Paper deep dive
Do 3D Medical Foundation Models See Through MRI Artifacts? A Controlled Study of Representation Robustness
Julia Anna Mielcarz, Daniel Klaaby, Mostafa Mehdipour Ghazi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/10/2026, 2:46:05 AM
Summary
This study evaluates the robustness of five self-supervised 3D medical foundation models (3DINO, BrainIAC, BrainFM, NeuroVFM, Neuro-SimCLR) against seven MRI artifacts (k-space spikes, ghosting, Gibbs ringing, Rician noise, Gaussian blur, bias field, gamma contrast). Using BraTS-Africa data, the authors assess representation stability via Centered Kernel Alignment (CKA), RankMe, and UMAP, alongside segmentation consistency. Results indicate that robustness is model- and artifact-dependent, with 3DINO showing the most stable representations. Artifacts often distort representation geometry without causing dimensional collapse, and larger pretraining scale does not guarantee artifact invariance.
Entities (20)
Relation Signals (10)
Julia Anna Mielcarz → authored → Do 3D Medical Foundation Models See Through MRI Artifacts? A Controlled Study of Representation Robustness
confidence 99% · Do 3D Medical Foundation Models See Through MRI Artifacts? A Controlled Study of Representation Robustness Julia Anna Mielcarz
3DINO → exhibits → most consistently stable representations
confidence 95% · 3DINO exhibits the most consistently stable representations
BrainIAC → is → highly sensitive
confidence 95% · BrainIAC is highly sensitive to several corruptions
Artifacts → cause → distortion of representation geometry
confidence 90% · artifacts often distort representation geometry without causing dimensional collapse
BraTS-Africa → isusedby → Controlled Study
confidence 90% · Using BraTS-Africa cases with four MRI sequences, we generate seven frequency- and image-domain artifacts
CKA → measures → Representation Geometry
confidence 90% · CKA is particularly suitable as it measures alignment between centered Gram structures
RankMe → measures → spectral effective rank
confidence 90% · RankMe estimates spectral effective rank from the entropy of normalized singular values
BrainIAC → uses →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Self-supervised 3D medical foundation models are increasingly used as general-purpose feature extractors, yet their sensitivity to MRI artifacts remains poorly understood. We present a controlled evaluation of representation robustness across five pretrained 3D encoders spanning different architectures, objectives, pretraining domains, and dataset scales. Using BraTS-Africa cases with four MRI sequences, we generate seven frequency- and image-domain artifacts at five predefined corruption settings. Robustness is assessed using linear centered kernel alignment (CKA), RankMe, and UMAP, complemented by an independent segmentation-consistency analysis. We find that robustness is strongly model- and artifact-dependent. 3DINO exhibits the most consistently stable representations, while BrainIAC is highly sensitive to several corruptions; NeuroVFM, BrainFM, and Neuro-SimCLR show intermediate but distinct artifact-specific profiles. Across many conditions, CKA decreases substantially while RankMe remains comparatively stable, indicating that artifacts often distort representation geometry without causing dimensional collapse. Segmentation consistency also degrades under corruption, particularly for ghosting and Rician noise, but aligns only partially with representation-level robustness. These findings show that larger-scale or domain-specific pretraining alone does not guarantee artifact invariance and motivate explicit robustness evaluation before deploying 3D foundation models in heterogeneous MRI settings.
Tags
Links
- Source: https://arxiv.org/abs/2608.06613v1
- Canonical: https://arxiv.org/abs/2608.06613v1
Trouble viewing inline? Open PDF directly →
Full Text
57,555 characters extracted from source content.
Expand or collapse full text
Do 3D Medical Foundation Models See Through MRI Artifacts? A Controlled Study of Representation Robustness Julia Anna Mielcarz , Daniel Klaaby, and Mostafa Mehdipour Ghazi Pioneer Centre for AI, University of Copenhagen, Copenhagen, Denmark ghazi@di.ku.dk Abstract. Self-supervised 3D medical foundation models are increas- ingly used as general-purpose feature extractors, yet their sensitivity to MRI artifacts remains poorly understood. We present a controlled evaluation of representation robustness across five pretrained 3D en- coders spanning different architectures, objectives, pretraining domains, and dataset scales. Using BraTS-Africa cases with four MRI sequences, we generate seven frequency- and image-domain artifacts at five prede- fined corruption settings. Robustness is assessed using linear centered kernel alignment (CKA), RankMe, and UMAP, complemented by an in- dependent segmentation-consistency analysis. We find that robustness is strongly model- and artifact-dependent. 3DINO exhibits the most con- sistently stable representations, while BrainIAC is highly sensitive to several corruptions; NeuroVFM, BrainFM, and Neuro-SimCLR show in- termediate but distinct artifact-specific profiles. Across many conditions, CKA decreases substantially while RankMe remains comparatively sta- ble, indicating that artifacts often distort representation geometry with- out causing dimensional collapse. Segmentation consistency also degrades under corruption, particularly for ghosting and Rician noise, but aligns only partially with representation-level robustness. These findings show that larger-scale or domain-specific pretraining alone does not guaran- tee artifact invariance and motivate explicit robustness evaluation before deploying 3D foundation models in heterogeneous MRI settings. Keywords: Foundation Models· Self-Supervised Learning· Represen- tation Analysis· MRI Artifacts· Neuroimaging 1 Introduction Self-supervised pretraining has enabled 3D medical-image encoders to learn reusable representations from large collections of unlabeled scans. Recent models span broad multi-organ pretraining [1], brain-specific contrastive learning [2,3], multi-task learning [4], and large-scale joint-embedding prediction [5]. These rep- resentations are increasingly used as general-purpose features for segmentation, classification, and disease analysis. Their practical value, however, depends not only on downstream accuracy but also on whether they remain stable under variations that are unrelated to the anatomy or pathology of interest. arXiv:2608.06613v1 [cs.CV] 6 Aug 2026 2J. A. Mielcarz et al. This requirement is particularly important in MRI, where scanner hardware, acquisition protocols, reconstruction procedures, and patient motion can sub- stantially alter image appearance. Common degradations include noise, intensity inhomogeneity, ghosting, Gibbs ringing, blurring, and k-space interference [6]. Motion-related degradation alone affects a substantial fraction of clinical MR acquisitions and can require repeated scanning [7]. Such artifacts may provide high-amplitude visual cues that