This project provides tools for preprocessing, filtering, labeling, and classifying high-resolution 3D cortical blood vessel datasets. It leverages semi-supervised learning, Zarr (Format 2), Dask, Napari, and various image processing techniques.
- Efficient loading of large volumetric datasets in Zarr format
- Filtering and per-voxel feature expansion using Frangi, Sato, Meijering vesselness filters
- Semi-supervised learning using SGDClassifier (via Dask-ML)
- Napari GUI integration for labeling and previewing predictions
Clone the repository using
git clone https://github.com/LandoRaDag/Capstone.git
Install dependencies using:
pip install -r requirements.txtRun the following command to launch the main GUI:
python python-code/main.py
Within the interface, you may:
-
Preprocessing Convert your .tiff file into a Zarr dataset. This step is compute-heavy — enabling GPU acceleration with CuPy (for NVIDIA) or PyOpenCL (for AMD) is recommended for speed.
-
Feature Expansion Embedded within preprocessing. Filters from make_dataset.py (Frangi, Sato, etc.) are applied to generate multi-channel voxel-level features. These features become the basis for classification.
-
Annotation & Segmentation Napari will open for manual labeling. Use the "Labels" layer: Red (1) for background, Green (2) for vessels
Once labeled, click the Predict button. An SGDClassifier will train and classify the rest of the image. Predictions will appear in the "predictions" layer.
Designed for research and prototyping, not production use.
Labels generated can be used to train downstream models like U-Net.
Feel free to expand or modify the pipeline to suit your specific imaging tasks
Developed by Charles Blancas, Landon Ra Dagenais, Imad Baida, and Mohamed Elsamadouny and team as part of a McGill capstone project in computer engineering and machine learning, under supervision from Professor Amir Shmuel of the Neuro Lab.