Image Anonymization Using Feature Separation to Preserve Utility
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for anonymizing patient medical images fail to preserve the utility of the images while effectively concealing patient identity, leading to potential privacy breaches.
Innovation Solution
A system utilizing machine learning models, including an encoder and decoder neural network, transforms identifying features irreversibly into identity-concealing features while maintaining residual features for utility attributes, such as gaze and facial expressions, using techniques like ensemble learning and variational encoder-decoder networks to ensure privacy and confidentiality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional anonymization methods (blurring, pixelation) are applied to medical images, then patient identity is concealed, but the utility of the image for diagnostic and AI tasks is degraded
Solution Approach 1:
The patent segments the image into two distinct feature sets: identifying features (extracted by the first ML model) and residual features (extracted by the second ML model). This segmentation allows selective transformation of only the identifying features while preserving the residual features that contain diagnostic utility, thereby resolving the contradiction between privacy protection and image utility preservation.
Solution Approach 2:
The patent extracts identifying features from the original medical image using a trained ML model, separates them from the residual features, and applies transformation only to the extracted identifying features. This extraction approach enables targeted anonymization without degrading the overall image quality or diagnostic utility retained in the residual features.
2Reliability
If identifying features are transformed irreversibly to conceal identity, then privacy is enhanced, but the ability to maintain image utility attributes may be compromised
Solution Approach 1:
The patent applies different quality requirements to different parts of the image processing pipeline: the identifying features are transformed with high security requirements (irreversible transformation, maximum distance in feature space), while the residual features are preserved with high fidelity to maintain diagnostic utility. This local differentiation of quality requirements resolves the contradiction between security and adaptability.
Solution Approach 2:
The patent changes the parameters of identifying features through irreversible transformation (applying stochastic operations, maximizing distance in feature space) while maintaining the parameters of residual features unchanged. This selective parameter transformation ensures both identity concealment security and preservation of image utility attributes.
3Reliability
If multiple ML models are used to extract and transform features, then anonymization effectiveness is improved, but system complexity increases
Solution Approach 1:
The patent performs preliminary training of the ML models offline before deployment. The first ML model is trained to extract identifying features, the second ML model is trained to extract residual features, and the transformation pipeline is pre-configured. This preliminary action reduces the complexity of the deployed system, as the models are already trained and ready for inference without requiring complex real-time training operations.
Data Source
AI summary
A person's privacy is protected by the law in many settings and disclosed herein are systems, methods, and instrumentalities associated with anonymizing an image of a person while still preserving the visual saliency and/or utility of the image for one or more downstream tasks. These objectives may be accomplished using various machine-learning (ML) techniques such as ML models trained for extracting identifying and residual features from the input image as well as ML models trained for transforming the identifying features into identity-concealing features and for preserving the utility features of the image. An output image may be generated based on the various ML models, wherein the identity of the person may be substantially disguised in the output image while the background and utility attributes of the original image may be substantially maintained in the output image.