pretrained encoders inadvertently represent, de- spite carrying no intended clinical information. This creates a risk that feature spaces change across acquisition conditions even when the underlying anatomy is unchanged, limiting transfer across scanners, sites, and patient populations [8,9]. Robustness to these effects cannot be inferred reliably from a single down- stream task. Task performance depends on the prediction head, available an- notations, and task-specific information, and may obscure substantial changes within the underlying representation. Direct representation analysis instead asks whether matched clean and corrupted images retain similar relational geometry and spectral structure. Although representation similarity has been extensively studied in general deep learning, e.g., using Representational Similarity Analy- sis (RSA) [10] or Singular Vector Canonical Correlation Analysis (SVCCA) [11], the sensitivity of recent 3D medical and neuroimaging encoders to controlled MRI artifacts remains poorly understood. In particular, it is unclear whether robustness is associated with broader pretraining data, greater training scale, brain-specific specialization, architectural choice, or pretraining objective. We address this gap through a controlled benchmark of five public 3D en- coders with different architectures, objectives, domains, and pretraining scales. Clean brain MRIs are perturbed using seven frequency- and image-domain arti- facts at five predefined corruption settings and evaluated across four modalities. We quantify changes in representation geometry using centered kernel alignment (CKA) [12], changes in spectral effective rank using RankMe [13], and qualitative organization using UMAP [14]. An independent segmentation-consistency exper- iment provides a task-level reference for assessing whether the relative effects of the same artifacts extend beyond the evaluated representation spaces. Our main contributions are three-fold: (1) We introduce a controlled eval- uation of artifact robustness across five recent pretrained 3D medical-image en- coders, seven MRI-relevant corruptions, five corruption settings, and four MRI modalities (T1, T1ce, T2, FLAIR). (2) We show that robustness is strongly model- and artifact-dependent: broad medical pretraining provides the most con- sistent stability overall, while brain-specific specialization or larger pretraining scale does not guarantee robustness to all corruption mechanisms. (3) We find that artifacts often induce substantial geometric drift without a correspond- ing collapse in effective rank [15], demonstrating that representation similarity and spectral dimensionality capture distinct failure modes. Comparison with an independent segmentation model further shows that representation-level and task-level sensitivities align only partially. 3D Foundation Models Under MRI Artifacts3 2 Related Work Pretrained 3D medical-image encoders. Early medical-image transfer learn- ing largely relied on two-dimensional networks pretrained on natural images, de- spite the different visual statistics and volumetric structure of medical data. Re- cent work instead pretrains 3D encoders directly on medical images. 3DINO uses self-distillation across approximately 100,000 multi-organ MRI, CT, and PET volumes; BrainIAC and Neuro-SimCLR apply contrastive learning to brain MRI; BrainFM learns a shared encoder through multiple brain-imaging tasks; and NeuroVFM uses volumetric joint-embedding prediction at substantially larger clinical scale. These models differ simultaneously in architecture, objective, pre- training domain, and dataset scale, but their relative sensitivity to MRI artifacts has not been systematically compared. MRI artifacts and robustness. MRI quality can be degraded by acquisition and reconstruction effects such as noise, intensity inhomogeneity, radiofrequency interference, motion ghosting, and finite k-space sampling. Prior studies have simulated such effects for data augmentation, quality assessment, artifact detec- tion, or robustness evaluation. Most evaluations, however, quantify image quality or downstream predictive performance. These measurements do not directly re- veal whether matched clean and corrupted images retain a similar organization within the representation space of a pretrained encoder. Our study instead uses controlled artifacts to compare representation robustness across multiple recent 3D medical-image models. Representation-level evaluation. Neural representations can be evaluated through their similarity, geometry, or spectral structure. RSA, SVCCA, and CKA compare feature spaces derived from matched inputs. CKA is particu- larly suitable as it measures alignment between centered Gram structures while remaining invariant to orthogonal transformations and isotropic scaling. Com- plementarily, RankMe estimates spectral effective rank from the entropy of nor- malized singular values, providing a label-free measure of whether embeddings become concentrated in fewer directions. We use them jointly, as an artifact may distort the relative geometry of the subjects without any dimensional collapse. UMAP is used only for qualitative inspection of artifact-dependent organization. 3 Methods We investigate how MRI artifacts alter the internal representations of pretrained 3D medical-image encoders and whether this sensitivity varies across pretraining strategies, artifact types, and corruption settings. To this end, we construct a controlled-perturbation benchmark and evaluate five frozen encoders using complementary measures of representational geometry and spectral structure. An independent tumor-segmentation consistency analysis provides a task-level reference for comparing the relative effects of the same artifacts. 4J. A. Mielcarz et al. 3.1 Pretrained 3D Encoders Model selection. We evaluate five publicly available pretrained models that process three-dimensional medical images. The selected models differ along four design dimensions that may influence robustness to imaging artifacts: (i) net- work architecture, including vision transformers (ViTs), convolutional encoders, and U-Net-like dense-prediction networks; (i) pretraining objective, including self-distillation, instance-contrastive learning, joint-embedding predictive learn- ing, and multi-task learning; (i) pretraining domain, ranging from multi-organ, multimodal medical imaging to brain-specific MRI; and (iv) pretraining scale, ranging from tens of thousands to several million images. The purpose of this comparison is not to rank the models by downstream accuracy, but to determine whether these design choices are associated with different degrees of representa- tional stability under controlled image corruptions. All pretrained models are frozen throughout the study. For each model m, an image volumeI is transformed by the model-specific preprocessing operatorP m and then passed through the frozen encoder f m . A fixed aggregation operator g m converts the token sequence or spatial feature map into a single vector: z (m) = φ m (I) = g m (f m (P m (I)))∈ R d m ,(1) where d m denotes the embedding dimensionality. For each model, identical pre- processing and feature extraction are applied to the clean and corrupted volumes. We evaluate five complementary pretrained 3D encoders. 