Medical Image Anonymization via Anatomical Region Modification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing medical imaging algorithms face challenges in anonymizing data used for training and validation, as identifiable features in image data can reveal the examination subject, limiting their use and consent compliance.

Innovation Solution

A computer-implemented method that modifies image or raw data to obscure anatomical areas easily identifiable, such as faces or internal organs, generating output data that prevents or significantly hinders identification, thus anonymizing the data effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If raw data or image data are used for training or validation, then the data reflects real examination objects and provides high training value, but the examination object can be identified from the data content

Engineering Contradiction:
Improvetraining data qualityVSAvoididentification risk
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent extracts and removes identifying features from medical image data while preserving the anatomical structures needed for algorithm training. This is achieved through automated detection and modification of identifiable anatomical regions, separating the useful training content from the harmful identification information.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies preliminary processing to counteract identification before the data is used for training. By pre-identifying and modifying potentially identifying features in the source data, the system prevents identification risks from arising in the first place, rather than attempting to address them after training.

Inventive Principle:
Principle #9Preliminary anti-action

2Object-affected harmful factors

If metadata such as person's name, birth dates are deleted to anonymize data, then personal data protection is improved, but identification can still occur through image content

Engineering Contradiction:
Improvepersonal data exposureVSAvoididentification capability
Core Design Contradiction:
Object-affected harmful factorsVSLoss of information

Solution Approach 1:

The patent introduces an automated intermediary system that acts between the original image data and the training dataset. This intermediary automatically detects, evaluates, and modifies identifying features based on configurable criteria, providing a systematic approach to anonymization that goes beyond simple metadata removal.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent modifies parameters of anatomical features in the image data, such as altering the shape, size, or position of identifiable structures like teeth or eyes. These parameter changes preserve the overall anatomical context for training while making specific identifying characteristics unrecognizable.

Inventive Principle:
Principle #35Parameter changes

3Object-affected harmful factors

If image data is modified to prevent identification, then anonymization is improved, but the data may become less suitable for training purposes

Engineering Contradiction:
Improveidentification riskVSAvoidtraining data quality
Core Design Contradiction:
Object-affected harmful factorsVSReliability

Solution Approach 1:

The patent applies partial modification only to specific identifying features rather than altering the entire dataset. By selectively targeting only the anatomical structures that pose identification risks while leaving other training-relevant features unchanged, the system maintains training quality while achieving anonymization.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent applies different levels and types of modification to different regions of the image data. Identifying features receive targeted modification while other anatomical structures remain intact, creating local variations in data quality that preserve overall training value while eliminating identification risks.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3798989B1Computer-implemented method for providing output data for training and / or validating an algorithm
Publication Date: 2022.07.20 SIEMENS HEALTHCARE GMBH
  • EP3798989B1 patent drawingFigure 1~2
  • EP3798989B1 patent drawingFigure 3
  • EP3798989B1 patent drawingFigure 4

AI summary

A computer-implemented method for providing output data (37) for training and/or validating an algorithm (41) whose processing result (47) relates to an anatomical region (52) of an object of investigation mapped by the output data (37), comprising the steps of: - receiving input data (33) that are image data (54) or raw data (63) of a medical imaging procedure, wherein the image data (54) or the raw data (63) at least partially map the anatomical region (52, 57, 81) and at least one further anatomical region (53, 58, 59, 82, 89) of the object of investigation, - generating output data (37) by modifying the image data (54) or the raw data (63) such that the mapping of the further anatomical region (52, 57, 81) by the image data (54) or the raw data (63) is altered, in particular, it is altered, and - providing the source data (37).