gfap-segmenter 1.0.0

A Bachelor Thesis project to allow Segmentation of Astrocytes on GFAP Staining using Machine Learning in napari

gfap-segmenter

A Bachelor Thesis project to allow segmentation of astrocytes on GFAP staining using machine learning in napari.

---

Features

• Load .tif/.tiff/.czi microscopy data via bioio and display them in napari.
• Run tiled U-Net inference that produces probability maps and binary masks.
• Curate training patches with ≥300×300 validation and immediate colour feedback.
• Synthesize 1000 augmented patches per curated region and train in the background with progress updates.
• Import/export checkpoints (including ONNX) and optionally bundle curated training data.

Installation

Install gfap-segmenter via pip:

pip install gfap-segmenter

If napari (Qt backend included) is not yet installed, use:

pip install "gfap-segmenter[all]"

GPU support

The plugin automatically uses GPU acceleration when available. GPU support significantly speeds up both model training and inference (typically 10-50x faster than CPU).

Installation

The plugin relies on PyTorch. By default, pip install gfap-segmenter installs the CPU-only version of PyTorch. To enable GPU acceleration, you need a CUDA-enabled PyTorch build.

Important: PyTorch bundles its own CUDA runtime, so you don't need to install the full CUDA toolkit separately. You only need NVIDIA GPU drivers (usually pre-installed on systems with NVIDIA GPUs).

For NVIDIA GPUs (CUDA):

Option 1: Fresh installation (recommended)

1. Install PyTorch with CUDA support from PyTorch's official installation guide:

   CUDA 11.8:

   pip install torch --index-url https://download.pytorch.org/whl/cu118
   

   CUDA 12.4:

   pip install torch --index-url https://download.pytorch.org/whl/cu124
   

   CUDA 12.8:

   pip install torch --index-url https://download.pytorch.org/whl/cu128
   

   CUDA 13.0:

   pip install torch --index-url https://download.pytorch.org/whl/cu130
   

2. Then install gfap-segmenter:

   pip install gfap-segmenter
   

Option 2: Upgrade existing CPU-only PyTorch installation

If you already have gfap-segmenter installed with CPU-only PyTorch, you can upgrade to CUDA-enabled PyTorch:

1. Uninstall the CPU-only PyTorch:

   pip uninstall torch
   

2. Install CUDA-enabled PyTorch (choose the version matching your driver):

   pip install torch --index-url https://download.pytorch.org/whl/cu121
   

3. The plugin will automatically use the CUDA-enabled PyTorch - no need to reinstall gfap-segmenter.

Note: You don't need to match the exact CUDA version. PyTorch includes its own CUDA runtime. As long as your NVIDIA driver is compatible (typically driver 11.0+ for CUDA 11.x, or 12.0+ for CUDA 12.x), you can use any recent PyTorch CUDA build. If unsure, use CUDA 12.1.

For Apple Silicon (macOS):

MPS (Metal Performance Shaders) support is automatically available if you have PyTorch 1.12+ installed. No additional steps required - the plugin will automatically use your Apple Silicon GPU.

Verification

To verify that GPU support is enabled:

1. Check in the plugin UI: The status bar at the bottom of the plugin widget shows the current device (e.g., "Device: CUDA (NVIDIA GeForce RTX 3090)" or "Device: CPU").

2. Check via Python:

   import torch
   print(f"CUDA available: {torch.cuda.is_available()}")
   if torch.cuda.is_available():
       print(f"CUDA device: {torch.cuda.get_device_name(0)}")
       print(f"CUDA version: {torch.version.cuda}")
   

3. Check PyTorch CUDA support (command line):

   python -c "import torch; print('CUDA available:', torch.cuda.is_available()); print('CUDA version:', torch.version.cuda if torch.cuda.is_available() else 'N/A')"
   

Checking your NVIDIA driver version (optional, for reference):

• Using nvidia-smi (if NVIDIA drivers are installed):

  nvidia-smi
  

  Look for the "CUDA Version" line (shows the maximum CUDA version your driver supports).

• Using nvcc (if CUDA toolkit is installed):

  nvcc --version
  

• Using PyTorch (after PyTorch is installed):

  python -c "import torch; print(torch.version.cuda)"
  

Troubleshooting

Problem: Plugin shows "Device: CPU" even though you have a GPU

• PyTorch CPU-only build: You may have installed the CPU-only version of PyTorch. Uninstall and reinstall with CUDA support:

  pip uninstall torch
  pip install torch --index-url https://download.pytorch.org/whl/cu121  # Adjust CUDA version
  

• CUDA driver mismatch: Your CUDA driver version may not match the PyTorch CUDA version. Check compatibility:

  nvidia-smi  # Shows driver version
  python -c "import torch; print(torch.version.cuda)"  # Shows PyTorch CUDA version
  

  The driver version should be >= the PyTorch CUDA version.

• CUDA not detected: Ensure your NVIDIA drivers are up to date and CUDA is properly installed.

Problem: "CUDA out of memory" errors

• Reduce batch size in training or prediction settings
• Use a smaller patch size
• Close other GPU-intensive applications

Problem: Training/inference is still slow

• Verify GPU is actually being used (check status bar or console output)
• Ensure you installed the correct CUDA version for your GPU
• Check GPU utilization: nvidia-smi should show high GPU usage during training

Performance Notes

• The plugin automatically uses GPU for both training and inference when available - no configuration needed. The current device is displayed at the bottom of the plugin.
• If no GPU is detected, the plugin automatically falls back to CPU execution (slower but functional).

Development version

pip install git+https://github.com/MMV-Lab/gfap-segmenter.git

Usage overview

1. Prediction tab – load the default checkpoint or import a custom one, choose an image layer, and run inference to create probability and mask layers.
2. Curate Patches tab – draw rectangles in the dedicated shapes layer. Patches ≥300×300 are tinted blue; smaller ones are red. Add valid regions to the training set.
3. Train tab – launch background training; progress and logs stream into the widget. Trained models are stored under ~/.gfap_segmenter/models.
4. Model IO tab – export/import checkpoint bundles. Bundles can optionally include the curated training data for reproducibility.

Contributing

Contributions are very welcome. Tests can be run with tox; please ensure coverage at least stays the same before submitting a pull request.

License

Distributed under the terms of the BSD-3 license, "gfap-segmenter" is free and open source software.

Issues

If you encounter any problems, please file an issue along with a detailed description.