fishROI: A specialized workflow for semi-automated muscle morphometry analysis in teleosts

This study introduces fishROI, a specialized FIJI2 plugin that combines cytoplasmic staining with deep learning to provide a robust, semi-automated workflow for analyzing and visualizing the unique muscle morphometry and hyperplastic growth dynamics of teleosts.

The Gist

Imagine you are a detective trying to count and measure the bricks in a massive, ever-changing wall. In the world of biology, this wall is muscle tissue, and the bricks are muscle fibers. Scientists have long wanted to understand how these walls grow, heal, or break down, but counting the bricks by hand is slow, boring, and prone to human error.

For a long time, scientists had a special "magic marker" (a specific stain) that worked perfectly on mammals (like humans and mice). It highlighted the edges of the bricks so computers could easily count them. But when they tried to use this same marker on fish (specifically zebrafish), it failed miserably. The fish bricks didn't light up the same way, and the computer got confused.

Worse yet, fish grow differently than mammals. Mammals mostly make their muscles bigger (like inflating a balloon), but fish keep adding new tiny bricks to the wall as they grow. This creates a chaotic mix of huge bricks and tiny, brand-new ones. Old computer programs, trained on the uniform bricks of mammals, would look at these tiny fish bricks and think, "That's just a speck of dust; ignore it." This meant scientists were missing a huge part of the story.

Enter "fishROI": The New Detective Tool

The authors of this paper built a new, specialized toolkit called fishROI (pronounced "fish-roy"). Think of it as a Swiss Army knife for fish muscle analysis that lives inside a free, popular image program called FIJI (which is like the "Photoshop" for scientists).

Here is how it works, broken down into simple steps:

1. The Right Lens (Cytoplasmic Staining)
Instead of trying to highlight the edges of the bricks (which was hard in fish), the researchers decided to color the inside of the bricks. They used a dye called phalloidin that fills the entire muscle fiber. It's like painting the inside of every brick a bright, solid color. Suddenly, even the tiniest, newest bricks are visible and distinct.

2. The Smart Brain (Deep Learning)
Once the bricks are colored, the tool uses Artificial Intelligence (AI) to do the counting.

• The "Off-the-Shelf" Brain: They found that a pre-made AI brain (called Cellpose) was already pretty good at recognizing these colored bricks, even without special training.
• The "Custom-Trained" Brain: For even better results, they taught the AI specifically how to look at zebrafish muscles. It's like taking a general driver's license test versus taking a specific course on driving a race car. The custom-trained AI became much better at ignoring "dust" (artifacts) and only counting real bricks.

3. The Cleanup Crew (Manual Refinement)
Even the smartest AI makes mistakes. Sometimes it might merge two bricks into one, or miss a tiny one. The fishROI tool gives scientists a user-friendly interface to quickly fix these errors. It's like having a "Find and Replace" button for your data, allowing you to delete bad counts or merge split ones with a few clicks.

4. The Heatmap (Seeing the Story)
This is the coolest part. Fish don't just grow evenly; they grow in specific zones. The tool can create a color-coded map (a heatmap) of the muscle.

• If an area has a mix of huge and tiny bricks, the map turns a specific color.
• This allows scientists to instantly see where the fish is growing new muscle (a process called hyperplasia).
• It's like looking at a weather map: instead of just knowing it's "raining," you can see exactly which neighborhoods are getting a storm.

Why Does This Matter?

Before this tool, studying fish muscle growth was like trying to read a book written in a language you don't speak, using a dictionary that only works for a different language.

fishROI changes the game by:

• Saving Time: What used to take days of manual counting now takes minutes.
• Being Accurate: It catches the tiny, new fibers that other tools miss.
• Being Flexible: It works on different types of fish, different ages, and different ways of preparing the samples.

In short, the authors built a universal translator for fish muscle. They combined a better way to "paint" the tissue with a smarter computer brain and a user-friendly interface. This allows scientists to finally understand the complex, dynamic way fish build their muscles, which could help us learn more about muscle growth, regeneration, and disease in general.

