Frozen Brain-MRI Foundation Models Are Site Fingerprints

• URL: http://arxiv.org/abs/2608.10295v1
• Date: Mon, 10 Aug 2026 23:03:21 GMT
• Title: Frozen Brain-MRI Foundation Models Are Site Fingerprints
• Abstract summary: Frozen foundation-model embeddings are increasingly used as off-the-shelf brain-MRI representations.<n>We audit what they actually encode and find that site is a large, intrinsic component of the representation.<n>We recommend site-audited use of frozen brain-MRI FMs and release an open audit toolkit.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Frozen foundation-model (FM) embeddings are increasingly used as off-the-shelf brain-MRI representations, on the assumption that they capture anatomy. We audit what they actually encode and find that acquisition site is a large, intrinsic component of the representation. Across two independent cohorts (ABIDE-I, ABIDE-II), three frozen 3-D encoders (brain-pretrained, CT-pretrained, and randomly initialized), and every network depth, site is linearly decodable at roughly 0.9 balanced accuracy at deep layers, exceeding the decodability of every clinical or demographic variable (sex, age, autism diagnosis) at every layer. The effect is intrinsic rather than learned: a randomly initialized encoder is already a ~0.9 site classifier on both cohorts and across three architecture families (Swin, ViT, ResNet), and site is decodable at ~0.95 directly from the raw downsampled image with no encoder, so the fingerprint reflects low-level image statistics that any encoder preserves rather than a product of pretraining. Residualizing measured population covariates leaves site decodability essentially unchanged, indicating an acquisition- rather than population-driven effect. A nonlinear probe matches the linear one, so the fingerprint is fully linearly accessible. The site subspace is removable post hoc by iterative null-space projection or ComBat (site decodability 0.94 -> 0.07/0.00), and is a site-attribution concern for shared or federated embeddings; but for dense segmentation this removal is not free, because site and anatomy occupy an entangled linear subspace (a matched-rank random-direction projection is Dice-neutral, whereas removing the site subspace is destructive). We recommend site-audited use of frozen brain-MRI FMs and release an open audit toolkit.

