MycorrhizaFinder: an efficient machine learning tool to quantify endomycorrhizal colonisation of real-world roots

MycorrhizaFinder is a user-friendly, semi-automated machine learning tool designed to efficiently quantify endomycorrhizal colonisation in diverse field roots, overcoming the time limitations of traditional microscopy to enable large-scale ecosystem monitoring.

The Gist

Imagine you are a detective trying to solve a mystery hidden inside a forest. The mystery isn't a crime, but a secret partnership between plants and fungi.

For decades, scientists have known that most plants have a "best friend" in the form of microscopic fungi living inside their roots. These fungi act like a super-charged extension of the plant's root system, helping it drink water and eat nutrients from the soil. In exchange, the plant gives the fungi sugar. This relationship is called mycorrhiza.

To understand the health of a forest, a farm, or a meadow, scientists need to know how many of these fungi are actually living inside the roots. But here's the problem: The old way of checking is incredibly slow and boring.

The Old Way: The "Human Magnifying Glass"

Traditionally, to see these fungi, a scientist has to:

1. Dig up a root.
2. Clean it and stain it with blue ink (so the fungi show up).
3. Put it under a microscope.
4. Squint at it for hours, manually counting tiny dots and lines on a grid.

It's like trying to count every single grain of sand on a beach by picking them up one by one. If you want to check 1,000 roots to understand a whole ecosystem, you'd need a team of people working for months. It's too slow to help us react quickly to climate change or pollution.

The New Hero: MycorrhizaFinder (MFKew)

This paper introduces a new tool called MycorrhizaFinder (or MFKew). Think of this as a super-smart, tireless robot assistant that can look at a picture of a root and instantly tell you, "Hey, this part has fungi, this part doesn't, and this part is just dirt."

Here is how it works, using some simple analogies:

1. The "Puzzle Piece" Strategy

Instead of looking at the whole root at once (which is too messy and complex), the tool cuts the image of the root into thousands of tiny square "tiles," like a jigsaw puzzle.

• The Old Tool (AMFinder): The previous version of this robot was trained in a "perfect world" (a lab with clean, simple plants). It was like a student who only studied in a quiet library. When you put it in a noisy, messy forest (real-world roots with dirt, clay, and weird shapes), it got confused and failed.
• The New Tool (MFKew): This robot was trained in the "messy classroom" of the real world. It learned to ignore dirt, clay, and blurry spots. It was taught to recognize not just one type of fungus, but many different kinds, including the tricky ones that look like dark, stringy endophytes (DSE).

2. The "Gossip" Feature (Context)

One of the coolest things about MFKew is its ability to use context.
Imagine you are looking at a single tile and you aren't sure if it's a fungus or just a weird stain.

• Without Context: The robot looks at the tile and guesses, "Maybe it's a fungus?"
• With Context: The robot looks at the four tiles surrounding it. It says, "Wait, the neighbors all look like fungus, so this tile is probably fungus too!"
  This is like a detective asking the neighbors for their opinion before making an arrest. This "gossip" feature made the robot much more accurate.

3. The "Human-in-the-Loop" Safety Net

The authors didn't just want a robot that guesses blindly. They built a system where a human expert can still be the boss.

• The robot does the heavy lifting, scanning hundreds of roots in the time it takes a human to scan one.
• If the robot is unsure (maybe the image is blurry), it highlights that spot and says, "I'm not 100% sure, human boss, please take a quick look."
• This means the human only has to fix the mistakes, not do the whole job. It's like having a spell-checker that highlights only the words you might have misspelled, rather than rewriting your whole essay for you.

Why Does This Matter?

• Speed: What used to take a human a whole day to analyze can now be done by the computer overnight.
• Scale: We can now check thousands of roots across entire countries, not just a few in a single lab. This helps us track how ecosystems are changing due to climate change or pollution.
• Accessibility: You don't need to be a computer programmer to use it. It's a user-friendly app that anyone with a laptop can run. Even better, if you are a scientist studying a specific type of plant, you can "teach" the robot your specific needs without writing a single line of code.

The Bottom Line

MycorrhizaFinder is like giving scientists a pair of X-ray glasses and a super-speed calculator. It turns a slow, tedious, manual counting job into a fast, automated, and highly accurate process. This allows us to finally understand the hidden underground world of plant-fungus partnerships on a scale that matches the size of our planet, helping us protect our natural capital for the future.

Technical Summary

1. Problem Statement

Traditional methods for assessing mycorrhizal root colonisation rely on manual microscopy (e.g., grid intersection or visual estimation), which is labor-intensive, time-consuming, and impractical for large-scale ecosystem assessments requiring hundreds or thousands of samples. While previous AI tools like AMFinder existed, they were trained primarily on controlled, laboratory-grown model plants with simple fungal structures. These tools failed when applied to field-sourced roots, which present complex challenges:

• Biological Complexity: Co-occurrence of Arbuscular Mycorrhizal (AM), Ericoid Mycorrhizal (ErM), and Dark Septate Endophytes (DSE).
• Environmental Artifacts: Soil debris, clay particles, irregular hyphal diameters, and variable staining quality.
• Structural Diversity: Field roots contain a mosaic of fungal morphotypes and life stages that binary classification models cannot distinguish.

