🔬 Cell Segmentation & Tracking

Interactive web app for cell segmentation and tracking using Cellpose + LAP tracking.

✨ Now with Apple M4 GPU Support!

🚀 Quick Start

Installation

# Clone repository
git clone https://github.com/YOUR_USERNAME/flim_tracker.git
cd flim_tracker

# Create conda environment
conda create -n flim python=3.10
conda activate flim

# Install dependencies
pip install -r requirements.txt

Run App

streamlit run streamlit_app.py

Open http://localhost:8501 in your browser.

💻 GPU Support

Apple Silicon (M4/M3/M2/M1)

✅ Automatic M4 Pro GPU acceleration via MPS (Metal Performance Shaders)

• 5-10x faster than CPU
• ~10-30 seconds per frame (vs 2-3 minutes on CPU)

NVIDIA GPUs

✅ Automatic CUDA GPU acceleration

• ~1-2 seconds per frame

CPU Fallback

⚠️ Automatically uses CPU if no GPU available (slower)

📊 Features

Interactive Parameter Tuning

• Real-time Preview: See segmentation results instantly
• Compact UI: No scrolling, all controls visible
• Video Playback: Play/pause frame-by-frame animation
• Arrow Navigation: Previous/Next buttons for frame selection

Preprocessing Options

• Brightness adjustment (0.1-3.0x)
• Contrast normalization (stretch 0-255)
• CLAHE enhancement
• Gamma correction

Cellpose Parameters

• Diameter: Expected cell size (10-60px)
• Flow Threshold: Cell boundary detection strictness (0-1)
• Cell Prob Threshold: Detection permissiveness (-6 to 6)

Filtering

• Min/Max cell area
• Minimum intensity threshold

Tracking

• LAP (Linear Assignment Problem) algorithm
• Max linking distance
• Gap frame bridging (handle temporary occlusions)

Export

• Download tracking CSV with:
  • Track IDs
  • Frame numbers
  • Cell centroids (x, y)
  • Cell areas
  • Mean intensities

📁 Data Format

Upload a ZIP file containing microscopy images:

• Supported formats: TIF, TIFF, PNG, JPG, JPEG
• Images should be grayscale
• Consistent dimensions across frames

🎯 Workflow

1. Upload Data: Drag & drop ZIP of images
2. Tune Parameters: Adjust preprocessing, Cellpose, and filtering in sidebar
3. Preview: Navigate frames with ⬅️/➡️ arrows
4. Play Video: Click ▶️ to see animated tracking preview
5. Run Pipeline: Process all frames with full tracking
6. Download CSV: Export results for analysis

📈 Performance

Hardware	Speed per Frame	Full Dataset (15 frames)
Apple M4 Pro	10-30 seconds	~3-7 minutes
NVIDIA T4	1-2 seconds	~15-30 seconds
CPU (Intel/AMD)	2-3 minutes	~30-45 minutes

🔧 Technical Details

• Segmentation: Cellpose v3.0+ (cyto2 model)
• Tracking: LAP algorithm with gap closing
• UI: Streamlit with custom compact CSS
• GPU: PyTorch MPS (Apple Silicon) or CUDA (NVIDIA)

📦 Example Data

Sample PBS_1 dataset included: pbs1_frames.zip (15 frames, 1.8MB)

🐛 Troubleshooting

"No cells detected"

• Try enabling Contrast Normalize (dim images)
• Lower Cell Prob Threshold (more permissive)
• Adjust Diameter to match your cell size

Slow performance on Mac

• Verify GPU is enabled: Look for "🚀 GPU: Apple M4 Pro (MPS)" message
• If showing "💻 CPU Mode": Reinstall PyTorch with MPS support

  pip uninstall torch torchvision
  pip install torch torchvision torchaudio

Too many false positives

• Increase Flow Threshold (stricter boundaries)
• Increase Cell Prob Threshold (stricter detection)
• Adjust Min/Max Area filters
• Increase Min Intensity threshold

Broken tracks

• Increase Max Linking Distance (cells move farther between frames)
• Increase Max Gap Frames (handle longer occlusions)

🙏 Credits

Built with Cellpose, Streamlit, and PyTorch

📄 License

MIT License