DICOM Trust Boundary Research

Toward Boundary-Aware DICOM Ingestion

Active Research Program Updated

This program studies data systems for medical imaging, with emphasis on DICOM ingestion, structural parsing, and where trust and policy boundaries belong in imaging pipelines. It asks a narrower version of a general question: how much of a sensitive data object must be interpreted before a system can decide what it is permitted to keep — and whether that decision can be made before the object is persisted at all.

Latest publication: Don't Land What You Don't Want to Own — Moving the DICOM trust boundary ahead of persistence

Research Thesis

For sufficiently sensitive input, the safest place to remove what shouldn't be retained may be before the persistence boundary, not after it — and that decision is only affordable if structural inspection can be done without fully materializing the object.

The work begins with structural parsing: understanding a DICOM object's tag, sequence, and bulk-data layout well enough to inspect and selectively transform it without decoding data that doesn't need to be interpreted. From there it moves toward the harder architectural question — what a pre-persistence trust boundary should look like once metadata-level policy, private-element handling, and pixel-level risk are all in play.

Guiding Principles

Don't land what you don't want to own

The safest place to remove sensitive data may be the point immediately before infrastructure ever has the opportunity to retain it.

Don't decode what you don't need

Structural inspection should not require materializing data — such as Pixel Data — that a policy doesn't require touching.

Structure before interpretation

A deterministic structural scan — tags, offsets, sequence and item boundaries — can support policy decisions without a full object model.

Irreversibility is a cost, not a default

Pre-persistence transformation trades recoverability for containment. That trade should be made deliberately, with bounded quarantine and auditable policy, not assumed.

Metadata policy is not the whole problem

Removing identifying metadata does not address identifying information that may be burned into image pixels; that is a distinct, unsolved problem this program does not claim to address.

Publication Roadmap

Published

01 — Don't Land What You Don't Want to Own

Moving the DICOM trust boundary ahead of persistence: pre-persistence structural inspection and policy transformation, validated against ~26,600 real-world CT DICOM objects with fastDICOMstructure.

This program studies trust and policy boundaries in medical-imaging data systems, as a peer to the Autonomous DataOps Research Series. Both examine how data systems behave at their boundaries — what evidence they emit, and what they should refuse to persist — from different sides of that question.

fastDICOMstructure on GitHub

The experimental C++/Python structural parser this program's first publication is built on, including corpus-validation tooling and benchmark methodology.