Save trained model and CNN architecture in DIR.	cwd
-e N	--epochs N	Perform N training cycles.	N = 100
-p N	--patience N	Wait for N epochs before early stopping.	N = 12
-lr X	--learning_rate X	Use X as learning rate for the Adam optimiser.	X = 0.001
-vf N	--validation_fraction N	Use N percents of total tiles as validation set.	N = 15%
-1	--CNN1	Train for root colonisation.	True
-2	--CNN2	Train for intraradical hyphal structures.	False

Training can benefit from high-performance computing (HPC) systems. Below is a template script for Slurm:

#! /bin/bash
#SBATCH -e <error_file>
#SBATCH -o <output_file>
#SBATCH --mem=<memory_GB>
#SBATCH -n <procs>

ROOT=/home/<user>/amf

source $ROOT/amfenv/bin/activate
$ROOT/amf train <parameters> <images>
deactivate

Annotation browser (amfbrowser)

The AMFinder standalone graphical interface amfbrowser enables the manual inspection of predictions and annotations. This tool can display predictions and annotations on top of the corresponding images. It enables users to edit values, to browse low-quality predictions, and to perform automatic conversion of amf predictions to annotations. You can run amfbrowser as follows:

$ amfbrowser <parameter> <image>

where <parameter> can be used to specify tile size (see below).

Short	Long	Description	Default value
-t N	--tile_size N	Optional. Use N pixels as tile size.	N = 126

A typical annotation pipeline

Below is a bash script describing a typical prediction/annotation pipeline:

#! /bin/bash

source amfenv/bin/activate

# Predict fungal colonisation (CNN1) on a bunch of JPEG images.
./amf predict ink_stained_image{1-5}.jpg

# Convert CNN1 predictions to annotations.
./amf convert ink_stained_image{1-5}.jpg

# Predict intraradical structures (CNN2) on the same images.
./amf predict --network CNN2v2.h5 ink_stained_image{1-5}.jpg

# Convert CNN2 predictions to annotations using a threshold of 0.6.
./amf convert --CNN2 --threshold 0.6 ink_stained_image{1-5}.jpg

deactivate

Computer predictions and annotations can be checked and amended by running amfbrowser on each image. Although not absolutely necessary, user supervision and validation of computer predictions are recommended for quality control.

How to batch stain plant roots?

An optimised ink-staining protocol with additional clearing is available in Evangelisti et al. (2021).

Batch staining plant roots is essential for high-throughput analyses. It can be achieved using a hand-crafted device composed of 10 cell strainers (100 µm nylon mesh) tied together with adhesive sealing film for PCR plates. A single sieve can accommodate a 4-week-old N. benthamiana root system. The sieves containing roots are immersed in 10% KOH, water, or ink/vinegar staining solution poured in a plastic lid. Plastic lids are floated in a hot (95°C) water bath to achieve the desired staining conditions. Sieves are pulled out of the lid containing the KOH solution and transferred to the washing solution, then to the ink staining solution without the need to manipulate roots, thereby reducing the risk of damage.

View from above	View from below	With plastic lid

Credit: devices from Dr Albin Teulet and Alex Guyon (Schornack lab), derived from an original idea by Dr Clément Quan.