Configuration guide

AUDIT uses YAML configuration files for feature extraction, metric evaluation, and the app. The files live in src/audit/configs/.

Every workflow separates the clinical modality from the file format:

modality: MRI
formats:
  image: .nii.gz
  segmentation: .nii.gz
  prediction: .nii.gz

modality selects domain behaviour. formats selects the storage reader. See Image formats and readers for all supported values and reader options.

Feature extraction

data_paths:
  BraTS: /data/BraTS/images

modality: MRI
formats:
  image: .nii.gz
  segmentation: .nii.gz

sequences:
  - _t1
  - _t2
  - _t1ce
  - _flair

labels:
  BKG: 0
  EDE: 3
  ENH: 1
  NEC: 2

features:
  statistical: true
  texture: true
  spatial: true
  tumor: true

preprocessing: {}

output_path: ./outputs/features
logs_path: ./logs/features
cpu_cores: 8

MRI requires an explicit sequence list. A single-image US dataset normally uses sequences: [], modality: US, and a raster or DICOM format.

preprocessing is passed to the selected modality before feature extraction. For example, US currently supports optional min-max normalization:

preprocessing:
  normalize: true

Metric extraction

data_path: /data/BraTS/images

modality: MRI
formats:
  segmentation: .nii.gz
  prediction: .nii.gz

model_predictions_paths:
  nnUNet: /data/BraTS/predictions/nnUNet

labels:
  BKG: 0
  EDE: 3
  ENH: 1
  NEC: 2

backend: audit

metrics:
  dice: true
  jaccard: true
  accuracy: true
  precision: true
  sensitivity: true
  specificity: true
  hausdorff_distance: true
  predicted_size: true

output_path: ./outputs/metrics
filename: BraTS
logs_path: ./logs/metric
cpu_cores: 8

The metric names above are the canonical keys for backend: audit. pymia and Metrics Reloaded expose their own documented metric keys so their additional capabilities remain available. Consult the metric availability table before switching backends.

App

modality: US
formats:
  image: .png
  segmentation: .png
  prediction: .png
sequences: []

labels:
  BKG: 0
  TUMOR: 255

datasets_path: ./datasets/US
features_path: ./outputs/features
metrics_path: ./outputs/metrics

raw_datasets:
  BUSI: "${datasets_path}/BUSI/images"

features:
  BUSI: "${features_path}/extracted_information_BUSI.csv"

metrics:
  BUSI: "${metrics_path}/extracted_information_BUSI.csv"

predictions:
  BUSI:
    nnUNet: "${datasets_path}/BUSI/predictions/nnUNet"

The app currently uses one global modality, format mapping, and label map. Mixed MRI/US datasets and dataset-specific labels are not supported in one execution.

Tip

Keep experimental dataset paths in a separate example or local YAML file. The general configuration should describe the reusable workflow rather than a particular experiment.