3DINO [1] extends DINO-style self-distillation [16] to volumetric, multi-organ medical imaging using a ViT backbone. BrainIAC [2] is a brain-MRI-specific ViT pretrained with the SimCLR objective [17]. BrainFM [4] adopts a U-Net-like architecture trained jointly across multiple brain-imaging tasks. NeuroVFM [5] is a large-scale ViT- based neuroimaging encoder pretrained on heterogeneous clinical MRI and CT volumes using Volumetric Joint-Embedding Predictive Architecture (Vol-JEPA). Finally, Neuro-SimCLR [3] uses a convolutional 3D ResNet-18 pretrained with SimCLR on structural brain MRI. Together, these models span different archi- tectures, pretraining objectives, domains, and dataset scales. Table 1 summarizes the properties of the original studies and the representa- tions extracted in this work. Because the source studies use different definitions of a training sample, we retain their original terminology, such as scans, volumes, or acquired images, rather than interpreting these counts as numbers of unique subjects. Full details on the pretraining data, model-specific preprocessing, and feature-extraction layers are provided in Appendix. 3.2 Artifact Simulation Framework We construct controlled corrupted versions of each clean MRI volume to evalu- ate how pretrained representations respond to image-quality degradation. The selected transforms are divided according to the domain in which they are imple- mented: (i) frequency-domain corruptions, which modify the Fourier representa- tion of the image, and (i) image-domain corruptions, which operate directly on 3D Foundation Models Under MRI Artifacts5 Table 1: Pretrained 3D encoders evaluated in this study. “Representation” denotes the frozen feature extracted in our experiments, with dimensionality d, and is not necessarily the representation used in the original downstream evaluations. ModelBackboneObjectiveDomainPretraining dataRepresentation (d) 3DINO [1]3D ViT-L/16Self-distillation Multi-organ MRI/CT/PET 100,000 volumesFinal class token (1024) BrainIAC [2]3D ViT-B/16ContrastiveBrain MRI32,015 scansFinal class token (768) BrainFM [4]3D U-Net-likeMulti-task learningBrain MRI/CT7,800 real images Globally pooled encoder features (2048) NeuroVFM [5]3D ViT-B/(4×16×16) Volumetric JEPA Neuroimaging MRI/CT 5.24M volumesMean-pooled patch tokens (768) Neuro-SimCLR [3]3D ResNet-18ContrastiveBrain MRI44,958 scans Globally pooled encoder features (512) voxel intensities. The former comprises k-space spikes, periodic k-space ghost- ing, and k-space truncation leading to Gibbs ringing. The latter comprises Rician noise, Gaussian blur, bias field, and gamma contrast transformation. Let I ∈ [0, 1] H×W×D denote a real-valued magnitude MRI volume after common intensity normalization, but before model-specific preprocessing. Each artifact transformation is applied to this common representation, after which the resulting volume is processed independently by the preprocessing operator P m of each encoder. Thus, all models are evaluated using corruptions generated from the same underlying image. For artifact type a and severity level ℓ, the corrupted image is written as I (a,ℓ) =T a (I;θ a,ℓ ,ξ a ),(2) where θ a,ℓ is the parameter controlling artifact severity andξ a contains any random variables, such as spike locations or noise realizations. Each artifact is applied independently so that changes in the resulting representations can be attributed to one corruption type and severity level. Frequency-Domain Corruptions LetF andF −1 denote the three-dimensional fast Fourier transform (FFT) and its inverse, respectively. We use the centered Fourier representation K = fftshift (F (I))∈ C H×W×D ,(3) where the zero-frequency component is located at the center of the array. Given a modified spectrum e K, the corresponding magnitude image is reconstructed as R( e K) = F −1 ifftshift( e K) . (4) K-space spikes. Transient radiofrequency interference or hardware instabil- ity can introduce isolated, high-amplitude samples into k-space. Because a sin- gle Fourier coefficient contributes a sinusoidal component over the full spatial domain, such disturbances appear as global stripe or herringbone patterns in the reconstructed image [6]. We sample a set of M distinct locations Ω sp = k 1 ,...,k M from the k-space grid. The corrupted image is defined as K sp [k] =K[k] + α sp ∥K∥ ∞ 1 [k∈ Ω sp ], T sp (I) =R(K sp )(5) 6J. A. Mielcarz et al. where ∥K∥ ∞ = max k |K[k]|, α sp > 0 controls the spike amplitude, and 1[·] is the indicator function. We fix the number of injected coefficients to M = 2 and vary only α sp across severity levels, thereby separating artifact amplitude from the number of corrupted k-space locations. Periodic k-space ghosting. Ghost replicas can arise when periodically incon- sistent measurements occur along the phase-encoding direction [6,18]. A common synthetic approximation is to periodically attenuate k-space lines, producing dis- placed copies of the image after inverse Fourier reconstruction. Let q ∈x,y,z denote the selected encoding axis, k q be the corresponding frequency index, and R denote the modulation period. The periodic modulation is obtained as K gh [k] = ( αK[k], k q ≡ 0 (mod R), K[k], otherwise, T gh (I) =R(K gh ).(6) We fix R = 3, set q to x-axis, and vary α gh ∈ (0, 1]. Smaller values cause stronger attenuation of the periodically selected lines and consequently more prominent replicas. This transformation is referred to as periodic k-space ghosting rather than a complete physical simulation of subject motion, because it does not model a time-dependent rigid transformation during acquisition. Gibbs ringing. Gibbs ringing results from the finite sampling or truncation of k-space. Multiplication of k-space by a rectangular window corresponds to convolution of the reconstructed image with a sinc-like point-spread function, producing oscillations near sharp intensity transitions [19,20]. Following previous MRI augmentation frameworks [21,22], we simulate this effect by retaining only a centered region of k-space. Let c q denote the number of retained coefficients along the axis q. The truncation mask is K gi [k] = Y q∈x,y,z 1 k q − N q − 1 2 ≤ c q − 1 2 ·K[k], T gi (I) =R(K gi ) (7) where (N x ,N y ,N z ) = (H,W,D). We use the retained width along the y-axis. Smaller cutoff values of c remove a larger fraction of peripheral high-frequency coefficients and therefore increase both ringing and loss of fine spatial detail. These two effects are inherent to rectangular frequency truncation and are not treated as separate corruptions. Image-Domain Corruptions Bias field. MRI intensity inhomogeneity is modeled as a smooth multiplicative field applied to the underlying image [23,24]. Let (u x ,u y ,u z ) denote voxel-index coordinates on the lattice Ω = 1,...,H×1,...,W×1,...,D, and let (x 0 ,y 0 ,z 0 ) denote the center of the field. Following previous MRI augmentation 3D Foundation Models Under MRI Artifacts7 studies that employ an elliptic gradient field [21,22,25], we define a paraboloidal spatial field and apply it to the image to obtain the corrupted one as B(u) = 1− λ (u x − x 0 ) 2 H 2 + (u y − y 0 ) 2 W 2 + (u z − z 0 ) 2 D 2 , T bf (I) = B⊙I, (8) where ⊙ denotes element-wise multiplication and λ ≥ 0 controls the strength of the intensity inhomogeneity. The gain equals 1 at the field center and, by fixing the paraboloid center at the volume midpoint, can decrease smoothly to a minimum value of approximately 1− 0.75λ at the most distant (corner) voxels. Rician noise. In conventional magnitude MRI, thermal noise can be approxi- mated as independent zero-mean Gaussian noise in the real and imaginary com- ponents of the complex signal. Taking the magnitude of the resulting complex image leads to Rician-distributed voxel intensities. Assuming a zero-phase noise- less signal, we simulate this process as T ri (I) = q (I + σ ri ε 1 ) 2 + (σ ri ε 2 ) 2 ,ε 1 ,ε 2 i.i.d. ∼ N (0,I), (9) where σ ri is expressed relative to the normalized intensity range and controls noise severity. Unlike additive Gaussian noise, this construction introduces a positive noise floor in low signal regions. Gaussian blur. We model smooth degradation of apparent spatial resolution by convolving the volume with a normalized isotropic 3D Gaussian kernel: G σ bl (u) = 1 (2πσ 2 bl ) 3/2 exp − ∥u∥ 2 2 2σ 2 bl , T bl (I) =G σ bl ∗I,(10) where ∗ denotes convolution and the parameter σ bl , measured in voxels, deter- mines the degree of smoothing. Since the Fourier transform of a Gaussian is also Gaussian, increasing σ bl suppresses high-spatial-frequency image content. Gamma contrast. We simulate global nonlinear changes in image contrast using a power-law transformation: T γ (I) =I γ ,(11) where the exponentiation is applied voxel-wise. For intensities normalized to [0, 1], γ < 1 increases intermediate intensities, whereas γ > 1 suppresses them. Gamma correction is used here as a controlled image-domain contrast perturba- tion and not as a physical forward model of an MRI pulse-sequence parameter. Output range. Some transformations, such as Rician noise and k-space spikes, may produce values outside the initial unit interval. To maintain a valid input range for subsequent processing, the transformed image is clipped voxel-wise: b T a (I) = min max(T a (I), 0), 1 .(12) The same operation is applied at all severity levels before model preprocessing. 8J. A. Mielcarz et al. Table 2: Parameters used for the five predefined corruption settings. For ghosting and Gibbs ringing, stronger corruption corresponds to smaller parameter values. The gamma settings span both sides of the identity transformation γ = 1 and therefore represent contrast conditions rather than a monotonic severity sequence. Artifact (parameter)L1 L2 L3 L4 L5 Rician noise (σ ri )0.07 0.08 0.09 0.10 0.15 Gaussian blur (σ bl )0.60 0.80 1.00 1.20 1.40 Bias field strength (λ)0.60 0.80 1.00 1.20 1.40 Gamma contrast (γ)0.40 0.60 0.80 1.75 2.00 Spike amplitude (α sp )0.45 0.60 0.75 0.90 1.05 Ghosting attenuation (α gh ) 0.70 0.60 0.55 0.50 0.45 Gibbs cutoff width (c)169 153 138 122 107 Severity levels. Each corruption is evaluated at five predefined severity lev- els, denoted L1–L5. The parameter values are summarized in Table 2. They were selected through a preliminary visual inspection to cover a broad range, from subtle degradation to pronounced corruption, while retaining recognizable anatomical structure. The levels should therefore be interpreted as controlled experimental perturbations rather than calibrated clinical image-quality grades. For Rician noise, Gaussian blur, k-space spikes, and the multiplicative bias field, larger parameter values produce stronger perturbations. For periodic k- space ghosting, smaller values of α gh produce stronger attenuation of the peri- odically selected k-space lines. Similarly, for Gibbs ringing, smaller values of the retained width c remove a larger portion of peripheral k-space and therefore pro- duce stronger truncation effects. Finally, gamma includes values both below and above the identity transform γ = 1. They represent different nonlinear contrast conditions rather than a single unidirectional severity sequence: values below one increase intermediate intensities, whereas values above one suppress them. 3.3 Representation-Level Robustness Metrics We evaluate artifact robustness in the representation space of each frozen en- coder. Downstream performance cannot distinguish whether a corruption al- ters the pretrained representation or whether the observed performance change arises from the task-specific prediction model. Representation-level analysis en- ables label-free comparison of the same images before and after corruption. We characterize two complementary properties: preservation of the relational ge- ometry of the embedding space using linear centered kernel alignment (CKA), and changes in its singular-value spectrum using RankMe. We then use Uniform Manifold Approximation and Projection (UMAP) for qualitative visualization. For a given encoder m, let z (m) n = φ m (I n ) ∈ R d m denote the embedding of clean image I n , and let z (m,a,ℓ) n = φ m (T a,ℓ (I n )) ∈ R d m denote the embedding of the corresponding image corrupted by artifact a at setting ℓ. For N matched 3D Foundation Models Under MRI Artifacts9 images, the clean and corrupted embeddings are stacked row-wise as Z (m) = h z (m) 1 ,...,z (m) N i ⊤ , Z (m,a,ℓ) = h z (m,a,ℓ) 1 ,...,z (m,a,ℓ) N i ⊤ . (13) Representational similarity. Linear CKA compares two representations via their centered pairwise inner-product structures [12]. It is invariant to isotropic scaling and orthogonal transformations of the feature coordinates, making it suitable for measuring whether the relative organization of samples is preserved despite changes in the feature basis. Let C N = I N − 1 N 11 ⊤ denote the centering matrix, and define X = C N Z (m) and Y = C N Z (m,a,ℓ) . Linear CKA is computed as CKA(X,Y ) = X ⊤ Y 2 F ∥X ⊤ X∥ F ∥Y ⊤ Y∥ F ,(14) where∥·∥ F denotes the Frobenius norm, and the numerator measures the align- ment between the centered linear Gram matrices of the two representations. CKA lies in [0, 1] for the positive-semidefinite linear kernels here. A value close to one indicates that the pairwise geometry of the clean embedding set is preserved after corruption, up to the invariances of linear CKA. Lower val- ues indicate increasing distortion of the relational structure among images. We compute CKA separately for every encoder, MRI modality, artifact type, and corruption setting, using matched clean and corrupted images in the same order. Spectral effective rank. CKA measures correspondence between clean and corrupted representations but does not describe how their variance is distributed across feature directions. A corruption may concentrate the embeddings into fewer dominant directions without destroying their pairwise organization. We therefore compute RankMe, an entropy-based effective rank derived from the singular-value spectrum of the representation matrix [13]. For a zero-centered embedding matrix Z ∈ R N×d m , let Z = UΣV ⊤ , with singular values Σ = diag(σ 1 ,...,σ r ), where r ≤ min(N,d m ) and σ j ≥ 0. The normalized singular values and RankMe scores are then calculated as p j = σ j P r k=1 σ k + ε,RankMe(Z) = exp − r X j=1 p j logp j , (15) where ε is a small constant used to avoid evaluating log 0. RankMe is defined as RankMe is largest when the singular-value mass is distributed uniformly across the available directions and decreases as the spectrum becomes concen- trated in a smaller number of dominant directions. We therefore use it to quan- tify corruption-associated spectral concentration or expansion of each model’s embedding space. A decrease relative to the clean representation indicates that fewer directions dominate the corrupted embeddings; it should not, however, be interpreted directly as loss of semantic information. 