MONICA: A Web Application for Automated Whole Optic Nerve Contour Extraction and Morphometric Analysis Validated Across Taxonomic Orders and Image Quality Levels

The authors present MONICA, a web-based application that combines deep learning with a novel morphology-based algorithm to automatically extract whole optic nerve boundaries and calculate morphometric metrics, achieving high accuracy across diverse species, strains, and image qualities without requiring local software installation.

The Gist

Imagine your eye is like a high-speed fiber-optic internet cable, and the optic nerve is the main cable bundle connecting your eye to your brain. When you have a condition like glaucoma, it's like someone is slowly crushing that cable, damaging the individual wires (axons) inside.

For a long time, scientists trying to measure this damage had a blind spot. They could count how many wires were broken, but they couldn't easily measure how crowded the cable was or how much of the space was filled with protective padding (glial cells) versus actual wires. To get those numbers, they had to manually trace the outer edge of the nerve bundle with a pen on a screen—a slow, tedious, and error-prone task that required expensive computers and coding skills.

Enter MONICA.

What is MONICA?

Think of MONICA as a super-smart, free, online "digital assistant" for eye researchers. It's a website you can open in any browser (no heavy software installation needed) that acts like a high-tech auto-drawing tool.

Here is how it works, using a simple analogy:

1. The Input (The Photo): You upload a microscopic photo of a nerve slice. It's like handing a blurry, complex map to a GPS.
2. The AI Brain (AxonDeepSeg): First, MONICA uses a deep learning "brain" (called AxonDeepSeg) to instantly identify every single wire (axon) and its protective coating (myelin) in the photo. It's like a detective spotting every single thread in a tangled ball of yarn.
3. The Magic Trick (Contour Extraction): This is the new, special part. Once the AI knows where all the wires are, it automatically draws a perfect line around the entire bundle.
   • The Analogy: Imagine a group of people holding hands in a circle. Previous tools could count the people, but they couldn't tell you how big the circle was. MONICA is like a robot that instantly sees the group and draws a perfect rope around the outside of the circle, no matter how squiggly or irregular the shape is.
4. The Output (The Report): Because it now knows the size of the circle (the nerve area) and the number of people inside (the axons), it can instantly calculate:
   • Density: How packed are the wires? (Are they squeezed too tight?)
   • Padding: How much space is taken up by the protective "glue" (glial cells) instead of wires?

Why is this a Big Deal?

Before MONICA, researchers had to do the "drawing the circle" part by hand. It was like trying to measure the size of a cloud by tracing it with a ruler while it was moving. It took forever and different people would draw different lines.

• Accessibility: MONICA runs in your web browser. You don't need a supercomputer or to know how to code. It's like using Google Maps instead of building your own GPS from scratch.
• Speed: It can process hundreds of images in a batch. It's the difference between hand-washing a pile of dishes and using a dishwasher.
• Accuracy: The researchers tested it on different animals (mice and rabbits) and different types of old and new microscope slides. It was as accurate as a human expert (98.7% agreement) but much faster and consistent.

The Bigger Picture

In glaucoma research, knowing just the number of broken wires isn't enough. Sometimes, the nerve gets "clogged" with too much protective padding, or the wires get too crowded even if they aren't broken yet.

MONICA gives scientists the full picture. It tells them not just how many wires are there, but how the whole cable is behaving. This helps them understand the disease better and test new drugs to stop the damage.

In short: MONICA is a free, easy-to-use web tool that automatically draws the outline of the eye's main nerve cable, allowing scientists to instantly measure its health, density, and structure without needing to be computer experts or spend hours drawing by hand.

Technical Summary

1. Problem Statement

Quantitative assessment of optic nerve health is critical for understanding glaucoma and evaluating neuroprotective interventions. While existing automated tools (e.g., AxonDeepSeg, AxoNet 2.0, AxonJ) excel at counting axons and segmenting individual axon morphometry, they suffer from a critical limitation: they do not automatically extract the whole optic nerve cross-sectional boundary.

This limitation creates several bottlenecks:

• Incomplete Metrics: Without the nerve boundary, researchers cannot calculate axon density (axons per unit area) or glial coverage fraction (non-axonal tissue ratio), which are increasingly recognized as having higher clinical relevance than total axon counts alone.
• Manual Labor & Inconsistency: Researchers must manually measure nerve area to derive density, introducing human error and labor-intensive workflows.
• High Barrier to Entry: Existing tools often require local software installation, command-line proficiency, or dedicated GPU hardware, limiting accessibility for non-computational biologists.

