prost-t2-classification 0.1.0

T2-only fastMRI prostate coil-selection and real-vs-complex classification pipeline.

fastMRI Prostate T2 Coil Classification

This project builds a T2-only classification experiment for the fastMRI prostate dataset. It follows the dataset paper's slice-level PI-RADS classification setup and official patient split, but replaces the released RSS images with selected coil images from the middle T2 acquisition.

The experiment compares:

• a real-valued CNN that receives only coil image amplitudes;
• a complex-valued CNN that receives the same selected coils as complex images.

PI-RADS labels are binarized as in the paper: PI-RADS > 2 is clinically significant prostate cancer.

Install

From PyPI:

python -m pip install prost-t2-classification

Confirm the console command is available:

prost-t2 --help

For development from a local checkout:

python -m venv .venv
.\.venv\Scripts\python -m pip install -U pip
.\.venv\Scripts\python -m pip install -e ".[dev]"

For CUDA training, install the PyTorch build that matches your GPU/driver before or after the editable install.

Full Pipeline

The full command prompts for storage locations if you omit them:

prost-t2 run --download-script .\prostate_download_script.txt

Equivalent non-interactive form:

prost-t2 run `
  --download-script .\prostate_download_script.txt `
  --download-dir D:\fastmri_prostate\archives `
  --extract-dir D:\fastmri_prostate\raw `
  --recon-dir D:\fastmri_prostate\recon_t2 `
  --npz-dir D:\fastmri_prostate\npz_t2_coils `
  --runs-dir D:\fastmri_prostate\runs

To stop after downloading, reconstruction, and NPZ preparation, skip training:

prost-t2 run `
  --download-script .\prostate_download_script.txt `
  --download-dir D:\fastmri_prostate\archives `
  --extract-dir D:\fastmri_prostate\raw `
  --recon-dir D:\fastmri_prostate\recon_t2 `
  --npz-dir D:\fastmri_prostate\npz_t2_coils `
  --skip-train

Individual Stages

Download labels and T2 tarballs only:

prost-t2 download --download-script .\prostate_download_script.txt --download-dir D:\fastmri_prostate\archives --no-extract

Download and extract labels plus T2 tarballs:

prost-t2 download --download-script .\prostate_download_script.txt --download-dir D:\fastmri_prostate\archives --extract-dir D:\fastmri_prostate\raw

Run GRAPPA/IFFT reconstruction with fastmri-tools:

prost-t2 reconstruct --raw-root D:\fastmri_prostate\raw --recon-dir D:\fastmri_prostate\recon_t2

Create compact NPZ samples from the middle acquisition and top-energy coils:

prost-t2 make-npz --labels D:\fastmri_prostate\raw --recon-dir D:\fastmri_prostate\recon_t2 --npz-dir D:\fastmri_prostate\npz_t2_coils

Prepare NPZ files from extracted raw data without training:

prost-t2 prepare-npz `
  --raw-root D:\fastmri_prostate\raw `
  --recon-dir D:\fastmri_prostate\recon_t2 `
  --npz-dir D:\fastmri_prostate\npz_t2_coils

Train both models:

prost-t2 train --manifest D:\fastmri_prostate\npz_t2_coils\manifest.csv --runs-dir D:\fastmri_prostate\runs --mode both

Data Decisions

• The official data_split column is used directly, and patient leakage across train/validation/test is checked before training.
• T2 kspace is reconstructed through fastmri-tools, producing complex image_complex arrays.
• The acquisition dimension is reduced by selecting the middle acquisition (shape[0] // 2).
• Up to five coils are selected per patient volume using highest image-space energy, measured on the selected acquisition across all slices.
• NPZ files store image_complex with shape (coils, height, width) plus patient, slice, split, and coil metadata.

Publishing

Releases are published to PyPI by GitHub Actions using PyPI Trusted Publishing, so the repository does not need a long-lived PyPI API token.

Create a pending publisher on PyPI with these values:

• PyPI project name: prost-t2-classification
• GitHub owner: meis-01
• GitHub repository: prost_t2_classification
• Workflow file: publish.yml
• GitHub environment: pypi

Then bump __version__ in src/prost_t2_classification/__init__.py, commit the change, and push a matching tag:

git tag v0.1.0
git push origin v0.1.0

The publish workflow checks that the tag matches the package version before it uploads the wheel and source distribution.