10J. A. Mielcarz et al. Because the maximum attainable effective rank depends on both the num- ber of images and the feature dimensionality, RankMe is interpreted primarily by comparing each model with its own baseline across artifact settings, rather than by treating absolute RankMe values as directly comparable measures of representation quality across different architectures. Qualitative visualization. To visualize corruption-associated changes that may not be fully summarized by CKA or RankMe, we use UMAP [14]. It con- structs a neighborhood graph in the original feature space and optimizes a low- dimensional representation intended to preserve its local neighborhood structure. For each encoder and MRI modality, clean and corrupted embeddings are included in a common UMAP fit so that their two-dimensional coordinates share the same projection. To reduce visual overlap, the qualitative analysis includes the clean images and corruption settings L3–L5. Embeddings generated by the same artifact type are assigned a common visual category, allowing inspection of whether artifact-corrupted images remain close to their clean counterparts or form artifact-specific regions in the projected space. UMAP is run with the default setting and is used only as an exploratory visualization. 3.4 Segmentation Robustness Analysis The representation-level metrics quantify changes within each pretrained encoder but do not establish whether the same image corruptions also affect a clinically relevant prediction task. We therefore conduct a complementary segmentation experiment using TumorSynth [26], an independently developed model for joint segmentation of brain anatomy and tumors. In whole-tumor mode, TumorSynth produces 18 tissue classes (labels 1–17 for healthy tissue and label 18 for whole tumor). It does not use the representations extracted from any of the five en- coders. Consequently, this experiment is not a downstream evaluation of those foundation models. Instead, it provides an independent task-level reference for examining whether artifacts that substantially perturb pretrained representa- tions also tend to disrupt an automated image-analysis pipeline. Let S(I n ) = b Y n denote the whole-label mask predicted by TumorSynth from clean image I n and S(T a,ℓ (I n )) = b Y (a,ℓ) n be the prediction obtained using the same TumorSynth model after applying artifact a at setting ℓ. Segmentation consistency. Because the simulated artifacts modify image intensities without changing the spatial coordinate system or anatomical geom- etry, we evaluate segmentation consistency using the Dice similarity coefficient between b Y (a,ℓ) n and b Y n . A value close to one indicates that TumorSynth produces similar masks from the clean and corrupted images. This measure is interpreted only as prediction stability under corruption and not as segmentation accuracy. 3.5 Evaluation Data We use 95 diffuse-glioma cases from the BraTS-Africa dataset [27], a retrospec- tive multicenter collection of preoperative brain MRI acquired during routine 3D Foundation Models Under MRI Artifacts11 clinical care between 2010 and 2022 at six diagnostic centers in Nigeria using 1.5T scanners. Each case includes T1, T1ce, T2, and FLAIR sequences. The cohort provides a relevant stress test because it captures variation in scanner manufacturer and acquisition protocol from a resource-constrained clinical set- ting; despite image-quality curation, the retained scans exhibit lower signal-to- noise ratio and a higher proportion of artifact-affected voxels than the Western BraTS comparison cohort reported by the dataset authors. This residual hetero- geneity makes BraTS-Africa well suited for evaluating how additional controlled corruptions affect pretrained representations. For each of the four MRI sequences, every clean volume is transformed in- dependently using seven artifact types at five predefined corruption settings. The resulting benchmark therefore contains 95× 4× 7× 5 = 13, 300 corrupted image volumes, in addition to 95× 4 = 380 clean volumes. The benchmark con- tains 3,325 corrupted volumes per MRI sequence. Representative examples of the seven transforms are shown in Appendix. 4 Results and Discussion 4.1 Artifact Sensitivity of Pretrained 3D Encoders Figures 1 and 2 summarize the representation robustness of the five encoders on T1 MRI. Results for T1ce, T2, and FLAIR are provided in Appendix and show broadly consistent model- and artifact-specific trends. Across the evaluated corruptions, 3DINO exhibits the most consistently sta- ble representations. Its linear CKA remains high under Rician noise, Gaus- sian blur, gamma transformation, periodic ghosting, and Gibbs ringing, with a more gradual reduction under k-space spikes and strong bias fields. NeuroVFM also shows high robustness to several structured corruptions, particularly blur, gamma transformation, ghosting, and Gibbs ringing, but is markedly more sen- sitive to k-space spikes. These results suggest that large-scale pretraining and broad anatomical or modality coverage may improve robustness, although archi- tecture, preprocessing, training objective, and dataset scale differ simultaneously and their individual effects cannot be isolated from this comparison. BrainIAC is consistently the most artifact-sensitive encoder. Its CKA is low for most corruptions and approaches zero under frequency-domain perturbations. BrainFM exhibits intermediate but strongly artifact-dependent robustness: it remains comparatively stable at mild corruption settings but degrades progres- sively under blur, bias field, ghosting, and Gibbs truncation. Neuro-SimCLR is comparatively robust to periodic ghosting and Gibbs ringing, but is more sensi- tive to Rician noise, bias-field variation, and k-space spikes. The relative ordering is largely preserved across modalities, although BrainFM and Neuro-SimCLR display greater modality dependence than 3DINO and NeuroVFM. The artifact type is as important as the nominal corruption level. K-space spikes produce substantial representation