2. Methodology

Data Collection and Preparation

• Source: Roots were collected from diverse field ecosystems (agricultural, grassland, heathland, bog) across England.
• Imaging: Roots were cleared, stained (blue-ink for AM, Trypan blue for ErM), and scanned at 37x magnification using a brightfield digital slide scanner. Extended focus imaging (7 focal planes) was used to capture 3D fungal networks.
• Dataset: 133 unique roots (45 AM, 88 ErM) were processed into tiles. The dataset was heavily imbalanced, with "Background" and "Uncolonised" tiles dominating.

Model Architecture and Classes

The authors moved beyond the original 3-class AMFinder model to a more granular classification system:

• Architecture: Evaluated ResNet, ResNeXt, and EfficientNetB5. EfficientNetB5 was selected for its balance of performance (~30M parameters) and efficiency.
• Expanded Class Structure: The model output layer was expanded to distinguish specific biological features:
  • AM Pathway: AM Colonised, DSE Colonised, Hybrid (AM+DSE), Uncolonised, Background, Unreadable.
  • ErM Pathway: Blue Coils, Brown Coils, Type Two, Hybrid ErM, Hybrid DSE, Main Root, Uncolonised, Background, Unreadable.
• Contextualisation: A Max-vote system was implemented where the prediction for a target tile is adjusted based on the majority vote of four partially overlapping neighboring tiles. This mimics human contextualization to resolve ambiguous structures.

User Interface and Workflow

• Platform: Migrated from a standalone OCaml application to a web-based React application.
• Features:
  • Single-workflow interface for image upload, annotation, training, and prediction.
  • Human-in-the-loop: Users can set confidence thresholds. Tiles below the threshold are flagged for manual review/annotation.
  • Database: Replaced CSV storage with PostgreSQL for better data governance.
  • Accessibility: No coding skills required; runs on Windows, MacOS, and Linux.

Performance Metrics

• Metric Selection: The authors rejected "Accuracy" due to dataset imbalance and adopted Macro F1 Score (the unweighted average of F1 scores across all classes) as the primary optimization metric. This ensures performance is measured fairly across rare classes (e.g., DSE) and common classes (e.g., Background).

3. Key Results

Performance Improvements

• Macro F1 Scores:
  • AM Pathway: Improved from 45.88% (CNN1 without context) to 66.47% (EfficientNetB5 with context).
  • ErM Pathway: Improved from 40.61% (CNN1 without context) to 57.03% (EfficientNetB5 with context).
• Contextual Impact: Enabling the contextual function increased Macro F1 by 5.0% for AM and 5.8% for ErM, significantly reducing false negatives and positives in complex root sections.
• Comparison to AMFinder: On field roots, the original AMFinder predicted only 0.5% AM colonisation (missing many true positives), whereas MFKew predicted 10.7%, closely aligning with expert human estimates of 12.7%.

Specific Class Performance

• Background/Uncolonised: Achieved high F1 scores (>98% and ~80% respectively) due to distinct visual features.
• Rare Classes: Despite low representation (e.g., DSE at 0.9% of tiles), the model achieved meaningful F1 scores (e.g., DSE F1 ~38% without context, improving with data balancing).
• Generalization: The model performed well on images captured by different devices (digital microscopes vs. slide scanners), though performance dipped slightly (Macro F1 58% vs 65%) on non-standard inputs.

Efficiency

• Throughput: MFKew significantly reduced analysis time. While manual microscopy takes 30–60 minutes per root section, MFKew allows for automated prediction of dozens to hundreds of roots overnight.
• Workflow: The ability to batch process and manually correct only low-confidence tiles drastically reduces the time required for expert validation compared to full manual assessment.

4. Key Contributions

1. First Field-Optimized Tool: Developed the first machine learning tool specifically trained and validated on field-sourced roots rather than controlled nursery samples, addressing the "real-world" complexity gap.
2. Dual-Pathway Architecture: Created a unified tool with toggleable pathways for both Arbuscular (AM) and Ericoid (ErM) mycorrhizae, including the ability to distinguish specific ErM morphotypes (Blue/Brown coils, Type Two).
3. Differentiation of Artifacts: Successfully integrated classes for Dark Septate Endophytes (DSE) and Unreadable tiles (obscured by debris), preventing these from being misclassified as colonised or uncolonised roots.
4. Contextual Learning: Demonstrated that incorporating spatial context (neighboring tiles) via a max-vote system significantly improves classification accuracy for complex biological structures.
5. Open-Source Accessibility: Released a user-friendly, non-coding interface with open-source code (MIT license), allowing ecologists to retrain the model on their own specific datasets.

5. Significance

• Scalability: MFKew enables medium-to-high throughput monitoring of mycorrhizal status across landscapes, facilitating large-scale ecosystem assessments (e.g., Defra's Natural Capital and Ecosystem Assessment programme).
• Standardization: Provides a repeatable, objective method for quantifying root colonisation, allowing for direct comparison across different studies, habitats, and time periods.
• Policy and Management: By linking root colonisation data to ecosystem health and natural capital, the tool supports evidence-based land management and environmental policy.
• Future-Proofing: The modular design allows for continuous improvement. As more data is labeled, the model can be retrained to improve performance on underrepresented classes, potentially moving toward fully automated workflows in the future.

In conclusion, MycorrhizaFinder represents a significant leap forward in mycorrhizal ecology, bridging the gap between traditional microscopy and modern computer vision to make large-scale, accurate assessment of root-fungal interactions feasible for the first time.