Related papers

• ReMAP: Self-supervised learning to unveil brain representations and vulnerability [3.3233194140188687]
  General anesthesia offers a rare opportunity to observe the human brain under a standardized, controlled perturbation.<n>EEG is almost always reduced to a single proprietary depth index, collapsing a rich trajectory into one number.<n>Here we ask whether the geometry of that trajectory, not merely the depth it reaches, carries clinically meaningful information.
  arXiv  Detail & Related papers  (2026-08-22T17:09:49Z)
• Sparse Concept Channels in Frozen 3D CT Vision Encoders [3.9173955300197973]
  Large vision-language models are becoming increasingly dominant in 3D medical image interpretation.<n>We first study this on a 3D chest vision-language model (Pillar-0) by probing its frozen vision embeddings.<n>We show that (i) each radiological finding is encoded by a i>sparse/i> set of 10 vision-encoder channels that match full-feature classification performance.
  arXiv  Detail & Related papers  (2026-07-23T07:17:11Z)
• Amplifying, Not Learning: Fine-Tuned AI Text Detectors Amplify a Pretrained Direction [51.56484100374058]
  Text detectors amplify a pretrained typicality axis; they do not construct an AI-vs-human boundary.<n>On raw encoders before any task supervision, projecting onto centroid(AI)-centroid(HC3) achieves NYT-vs-HC3 AUROC 0.806/0.944/0.834 across three architectures.<n>On RoBERTa-base, raw projection exceeds fine-tuning; on RoBERTa-base, full fine-tuning reduces discrimination below raw on both fluent-formal populations tested.
  arXiv  Detail & Related papers  (2026-05-20T19:08:38Z)
• VoxShield: Protecting 3D Medical Datasets from Unauthorized Training via Frequency-Aware Inter-Slice Disruption [25.533159285300176]
  Unlearnable Examples (UE) offer protection by injecting imperceptible perturbations to prevent effective model learning.<n>We propose VoxShield, a UE framework that explicitly targets the inductive biases of 3D networks.<n> Experiments on BraTS19 and FLARE21 demonstrate that VoxShield successfully degrades 3D segmentation performance.
  arXiv  Detail & Related papers  (2026-05-17T09:35:19Z)
• Hide-and-Seek Attribution: Weakly Supervised Segmentation of Vertebral Metastases in CT [68.09387763135236]
  We introduce a weakly supervised method trained solely on vertebra-level healthy/malignant labels, without any lesion masks.<n>We achieve strong blastic/lytic performance despite no mask supervision.
  arXiv  Detail & Related papers  (2025-12-07T14:03:28Z)
• Self-supervised learning for classifying paranasal anomalies in the maxillary sinus [31.45131665942058]
  Self-supervised learning can be used to learn representations from unlabelled data. There are no SSL methods designed for the downstream task of classifying paranasal anomalies in the maxillary sinus. Our approach uses a 3D Convolutional Autoencoder trained in an unsupervised anomaly detection framework.
  arXiv  Detail & Related papers  (2024-04-29T11:14:11Z)
• Fighting the scanner effect in brain MRI segmentation with a progressive level-of-detail network trained on multi-site data [1.6379393441314491]
  LOD-Brain is a 3D convolutional neural network with progressive levels-of-detail able to segment brain data from any site. It produces state-of-the-art results, with no significant difference in performance between internal and external sites. Its portability opens the way for large scale application across different healthcare institutions, patient populations, and imaging technology manufacturers.
  arXiv  Detail & Related papers  (2022-11-04T12:15:18Z)
• Building Brains: Subvolume Recombination for Data Augmentation in Large Vessel Occlusion Detection [56.67577446132946]
  A large training data set is required for a standard deep learning-based model to learn this strategy from data. We propose an augmentation method that generates artificial training samples by recombining vessel tree segmentations of the hemispheres from different patients. In line with the augmentation scheme, we use a 3D-DenseNet fed with task-specific input, fostering a side-by-side comparison between the hemispheres.
  arXiv  Detail & Related papers  (2022-05-05T10:31:57Z)
• Adversarial Distortion Learning for Medical Image Denoising [43.53912137735094]
  We present a novel adversarial distortion learning (ADL) for denoising two- and three-dimensional (2D/3D) biomedical image data. The proposed ADL consists of two auto-encoders: a denoiser and a discriminator. Both the denoiser and the discriminator are built upon a proposed auto-encoder called Efficient-Unet.
  arXiv  Detail & Related papers  (2022-04-29T13:47:39Z)
• FedMed-ATL: Misaligned Unpaired Brain Image Synthesis via Affine Transform Loss [58.58979566599889]
  We propose a novel self-supervised learning (FedMed) for brain image synthesis. An affine transform loss (ATL) was formulated to make use of severely distorted images without violating privacy legislation. The proposed method demonstrates advanced performance in both the quality of synthesized results under a severely misaligned and unpaired data setting.
  arXiv  Detail & Related papers  (2022-01-29T13:45:39Z)
• Localized Perturbations For Weakly-Supervised Segmentation of Glioma Brain Tumours [0.5801621787540266]
  This work proposes the use of localized perturbations as a weakly-supervised solution to extract segmentation masks of brain tumours from a pretrained 3D classification model. We also propose a novel optimal perturbation method that exploits 3D superpixels to find the most relevant area for a given classification using a U-net architecture.
  arXiv  Detail & Related papers  (2021-11-29T21:01:20Z)
• 3-Dimensional Deep Learning with Spatial Erasing for Unsupervised Anomaly Segmentation in Brain MRI [55.97060983868787]
  We investigate whether using increased spatial context by using MRI volumes combined with spatial erasing leads to improved unsupervised anomaly segmentation performance. We compare 2D variational autoencoder (VAE) to their 3D counterpart, propose 3D input erasing, and systemically study the impact of the data set size on the performance. Our best performing 3D VAE with input erasing leads to an average DICE score of 31.40% compared to 25.76% for the 2D VAE.
  arXiv  Detail & Related papers  (2021-09-14T09:17:27Z)
• Fader Networks for domain adaptation on fMRI: ABIDE-II study [68.5481471934606]
  We use 3D convolutional autoencoders to build the domain irrelevant latent space image representation and demonstrate this method to outperform existing approaches on ABIDE data.
  arXiv  Detail & Related papers  (2020-10-14T16:50:50Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.