changes for all encoders except 3DINO, whereas periodic ghosting and Gibbs ringing sharply separate the models: 3DINO, 12J. A. Mielcarz et al. NeuroVFM, and Neuro-SimCLR remain comparatively stable, while BrainFM and especially BrainIAC deteriorate strongly. Among the image-domain corrup- tions, bias field produces the most consistent severity-dependent decline across models. Rician noise strongly affects BrainIAC and BrainFM but has a smaller effect on 3DINO and NeuroVFM. Gaussian blur causes a gradual loss of simi- larity, particularly for BrainIAC and BrainFM. The gamma results should be interpreted differently from the other curves. Because the evaluated parameters include values below and above the identity transformation γ = 1, the five settings do not form a single monotonic severity axis. CKA is generally highest for the conditions closest to the identity transfor- mation and decreases for stronger brightening or darkening, indicating sensitivity to the magnitude of the contrast change rather than to the level index itself. L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Rician noise (CKA) L1L2L3L4L5 Severity level 0 20 40 60 RankMe Rician noise (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Blur (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 RankMe Blur (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Bias field (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 RankMe Bias field (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Gamma contrast (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 RankMe Gamma contrast (RankMe) BrainIACBrainFMNeuroSimCLRNeuroVFM3DINO Fig. 1: Representation robustness to image-domain artifacts on T1 MRI. Linear CKA and RankMe are shown across five corruption settings for each encoder. 3D Foundation Models Under MRI Artifacts13 L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA K-space spike (CKA) L1L2L3L4L5 Severity level 0 20 40 60 RankMe K-space spike (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Motion ghosting (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 RankMe Motion ghosting (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Gibbs ringing (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 RankMe Gibbs ringing (RankMe) BrainIACBrainFMNeuroSimCLRNeuroVFM3DINO Fig. 2: Representation robustness to frequency-domain artifacts on T1 MRI. Linear CKA and RankMe are shown across five corruption settings for each encoder. 4.2 Geometric Drift Without Consistent Dimensional Collapse CKA and RankMe reveal complementary aspects of representation robustness. Across many artifact–model combinations, CKA decreases substantially with corruption strength while RankMe remains nearly constant. This is clear for blur, bias field, k-space spikes, and ghosting, where the relational geometry of the representation space changes without a corresponding reduction in rank. This divergence indicates that the dominant effect of MRI artifacts is not necessarily a collapse onto a lower-dimensional subspace. Instead, the artifacts reorganize the relative positions of subjects within the embedding space while preserving a similar spread of singular values. RankMe alone would fail to iden- tify several prominent robustness failures detected by CKA. Conversely, modest RankMe changes should not be interpreted as information loss, since it measures spectral dispersion rather than the clinical relevance of the retained directions. Absolute RankMe values differ substantially across models, with 3DINO showing the highest values and BrainIAC and Neuro-SimCLR generally showing lower values. These differences should not be interpreted as a ranking of represen- tation quality, because RankMe is affected by feature dimensionality, sample size, and the baseline spectrum of each encoder. The more informative observation is the within-model trajectory across corruption settings. Most models exhibit only modest changes, further supporting the conclusion that artifact sensitivity is expressed primarily as geometric drift rather than dimensional collapse. 14J. A. Mielcarz et al. BrainIACBrainFMNeuroSimCLR NeuroVFM3DINO Rician noise Gaussian blur Bias field Gamma contrast K-space spike Motion ghosting Gibbs ringing Clean Fig. 3: UMAP projections of T1 representations from five encoders under seven arti- facts. Clean and artifact-corrupted samples are embedded jointly for each encoder. The T1 UMAP projections in Figure 3 support this interpretation. BrainFM and NeuroVFM form distinct artifact-associated groups, indicating that their embeddings encode corruption-specific variation. BrainIAC shows strong sepa- ration for spike and Gibbs-corrupted samples, consistent with its low CKA. In contrast, 3DINO displays greater overlap between clean and corrupted samples, in agreement with its generally higher representational similarity. The UMAP and CKA results are not expected to agree perfectly. Linear CKA is computed after centering and is invariant to global translation, isotropic scal- ing, and orthogonal changes of basis, whereas UMAP may visibly separate groups undergoing global distribution shifts. An encoder can therefore retain high CKA while still exhibiting artifact-dependent offsets in a nonlinear two-dimensional projection. The UMAPs are consequently used as qualitative support rather than as an independent quantitative robustness measure. 4.3 Relation to Segmentation Stability Figure 4 reports the Dice similarity between TumorSynth predictions from clean and corrupted inputs. Since TumorSynth is independent of the five evaluated encoders, these results do not constitute downstream validation of the model representations. Instead, they provide a task-level reference for the relative effects of the artifacts reflected in an automated segmentation pipeline. Periodic ghosting produces the largest reduction in prediction consistency, with mean Dice decreasing steadily from 0.79 at L1 to 0.69 at L5. Rician noise causes the second-largest degradation, followed by k-space spikes. Blur, bias field, Gibbs ringing, and gamma transforms retain comparatively high consistency over the evaluated settings with some degradation at stronger corruptions. The strong effect of ghosting is also visible in the representative segmentation maps provided in Appendix, where ghost replicas lead to spatially implausible 3D Foundation Models Under MRI Artifacts15 L1L2L3L4L5 Severity level 0.5 0.6 0.7 0.8 0.9 1.0 Mean Dice Rician noise Gaussian blur Bias field Gamma contrast K-space spike Motion ghosting Gibbs ringing Fig. 4: Segmentation consistency under artifact corruption. Dice similarity between clean and corrupted TumorSynth predictions is averaged across labels 1–18 and cases. predicted regions outside the main anatomy. The remaining corruptions primar- ily induce smaller boundary and regional changes. The task-level ranking only partially matches the representation-level find- ings. Ghosting strongly affects the segmentation model and also produces large representation changes for BrainFM and BrainIAC, but it has little effect on 3DINO, NeuroVFM, and Neuro-SimCLR. Conversely, k-space spikes strongly perturb several encoder representations while producing a moderate segmenta- tion consistency loss. These differences show that artifact sensitivity is model- and task-dependent: an artifact may alter a pretrained representation without producing the same degree of instability in a separate segmentation pipeline. 