Process case folders with the batch API

Use a directory containing one subfolder per case, with a consistently named NIfTI image and ROI mask in each. This example preprocesses the cases, extracts features into one CSV file, and checks for failed cases or skipped structures. Replace the example paths and names with your dataset layout; see Expected data structure.

import csv
from pathlib import Path

from zrad.batch import BatchPreprocessor, BatchRadiomicsExtractor

# Each case folder contains phantom.nii.gz and GTV-1.nii.gz.
preprocessed = BatchPreprocessor(
    input_directory="path/to/nifti_cases",
    output_directory="output/preprocessed",
    input_data_type="nifti",
    modality="CT",
    nifti_image_name="phantom",
    structures=["GTV-1"],
    resample_resolution=2.0,
    resample_dimension="3D",
    image_interpolation_method="linear",
    mask_interpolation_method="linear",
    mask_interpolation_threshold=0.5,
).run()

extracted = BatchRadiomicsExtractor(
    input_directory="output/preprocessed",
    output_directory="output/features",
    input_data_type="nifti",
    modality="CT",
    nifti_image_name="image",  # Name written by BatchPreprocessor.
    structures=["GTV-1"],
    aggregation_dimension="3D",
    aggregation_method="AVER",
    discretization_method="Number of Bins",
    number_of_bins=32,
).run()

# A processed case can still contain a skipped structure.
for stage, result in (("preprocessing", preprocessed), ("radiomics", extracted)):
    print(f"{stage}: {result.processed_count} processed, {result.failed_count} failed")
    for case in result.case_results:
        if case.error or case.skipped_structures:
            print(case.case_name, case.error or case.skipped_structures)

with Path("output/features/radiomics.csv").open(newline="") as csv_file:
    for row in csv.DictReader(csv_file):
        print(row["pat_id"], row["mask_id"])

Batch preprocessing writes one image and mask per case under output/preprocessed. Extraction writes output/features/radiomics.csv. Check the output rows against the requested cases and structures; see Python batch workflows and Understanding results.