Minimal Dicom de-identification

Description

This pipeline performs the least intrusive form of DICOM de-identification, removing only the most critical personal identifiers while preserving as much metadata as possible. It ensures that Protected Health Information (PHI) is stripped from both the image and metadata, but all non-sensitive details remain intact for research and analysis.

Minimal removal: Eliminates only Personally Identifiable Information (PII) from images and the most essential metadata fields while keeping the majority of the DICOM tags untouched.

Predicted Entities

Live Demo Open in Colab

How to use

dicom_df = spark.read.format("binaryFile").load(dicom_path)

pipeline = PretrainedPipeline("dicom_deid_generic_augmented_minimal", "en", "clinical/ocr")

result = pipeline.transform(dicom_df).cache()

val dicom_df = spark.read.format("binaryFile").load(dicom_path)

val pipeline = new PretrainedPipeline("dicom_deid_generic_augmented_minimal", "en", "clinical/ocr")

val result = pipeline.transform(dicom_df).cache()

Example

Input:

Output:

Model Information