CT Phase Classification Pipeline

This project provides a pipeline for automatic CT contrast phase classification using organ segmentation, radiomics feature extraction, and a trained machine learning model.

The pipeline takes a CT scan as input, segments selected anatomical structures using TotalSegmentator, extracts radiomic features from the segmented regions, and uses a trained XGBoost classifier to predict the CT phase.

---

Predicted CT Phases

The model classifies scans into the four main CT contrast phases:

• NP: Non-Contrast Phase
• AP: Arterial Phase
• VP: Portal-Venous Phase
• DP: Delayed Phase

---

Pipeline Overview

The pipeline consists of the following steps:

1. Read a CT image from a DICOM folder or a NIfTI file.
2. Clip CT intensity values to the range [-200, 200].
3. Segment selected anatomical structures using TotalSegmentator.
4. Extract radiomic features from each segmentation mask using PyRadiomics.
5. Calculate selected organ-level and organ-difference features.
6. Predict the CT contrast phase using the trained XGBoost model.

The following anatomical structures are used for feature extraction:

• Aorta
• Portal vein and splenic vein
• Urinary bladder
• Kidneys, using both left and right kidney masks
• Spleen
• Liver

---

Project Files

• simple_phase_pipeline.py
  Runs the full prediction pipeline for one CT scan.

• feature_extractor.py
  Reads the CT image, runs segmentation, extracts radiomic features, and prepares the final feature table.

• run_totalsegmentator.py
  Runs TotalSegmentator for the selected anatomical structures.

• final_model.pkl
  Trained XGBoost model used for CT phase classification.

• features.pkl
  List of selected features expected by the trained model.

• label_encoder.pkl
  Label encoder used to convert between numeric model outputs and CT phase labels.

• requirements.txt
  Python packages needed to run the project.

---

Requirements

This project requires Python 3.10 or 3.11.

Install the required packages using:

pip install -r requirements.txt

The main packages used in this project are:

• TotalSegmentator
• PyRadiomics
• SimpleITK
• pandas
• scikit-learn
• XGBoost
• joblib

---

How to Run

Run the pipeline from the command line:

python simple_phase_pipeline.py "input_path"

The input_path can be either:

• a DICOM folder
• a NIfTI file, for example .nii.gz

Example with a NIfTI file:

python simple_phase_pipeline.py "path/to/ct_scan.nii.gz"

Example with a DICOM folder:

python simple_phase_pipeline.py "path/to/dicom_folder"

---

Optional Arguments

Use the fast TotalSegmentator option:

python simple_phase_pipeline.py "input_path" --fast

Use a local TotalSegmentator model directory:

python simple_phase_pipeline.py "input_path" --totalseg-model-dir "path/to/totalseg_models"

Save the extracted model features to a CSV file:

python simple_phase_pipeline.py "input_path" --save-features extracted_features.csv

---

Output

The pipeline prints:

• the predicted CT phase
• class probabilities, if available
• a feature check showing the number of expected and extracted features

If --save-features is used, the model input features are also saved as a CSV file.

---

References

• TotalSegmentator: https://github.com/wasserth/TotalSegmentator
• PyRadiomics: https://pyradiomics.readthedocs.io/
• Wasserthal et al., TotalSegmentator: Robust segmentation of 104 anatomical structures in CT images, Radiology: AI (2023)