A Preset-Based Privacy–Utility Framework for Re-Identification Risk and Disease Classification in Structural MRI
Year: 2026
Type: article
Abstract: Structural magnetic resonance imaging (sMRI) contains both disease-related and subject-specific anatomical information, making it necessary to consider disease classification utility alongside re-identification risk in neuroimaging data sharing. We propose a preset-based privacy–utility framework for quantitatively comparing these factors in sMRI. The framework combines key-based block scrambling, downsampling, intensity modulation, and Gaussian noise injection to define three protection presets: Utility-First, Balanced, and Privacy-First. Using NeuroID-Swin3D, we evaluated cognitively normal (CN) versus Alzheimer’s disease (AD) classification using Macro F1, Macro Recall, and area under the receiver operating characteristic curve (AUROC), and repeated-scan linkability using mean average precision at rank 3 (mAP@3) and cumulative matching characteristic at rank 3 (CMC@3). On the final test set, Utility-First retained utility comparable to the RAW baseline but produced only limited changes in the privacy metrics. Balanced retained 99.51%, 99.05%, and 97.87% of the RAW values for Macro F1, Macro Recall, and AUROC, respectively, while reducing mAP@3 and CMC@3 by 75.00% and 69.63%. Privacy-First reduced mAP@3 and CMC@3 by 93.93% and 91.31%, respectively, while exhibiting lower classification utility. These results demonstrate a staged privacy–utility trade-off across the tested protection levels and provide a framework for comparing protection levels according to data-use objectives.
Source: Applied Sciences
DOI: 10.3390/app162211474
Institution: Jeonbuk National University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7236810915
