Anonymization
AI/ML training corpus prep
Build a de-identified training corpus for AI/ML model development.
Problem
You are building a training dataset for an AI/ML model. The DICOM files must be de-identified and optionally converted to a non-DICOM format (e.g. NIfTI) for ingestion into your ML pipeline.
Steps
- Open a DICOM file. Click Files on the start screen (⌘O) to load a representative sample.
- Switch to Anonymization mode. Click the Anon tab (Ctrl+3 / ⌘3 in the desktop app, Ctrl+Shift+A / ⌘⇧A in the browser).
- Configure anonymization. Select the profile your IRB protocol requires: Basic, +Pixel, HIPAA, Teaching or Custom. There is no Limited Data Set profile. If your protocol allows a limited data set that keeps dates, combine a profile with a date shift as described in Limited Dataset (dates shifted).
- Apply. Click Apply & Export to generate a de-identified copy.
- Convert to ML format. Use the Export menu to convert the series to NIfTI-1 (
.nii.gz). Verify the output preserves spatial geometry (affine matrix) and volume dimensions. - Process the whole corpus. In the desktop app with Pro, Anonymize a whole folder of studies applies one profile to every DICOM file in a folder, and Anonymize a folder from Terminal on a Mac starts the same run from Terminal. In the browser, and in the desktop app without Pro, you anonymize one file at a time. A folder run does not convert to NIfTI, so convert its output afterwards.
Result
Each study is de-identified consistently across the corpus and converts to NIfTI-1 with its spatial orientation and volume dimensions intact.