rai-audit-dl 0.1.9

Deep learning audits for image classification, medical imaging, and scientific AI

rai-audit-dl

Deep learning audits for image classification, medical imaging, and scientific AI.

Features

• Image classification accuracy and per-class recall checks
• Robustness checks for recorded or callback-generated transformation predictions
• Built-in brightness, contrast, horizontal-flip, and Gaussian-noise transformations
• Grad-CAM heatmaps with PyTorch hooks or TensorFlow GradientTape
• Medical imaging checks for patient leakage across splits and site-level accuracy bias
• Segmentation Dice, IoU, mask-quality, and subgroup checks
• Object-detection mAP, IoU-threshold, class-coverage, and subgroup checks
• DICOM metadata, scanner/protocol bias, near-duplicate leakage, calibration, and patient aggregation
• Grad-CAM transformation stability and optional localization-overlap checks

Python API

from rai_audit.dl import ImageClassificationAudit

report = ImageClassificationAudit(
    y_true=y_true,
    y_pred=y_pred,
    transformed_predictions={"sensor_noise": noisy_predictions},
    persist=False,
).run()

To evaluate transformations directly, pass images and a predictor callback.

CLI

Audit recorded predictions from CSV:

rai-audit dl run --data predictions.csv --task image --format html

Medical imaging CSV audits can add --patient-id, --split, and --site.

Examples

• examples/scientific_ai/microscopy_audit.py
• examples/medical_imaging/audit_example.py