ctkit

CT image processing for AI and radiomics. Makes CT processing simple and reproducible.

Installation

pip install ctkit

To install pyradiomics and TotalSegmentator:

pip install 'ctkit[all]'

Quick start

import ctkit

ctkit.download("tcga-kirc", "data/tcga_kirc_raw", limit=20)

data = ctkit.Dataset("data/tcga_kirc_raw")
data.filter(min_slices=25).process("tcga-kirc", out_dir="data/processed")
# the tcga-kirc protocol segments the kidneys, so this one needs ctkit[all]

The pipeline

Reproducibility

Every run that writes a cohort to disk writes processing_config.yaml next to it. To reproduce a dataset, or to hand one to a collaborator:

ctkit.Dataset("data/raw").process(
    "data/processed/processing_config.yaml", out_dir="rerun"
)

process takes a protocol as a ProcessingConfig, a path to a saved one, or the name of a collection whose curated protocol to use.

Notebooks

notebooks/quickstart.ipynb walks through the package end to end: pick a collection, download it, filter it, process it, and extract features.

notebooks/scanner_confounding.ipynb is the argument for the pipeline. A cohort is built on two scanners whose patient populations differ, so the scanner correlates with the label. A classifier trained on the raw volumes scores 0.89 on held-out data from the same study and 0.58 at another hospital, and on that external set its answers agree with the scanner 88% of the time — it learned the machine, and internal validation could not tell. The same classifier on the same scans put through a ctkit protocol scores 0.96 externally. One ablation per step shows what each is worth, and what preprocessing cannot fix. Runs in a few minutes, no GPU or download needed.

Relationship to tcia-radiology-processing

This package grew out of the protocol in pachterlab/tcia-radiology-processing, which documents the same pipeline as a step-by-step notebook. That repository remains the written protocol; ctkit is the library implementation of it.

License

BSD 2-Clause. See LICENSE.

Issues and pull requests welcome.