Anatomical Image Encryption With Patch-Based AI Anonymization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI training methods using clinical patient images face challenges with patient privacy regulations due to the potential identification of individuals from anonymized images, and synthesized images may introduce systematic errors.
Innovation Solution
A method involving spatial mapping of patient images to a common reference frame, partitioning into spatial regions, and applying statistical inverse spatial mappings to generate anonymized training images, ensuring patient-specific information is not extractable.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real patient images are used for AI training, then training accuracy is improved, but patient privacy is compromised
Solution Approach 1:
The patent segments patient images into multiple patches and recombines them to create synthetic training images. This segmentation approach allows the training data to capture anatomical variations and pathological features from multiple patients while preventing identification of any single patient, thus resolving the contradiction between training accuracy and privacy protection.
Solution Approach 2:
The patent creates synthetic copies of patient images by combining patches from different sources. These synthetic copies preserve the statistical properties and diagnostic features of real images for training purposes while being inherently anonymized, as they do not directly correspond to any single patient's actual image.
2Reliability
If patient consent is obtained for image use, then privacy compliance is improved, but training set size and diversity are limited
Solution Approach 1:
By creating synthetic copies through patch recombination, the method generates large volumes of training data without requiring additional patient consents. The synthetic images can be generated indefinitely from the pooled patch library, expanding training set size while maintaining privacy compliance through the anonymization inherent in the synthesis process.
3Object-affected harmful factors
If synthesized images are used for training, then privacy protection is improved, but systematic errors are introduced
Solution Approach 1:
The segmentation of images into patches followed by random recombination preserves the statistical distribution of anatomical structures and pathological features present in the original patient images. This approach maintains training accuracy by ensuring that the synthetic images reflect real-world variations without introducing systematic biases, while still achieving privacy protection through anonymization.
Data Source
Figure 1
Figure 2
Figure 3A~4
AI summary
An apparatus (10) for generating a training set of anonymized images (40) for training an artificial intelligence (AI) component (42) from images (11) of a plurality of persons. The apparatus includes at least one electronic processor (20) programmed to: spatially map the images of the plurality of persons to a reference image (30) to generate images (32) in a common reference frame; partition the images in the common reference frame into P spatial regions (34) to generate P sets of image patches (36) corresponding to the P spatial regions; assemble a set of training images (38) in the common reference frame by, for each training image in the common reference frame, selecting an image patch from each of the P sets of image patches and assembling the selected image patches into the training image in the common reference frame; and process the training images in the common reference frame to generate the training set of anonymized images including applying statistical inverse spatial mappings to the training images in the common reference frame, wherein the statistical inverse spatial mappings are derived from spatial mappings (33) of the images of the plurality of persons to the reference image.