5 Conclusion We presented a controlled study of how MRI artifacts affect the representa- tions learned by five pretrained 3D medical-image encoders. Robustness varied substantially across models and corruption types: 3DINO was the most con- sistently stable, whereas other encoders showed distinct artifact-specific sensi- tivities. Across many conditions, CKA declined while RankMe remained com- paratively stable, indicating that artifacts often distort representation geometry without causing dimensional collapse. The segmentation analysis further showed that task-level sensitivity only par- tially follows representation-level robustness. These findings demonstrate that neither pretraining scale nor domain specialization alone guarantees robustness to MRI artifacts, and highlight the need to assess representation stability before deploying foundation models across heterogeneous clinical imaging conditions. Acknowledgements This project is supported by the Pioneer Centre for AI, funded by the Danish National Research Foundation (grant number P1). 16J. A. Mielcarz et al. References 1. Xu, T., Hosseini, S., Anderson, C., Rinaldi, A., Krishnan, R.G., Martel, A.L., Goubran, M.: A generalizable 3D framework and model for self-supervised learning in medical imaging. npj Digital Medicine 8(1) (2025) 639 1, 4, 5 2. Tak, D., Garomsa, B.A., Zapaishchykova, A., Chaunzwa, T.L., Climent Pardo, J.C., Ye, Z., Zielke, J., Ravipati, Y., Pai, S., Vajapeyam, S., et al.: A generalizable foundation model for analysis of human brain MRI. Nature Neuroscience (2026) 1–12 1, 4, 5 3. Kaczmarek, E., Szeto, J., Nichyporuk, B., Arbel, T.: Building a general SimCLR self-supervised foundation model across neurological diseases to advance 3D brain MRI diagnoses. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. (2025) 1310–1319 1, 4, 5 4. Liu, P., Puonti, O., Hu, X., Gopinath, K., Sorby-Adams, A., Alexander, D.C., Kimberly, W.T., Iglesias, J.E.: A modality-agnostic multi-task foundation model for human brain imaging. arXiv preprint arXiv:2509.00549 (2025) 1, 4, 5 5. Kondepudi, A., Rao, A., Zhao, C., Lyu, Y., Harake, S., Banerjee, S., Ogle, J., Joshi, R., Meissner, A.K., Hou, X., et al.: Health system learning enables generalist neuroimaging models. Nature Medicine (2026) 1–7 1, 4, 5 6. Krupa, K., Bekiesińska-Figatowska, M.: Artifacts in magnetic resonance imaging. Polish Journal of Radiology 80 (2015) 93 2, 5, 6 7. Andre, J.B., Bresnahan, B.W., Mossa-Basha, M., Hoff, M.N., Smith, C.P., Anzai, Y., Cohen, W.A.: Toward quantifying the prevalence, severity, and cost associated with patient motion during clinical MR examinations. Journal of the American College of Radiology 12(7) (2015) 689–695 2 8. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11) (2020) 665–673 2 9. Glocker, B., Robinson, R., Castro, D.C., Dou, Q., Konukoglu, E.: Machine learning with multi-site imaging data: An empirical study on the impact of scanner effects. arXiv preprint arXiv:1910.04597 (2019) 2 10. Kriegeskorte, N., Mur, M., Bandettini, P.A.: Representational similarity analysis- connecting the branches of systems neuroscience. Frontiers in Systems Neuro- science 2 (2008) 249 2 11. Raghu, M., Gilmer, J., Yosinski, J., Sohl-Dickstein, J.: SVCCA: Singular vec- tor canonical correlation analysis for deep learning dynamics and interpretability. Advances in Neural Information Processing Systems 30 (2017) 2 12. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network representations revisited. In: International Conference on Machine Learning, PMlR (2019) 3519–3529 2, 9 13. Garrido, Q., Balestriero, R., Najman, L., Lecun, Y.: RankMe: Assessing the down- stream performance of pretrained self-supervised representations by their rank. In: International Conference on Machine Learning, PMLR (2023) 10929–10974 2, 9 14. McInnes, L., Healy, J., Saul, N., Großberger, L.: UMAP: Uniform manifold ap- proximation and projection. Journal of Open Source Software 3(29) (2018) 861 2, 10 15. Del Giudice, M.: Effective dimensionality: A tutorial. Multivariate Behavioral Research 56(3) (2021) 527–542 2 16. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINOv2: Learning 3D Foundation Models Under MRI Artifacts17 robust visual features without supervision. Transactions on Machine Learning Research Journal (2024) 4 17. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: International Conference on Machine Learning, PmLR (2020) 1597–1607 4 18. Wood, M.L., Henkelman, R.M.: MR image artifacts from periodic motion. Medical Physics 12(2) (1985) 143–151 6 19. Block, K.T., Uecker, M., Frahm, J.: Suppression of MRI truncation artifacts using total variation constrained data extrapolation. International Journal of Biomedical Imaging 2008(1) (2008) 184123 6 20. Kellner, E., Dhital, B., Kiselev, V.G., Reisert, M.: Gibbs-ringing artifact removal based on local subvoxel-shifts. Magnetic Resonance in Medicine 76(5) (2016) 1574– 1581 6 21. Mehdipour Ghazi, M., Nielsen, M.: FAST-AID Brain: Fast and accurate seg- mentation tool using artificial intelligence developed for brain. In: Scandinavian Conference on Image Analysis, Springer (2025) 161–176 6, 7 22. Van Eeden Risager, K., Gholamalizadeh, T., Mehdipour Ghazi, M.: Non-reference quality assessment for medical imaging: Application to synthetic brain MRIs. In: MICCAI Workshop on Deep Generative Models, Springer (2024) 191–201 6, 7 23. Sled, J.G., Zijdenbos, A.P., Evans, A.C.: A nonparametric method for automatic correction of intensity nonuniformity in MRI data. IEEE Transactions on Medical Imaging 17(1) (1998) 87–97 6 24. Pérez-García, F., Sparks, R., Ourselin, S.: TorchIO: a Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images in deep learning. Computer Methods and Programs in Biomedicine 208 (2021) 106236 6 25. Hui, C., Zhou, Y.X., Narayana, P.: Fast algorithm for calculation of inhomogeneity gradient in magnetic resonance imaging data. Journal of Magnetic Resonance Imaging 32(5) (2010) 1197–1208 7 26. Wu, J., Billot, B., Zhao, F., Li, J., Liu, Y., Zhuo, Z., Pontillo, G., Liu, X., Lin, S., Zhang, X., et al.: TumorSynth: Integrated brain tumor and tissue segmentation on brain MRI scans of any resolution and contrast. Radiology: Imaging Cancer 8(2) (2026) e250222 10 27. Adewole, M., Rudie, J.D., Gbadamosi, A., Zhang, D., Raymond, C., Ajigboto- shso, J., Toyobo, O., Aguh, K., Omidiji, O., Akinola, R., et al.: The BraTS-Africa dataset: expanding the brain tumor segmentation data to capture African popula- tions. Radiology: Artificial Intelligence 7(4) (2025) e240528 10 Appendix A Utilized Encoders A.1 3DINO 3DINO adapts the DINO self-distillation framework to volumetric medical im- ages using complementary image-level and masked patch-level