2. Methodology

The authors developed MONICA (Morphometrics from Optic Nerve Imaging Contour Analysis), a web-based application designed to automate the entire workflow from image upload to comprehensive morphometric analysis.

A. Core Architecture

• Platform: A GPU-accelerated web application accessible via browser (no local installation required).
• Backend: Python 3.11, utilizing PyTorch, nnU-Net, OpenCV, and AxonDeepSeg.
• Hardware: Server-side processing on NVIDIA RTX 3070 GPUs with 64GB RAM.

B. Technical Pipeline

1. Axon/Myelin Segmentation:

   • Utilizes AxonDeepSeg (v5.3.0) with a generalist bright-field model.
   • Performs three-class semantic segmentation: Axon (intra-axonal space), Myelin, and Background.
   • Uses nnU-Net with a cached predictor architecture for low-latency inference.

2. Morphology-Based Contour Extraction (Novel Algorithm):

   • Mask Consolidation: Merges binary axon and myelin masks into a unified "nerve tissue" mask.
   • Morphological Operations: Applies morphological closing (to fill gaps between adjacent units) and opening (to remove isolated noise pixels).
   • Boundary Detection: Uses OpenCV findContours to detect the outer boundary. If multiple contours exist, the one enclosing the largest area is selected.
   • Smoothing: Applies spline interpolation to generate anatomically plausible nerve outlines.
   • Filtering: Removes spurious segmentation detections (e.g., background artifacts) by filtering components whose centroids fall outside the derived nerve boundary.

3. Metric Calculation:

   • Summary Metrics: Cross-sectional area, total axon count, axon density, glial coverage area, and glial coverage fraction.
   • Per-Axon Metrics: Axon area, equivalent circular diameter, perimeter, circularity, eccentricity, g-ratio, myelin thickness, etc.

4. User Interface Features:

   • Supports batch processing and standard image formats (PNG, TIFF).
   • Includes a "Compare Settings" mode and five parameter presets (Default, Aggressive, Conservative, Scattered, Dense) for varying tissue qualities.
   • Offers an interactive contour editor allowing manual refinement via 200 draggable control points.
   • Provides a REST API and Python client for programmatic integration.

3. Validation Strategy

The system was validated against manual ground truth annotations created by an expert observer using polygon tools.

• Dataset: 15 bright-field microscopy images of PPD-stained, plastic-embedded optic nerve cross-sections.
• Diversity:
  • Taxonomic Orders: Mouse (Mus musculus) and Rabbit (Oryctolagus cuniculus).
  • Strains: BXD29 (contemporary, high-quality) and BXD51 (archival, variable quality).
  • Conditions: Varied staining intensity, tissue preservation, and nerve dimensions (mouse vs. larger rabbit nerves).

4. Key Results

• Segmentation Accuracy: MONICA demonstrated excellent agreement with manual ground truth.
  • Dice Similarity Coefficient (DSC): 0.987 ± 0.009 (overall mean).
  • Intersection over Union (IoU): 0.975 ± 0.017.
  • Precision/Recall: Balanced at 0.985 and 0.989 respectively, indicating no systematic over- or under-segmentation.
• Robustness:
  • Rabbit Samples: Highest performance (DSC = 0.994), suggesting scalability to larger nerves (potentially human).
  • Archival Samples: BXD51 samples (poor preservation/staining) achieved DSC = 0.987, proving robustness to image quality variations.
  • Lowest Performance: Even the worst-performing sample (BXD29 M639) achieved a DSC of 0.967.
• Metric Derivation: The tool successfully calculated axon density and glial coverage fraction, metrics previously inaccessible via automated pipelines.

5. Significance and Contributions

• Filling a Critical Gap: MONICA is the first tool to provide automated whole-nerve boundary extraction alongside axon segmentation, enabling the calculation of axon density and glial coverage without manual intervention.
• Accessibility: By running entirely in the browser with no local dependencies, it democratizes advanced morphometry for researchers lacking computational expertise or GPU hardware.
• Generalizability: The morphology-based approach is not anatomically specific, suggesting potential applicability to other myelinated nerve tracts (e.g., spinal cord, peripheral nerves).
• High-Throughput Capability: The batch processing and REST API support large-scale phenotyping studies in glaucoma genetics and drug discovery.
• Clinical Relevance: By enabling the analysis of glial coverage and axon density, MONICA supports the shift toward metrics that may correlate more strongly with visual function and intraocular pressure than simple axon counts.

Availability: The tool is freely accessible at monica.jablonskilab.org, with source code and data available upon request.