Entangled Filament Detection

Exploring methodologies for detecting and annotating highly entangled filamentous instances in images with partial annotations.

Project Overview

This project investigates different approaches for detecting thin, continuous filamentous structures (like neurons, fibers, or roads) that exhibit:

• Crossings and occlusions
• Partial annotations (non-exhaustive centerline annotations)
• No reliable negative background

Key Constraints

• Never penalize for failing to detect unannotated filaments
• Work with polyline/centerline annotations
• Handle variable filament thickness and blur

Installation

# Clone the repository
git clone https://github.com/5TuX/entangled-filament-detection.git
cd entangled-filament-detection

# Create virtual environment
uv venv
source .venv/bin/activate  # Linux/Mac
# or
.venv\Scripts\activate  # Windows

# Install dependencies
uv pip install -e .

Quick Start

Generate Synthetic Data

from synthetic.generator import FilamentGenerator, GeneratorConfig

config = GeneratorConfig(
    image_size=(512, 512),
    num_filaments=(50, 100),
    blur_sigma_range=(0.5, 2.0),
    convergence_zones=(2, 5),
    seed=42
)

generator = FilamentGenerator(config)
result = generator.generate_image_with_annotations()

# Access generated data
image = result['image']
annotated_polylines = result['annotated_polylines']
unannotated_polylines = result['unannotated_polylines']

Run Visualizer

streamlit run notebooks/visualizer_app.py

Datasets

Identified Public Datasets

Dataset	Domain	Annotation Type
FISBe (CVPR 2024)	Neuron microscopy	Instance masks
CREMI	EM neurites	Instance IDs
DeepGlobe Roads	Satellite imagery	Centerlines
Fluorescent Neuronal Cells v2	Microscopy	Polylines, masks

Synthetic Data Generator

The FilamentGenerator class creates realistic entangled filament images with:

• Configurable blur and noise
• Convergence zones (high-density areas)
• Partial annotation simulation
• Variable filament thickness

Methodologies

1. Direct Polyline Regression - CNN + sequential point prediction
2. DETR-like Set Prediction - Transformer with learnable queries
3. Segmentation + Embedding - Shared encoder with clustering
4. PU Learning - Positive-Unlabeled learning for partial annotations

Project Structure

entangled-filament-detection/
├── data/
│   ├── synthetic/           # Generated synthetic data
│   └── fisbe/              # Downloaded FISBe data
├── src/
│   ├── models/             # Model implementations
│   ├── losses/             # Custom loss functions
│   └── utils/              # Utilities
├── synthetic/
│   └── generator.py        # Synthetic data generator
├── notebooks/
│   └── visualizer_app.py   # Streamlit visualizer
├── configs/                 # Configuration files
└── pyproject.toml          # Dependencies

Evaluation Metrics

• Annotated Recall - Detection rate on annotated filaments
• Geometric Accuracy - Chamfer/Fréchet distance
• Structural Error - Miss rate, fragmentation

License

MIT License

References

• FISBe: A Real-World Benchmark Dataset for Instance Segmentation of Long-Range Thin Filamentous Structures (CVPR 2024)
• TREXplorer: Recurrent DETR for Topologically Correct Tree Centerline Tracking (MICCAI 2024)
• PU Learning surveys and selective labeling methods