Medical imaging datasets are essential for training artificial intelligence systems, but sharing brain magnetic resonance imaging (MRI) scans can expose patient-specific anatomical patterns. Anonymization must therefore do more than remove identifiers: it must make the overall image harder to link back to a patient while retaining the lesion information needed for segmentation and classification. Existing generative approaches often emphasize global image quality or facial de-identification, leaving lesion-centric utility and re-identification risk insufficiently addressed.

In a new research article, led by Hairong Lv from Tsinghua University, presents a lesion-aware anonymization framework in Quantitative Biology. The method combines variational autoencoders (VAEs) with a latent diffusion model. Two VAEs separately encode lesion and non-lesion regions; a diffusion model conditioned on the class label, lesion mask, and slice depth generates varied latent representations for non-lesion tissue; and a decoder recombines the generated background with the original lesion representation. This division of labor is designed to vary patient-specific context while preserving clinically relevant lesion structure.

Experiments on the public BraTS2020 brain-tumor dataset show that the pre-trained VAEs reconstructed 20,000 samples with a structural similarity index (SSIM) of at least 98.2%. The generated images achieved an 86% anonymity rate under the study’s PSNR-based criterion. In low-data settings, adding anonymized samples increased tumor-segmentation mIoU from 61.40% to 66.00% - a gain of 4.60 percentage points - and classification accuracy from 61.00% to 69.75%, a gain of 8.75 points. With the same number of training samples, synthetic-only classification reached 95.75%, close to 99.25% with real images. As illustrated in Figure 1, the framework offers a route to privacy-aware MRI data synthesis for model development and evaluation, while future work will examine 3D sequences, stronger anonymization of cranial contours, and the privacy risks of rare or distinctive lesion morphologies.
DOI :10.1002/qub2.70033