autorad 0.1.2

Simple pipeline for experimenting with radiomics features

Installation from source

git clone https://github.com/pwoznicki/AutoRadiomics.git
cd AutoRadiomics
pip install -e .

Example - Hydronephrosis detection from CT images:

Extract radiomics features

df = pd.read_csv(table_dir / "paths.csv")
image_dataset = ImageDataset(
    df=df,
    image_colname="image path",
    mask_colname="mask path",
    ID_colname="patient ID"
)
extractor = FeatureExtractor(
    dataset=image_dataset,
    out_path=(table_dir / "features.csv"),
)
extractor.extract_features()

Load, split and preprocess extracted features

# Create a dataset from the radiomics features
feature_df = pd.read_csv(table_dir / "features.csv")
feature_dataset = FeatureDataset(
    dataframe=feature_df,
    target="Hydronephrosis",
    task_name="Hydronephrosis detection"
)

# Split data and load splits
splits_path = result_dir / "splits.json"
feature_dataset.full_split(save_path=splits_path)
feature_dataset.load_splits_from_json(splits_path)

# Preprocessing
preprocessor = Preprocessor(
    normalize=True,
    feature_selection_method="boruta",
    oversampling_method="SMOTE",
)
feature_dataset._data = preprocessor.fit_transform(dataset.data)

Train the model for hydronephrosis classification

# Select classifiers to compare
classifier_names = [
    "Gaussian Process Classifier",
    "Logistic Regression",
    "SVM",
    "Random Forest",
    "XGBoost",
]
classifiers = [MLClassifier.from_sklearn(name) for name in classifier_names]

model = MLClassifier.from_sklearn(name="Random Forest")
model.set_optimizer("optuna", n_trials=5)

trainer = Trainer(
    dataset=dataset,
    models=[model],
    result_dir=result_dir,
    experiment_name="Hydronephrosis detection"
)
trainer.run()

Create an evaluator to train and evaluate selected classifiers

evaluator = Evaluator(dataset=data, models=classifiers)
evaluator.evaluate_cross_validation()
evaluator.boxplot_by_class()
evaluator.plot_all_cross_validation()
evaluator.plot_test()

Commands

MLFlow

mlflow server -h 0.0.0.0 -p 5000 --backend-store-uri <result_dir>

Dependencies:

• MONAI
• pyRadiomics
• MLFlow
• Optuna
• scikit-learn
• imbalanced-learn
• XGBoost
• Boruta
• Medpy
• NiBabel
• nilearn
• plotly
• seaborn

App dependencies:

• Streamlit
• Docker