Technical Summary

1. Problem Statement

Quantitative histological analysis of skeletal muscle is essential for studying growth, regeneration, and disease. However, current automated tools face significant limitations when applied to teleosts (bony fish), particularly zebrafish:

• Species Specificity: Existing tools are optimized for mammalian models (e.g., mice, humans) and rely on specific extracellular matrix (ECM) or membrane markers (e.g., WGA) that perform poorly in teleost tissues, especially during early developmental stages.
• Morphological Heterogeneity: Mammalian algorithms assume relatively uniform fiber sizes. Teleosts exhibit mosaic hyperplasia (continuous addition of new, small fibers throughout development), creating a broad distribution of fiber sizes. Algorithms often misclassify these small, nascent fibers as imaging artifacts.
• Lack of Visualization Tools: There is a scarcity of tools capable of visualizing spatial organization and morphometric variability (heterogeneity) specific to teleost muscle growth dynamics.
• Workflow Fragmentation: Researchers currently lack a unified, user-friendly pipeline that integrates segmentation, quantification, and visualization for teleost-specific data.

2. Methodology

The authors developed fishROI, a modular plugin for the ImageJ2/FIJI platform (implemented in Jython) designed to streamline muscle morphometry analysis.

• Staining Strategy: The study evaluated cytoplasmic staining (using Phalloidin) as a superior alternative to ECM/membrane markers (like WGA) for teleosts. Phalloidin provides consistent labeling across all developmental stages, including early larvae where ECM markers are weak.
• Segmentation Approaches:
  • Shallow Learning: Utilized Labkit for rapid, resource-light segmentation. Users can train models on individual images within minutes.
  • Deep Learning: Utilized Cellpose (specifically the cyto3 model). The authors tested the pre-trained model and trained a custom, zebrafish-specific model ("rerio") using a dataset of 54 confocal images and 19,840 manually curated Regions of Interest (ROIs).
• Workflow Pipeline:
  1. Image Segmentation: Users can select Labkit or Cellpose. The plugin handles ROI generation from binary masks or probability maps.
  2. Manual Refinement: Tools for visualizing merged fibers (random coloring), bulk ROI removal, and autosaving to prevent data loss.
  3. Quantification & Visualization:
     • Morphometric Heatmaps: Color-coding ROIs based on parameters like area, circularity, and intensity.
     • Variation Heatmaps: A novel method to calculate the Coefficient of Variation (CoV) of fiber area within a sliding window (radius = 3x mean Feret diameter) to quantify local heterogeneity. This requires an external Julia script triggered by the plugin.
• Data Handling: The plugin supports diverse file formats via Bio-Formats and exports comprehensive morphometric data to spreadsheets for downstream analysis.

3. Key Results

• Staining Superiority: Phalloidin (cytoplasmic) staining consistently outperformed WGA (ECM) staining, particularly in larval stages (3–7 dpf) where ECM boundaries are indistinct.
• Segmentation Performance:
  • Labkit: Struggled to distinguish tightly packed fibers in early larval stages, leading to high false negatives. Performance improved in juvenile stages.
  • Cellpose (Pre-trained): The generic cyto3 model significantly outperformed Labkit without retraining, correctly segmenting ~80% of fibers in late larval/juvenile stages with a false positive rate <20%.
  • Cellpose (Custom Model): Training a custom "rerio" model on zebrafish data did not significantly increase the true positive rate but significantly reduced the false positive rate across all stages, improving the rejection of artifacts. A single universal model was found sufficient for most developmental stages.
• Quantification of Hyperplasia: The CoV-based heatmap successfully identified zones of mosaic hyperplasia. Juvenile zebrafish (active hyperplasia) showed significantly elevated CoV in the deep myotome compared to early larvae, validating the metric as a quantitative measure of growth dynamics.

4. Key Contributions

• fishROI Plugin: A free, open-source, modular FIJI2 plugin that integrates shallow and deep learning segmentation with specialized visualization tools for teleosts.
• Species-Agnostic Workflow: A flexible pipeline that accommodates various staining protocols, image qualities, and species, moving away from the rigid assumptions of mammalian-specific tools.
• Novel Quantitative Metric: Introduction of a local Coefficient of Variation (CoV) heatmap to objectively quantify mosaic hyperplasia and spatial heterogeneity, replacing subjective visual assessments.
• Accessibility: The tool bridges the gap between advanced machine learning (Cellpose) and biologists by providing a GUI, conversion tools for ROI formats, and clear instructions for training custom models.

5. Significance

This study addresses a critical bottleneck in comparative muscle biology. By validating cytoplasmic staining and deep learning segmentation for teleosts, fishROI enables high-throughput, accurate analysis of muscle growth, regeneration, and disease in zebrafish and other teleost models. The ability to quantitatively map mosaic hyperplasia opens new avenues for understanding developmental plasticity and muscle regeneration mechanisms that are distinct from mammalian hypertrophy. The modular design ensures the tool remains future-proof, capable of integrating emerging segmentation algorithms and adapting to diverse experimental needs.