Medical Image De-Identification Using Relevance-Guided Pixel Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical images are needed for deep learning (DL) model training but transferring them from hospitals to third parties is often denied due to concerns that patient identity could be revealed through pixel data, preventing effective data utilization.
Innovation Solution
A computer-implemented method for de-identifying medical images by determining pixel relevance based on in-image annotations, applying localized modifications such as blurring or noise addition to reduce patient re-identification risk while maintaining image usefulness for DL training, with an internal check using an existing DL model to ensure equivalence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If pixel data from medical images is transferred to third parties for DL training, then the quality and effectiveness of model training is improved, but the risk of patient re-identification increases
Solution Approach 1:
The patent applies different processing intensities to different regions of the medical image based on a relevance map. Regions with high diagnostic relevance (such as anatomical structures and pathologies) are preserved with minimal modification, while regions with low relevance (such as background areas and less critical anatomical regions) undergo stronger de-identification transformations like noise addition, blurring, or pixel value modification. This localized differential processing resolves the contradiction by maintaining training quality in relevant regions while reducing re-identification risk in irrelevant regions.
2Object-affected harmful factors
If de-identification manipulation is applied to pixel data, then patient re-identification risk is reduced, but the usefulness of the image for DL model training deteriorates
Solution Approach 1:
The patent generates a relevance map that identifies and weights different regions of the medical image based on their diagnostic importance. This relevance map guides the de-identification process by applying transformation intensities that are locally adapted to each region's importance. Consequently, the image maintains its usefulness for DL training in critical regions while achieving adequate de-identification in non-critical regions, thus resolving the contradiction between privacy protection and training utility.
Solution Approach 2:
The patent modifies pixel parameters (such as intensity values, noise levels, and blur radii) based on the relevance map without fundamentally changing the image structure or diagnostic features. By adjusting these parameters locally rather than applying uniform transformations, the patent preserves the statistical properties and visual characteristics needed for DL training while reducing re-identification risk through controlled parameter modifications in less relevant regions.
3Object-affected harmful factors
If uniform de-identification is applied to the entire image, then patient re-identification risk is reduced, but diagnostic relevance and training value are lost
Solution Approach 1:
The patent explicitly avoids uniform de-identification by implementing region-specific processing based on the relevance map. Each region receives a customized de-identification intensity proportional to its diagnostic importance, ensuring that critical diagnostic information is preserved while non-critical regions are adequately anonymized. This localized approach directly resolves the contradiction by preventing information loss in diagnostically relevant areas.
Data Source
AI summary
A computer-implemented method, comprising: receiving input data including a medical image and an in-image annotation; applying a first function to the input data to determine a relevance value of pixels in the image and a relevance map; applying a second function to the medical image to generate a de-identified medical image; applying a trained function to the medical image and the de-identified medical image to determine a first property in the medical image and a second property in the de-identified medical image; applying a comparison function to the first property and the second property to determine a similarity value, wherein in response to the similarity value being below a similarity threshold, the relevance map is adjusted and the applying of the second function, the applying of the trained function and the applying of the comparison function are repeated; and providing the de-identified medical image and the in-image annotation.