objectives. Its 3D ViT-L/16 encoder was pretrained on approximately 100,000 unlabeled vol- umes from 35 studies, comprising MRI, CT, and brain PET images from more than 10 anatomical regions. It therefore represents a broadly pretrained medi- cal encoder rather than a brain- or MRI-specific model. Following the released inference pipeline, each volume is resized to 112 3 voxels and intensities are lin- early rescaled to [−1, 1]. We use the 1024-dimensional class token from the final transformer layer, that is φ 3DINO (I)∈ R 1024 . A.2 BrainIAC BrainIAC is a brain-MRI-specific foundation model based on a 3D ViT-B/16 encoder pretrained with the SimCLR objective. The model was pretrained on 32,015 multiparametric brain MRI scans selected from 16 datasets spanning 10 neurological conditions. The larger collection of 48,965 scans reported in the study includes data used across model development and downstream evaluation and does not correspond solely to the self-supervised pretraining set. Following the released inference pipeline, volumes are resized to 96 3 voxels and z-score normalized using the non-zero voxels. We extract the 768-dimensional class token from the final transformer layer, so φ BrainIAC (I)∈ R 768 . A.3 BrainFM BrainFM differs from the other evaluated models by using a 3D U-Net-like, fully convolutional architecture. It is trained jointly for image synthesis, anatomical segmentation, cortical-distance estimation, bias-field estimation, and atlas reg- istration. Its training strategy combines on-the-fly intra-subject synthesis with real–synthetic mixing to expose the model to variations in modality, contrast, resolution, orientation, deformation, and image quality. In Setup I, the 11 source datasets contain 8,675 modality-specific images before the dataset-wise 90/10 training–test split, corresponding to approximately 7,800 real training images. This count treats different modalities from the same subject as separate images and excludes the unlimited synthetic samples generated during training. Following the published preprocessing protocol, input images are skull-stripped and resampled to 1-m isotropic resolution. BrainFM is fully convolutional and consequently does not require a single fixed matrix size; the released implemen- tation supports whole-volume and tiled inference. Input volumes are min–max 3D Foundation Models Under MRI Artifacts19 normalized over the full volume range. Since BrainFM is not originally defined as a global image encoder, we extract the released encoder feature map and average it over its spatial dimensions. The resulting representation for the bot- tleneck feature map (deepest encoder layer) is therefore φ BrainFM (I) ∈ R 2048 . The global pooling operation is specific to our representation analysis and is not part of BrainFM’s original task-specific inference procedure. A.4 NeuroVFM NeuroVFM is a generalist neuroimaging encoder pretrained on 5.24 million clin- ical MRI and CT volumes from 566,915 studies acquired over more than two decades of routine clinical care. It uses a 3D ViT-B encoder with 4× 16× 16 volumetric patches and is trained using volumetric JEPA. Vol-JEPA learns to predict latent representations of masked target regions from visible context re- gions. The preprocessing pipeline quantizes the clinical volumes before tokeniza- tion. Because the model was designed for heterogeneous clinical series, it does not require all input volumes to be resized to one fixed 3D matrix. We use the official preprocessing object associated with the released checkpoint and ob- tain a volume-level representation by averaging the final-layer patch tokens as φ NeuroVFM (I)∈ R 768 . A.5 Neuro-SimCLR Neuro-SimCLR uses a 3D ResNet-18 encoder pretrained with SimCLR on 44,958 structural brain MRI scans from 18,759 patients across 11 public datasets and multiple neurological conditions. It complements BrainIAC by combining an instance-contrastive objective with a convolutional rather than transformer- based backbone. We follow the released TurboPrep-based preprocessing and in- ference pipeline. The image axes are reordered to the orientation expected by the encoder, the brain volume is cropped using the coordinates prescribed by the released implementation, and intensities are z-score normalized over foreground voxels. No additional isotropic resizing is introduced in our experiments. The 512-dimensional output of the encoder’s global average-pooling layer is used as the representation to obtain φ NeuroSimCLR (I)∈ R 512 . B Additional Results 20J. A. Mielcarz et al. L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Rician noise (CKA) L1L2L3L4L5 Severity level 0 20 40 60 RankMe Rician noise (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Blur (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 60 RankMe Blur (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Bias field (CKA) L1L2L3L4L5 Severity level 20 40 60 RankMe Bias field (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Gamma contrast (CKA) L1L2L3L4L5 Severity level 20 40 60 RankMe Gamma contrast (RankMe) BrainIACBrainFMNeuroSimCLRNeuroVFM3DINO Fig. 5: Representation robustness to image-domain artifacts on T1ce MRI. Linear CKA and RankMe are shown across five corruption settings for each encoder. 3D Foundation Models Under MRI Artifacts21 L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA K-space spike (CKA) L1L2L3L4L5 Severity level 20 40 60 RankMe K-space spike (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Motion ghosting (CKA) L1L2L3L4L5 Severity level 20 40 60 RankMe Motion ghosting (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Gibbs ringing (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 60 RankMe Gibbs ringing (RankMe) BrainIACBrainFMNeuroSimCLRNeuroVFM3DINO Fig. 6: Representation robustness to frequency-domain artifacts on T1ce MRI. Linear CKA and RankMe are shown across five corruption settings for each encoder. 22J. A. Mielcarz et al. L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Rician noise (CKA) L1L2L3L4L5 Severity level 0 20 40 60 RankMe Rician noise (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Blur (CKA) L1L2L3L4L5 Severity level 0 10 20 30 40 50 RankMe Blur (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Bias field (CKA) L1L2L3L4L5 Severity level 0 20 40 60 RankMe Bias field (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Gamma contrast (CKA) L1L2L3L4L5 Severity level 0 10 20 30 40 50 RankMe Gamma contrast (RankMe) BrainIACBrainFMNeuroSimCLRNeuroVFM3DINO Fig. 7: Representation robustness to image-domain artifacts on T2 MRI. Linear CKA and RankMe are shown across five corruption settings for each encoder. 3D Foundation Models Under MRI Artifacts23 L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA K-space spike (CKA) L1L2L3L4L5 Severity level 0 20 40 60 RankMe K-space spike (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Motion ghosting (CKA) L1L2L3L4L5 Severity level 0 10 20 30 40 50 RankMe Motion ghosting (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Gibbs ringing (CKA) L1L2L3L4L5 Severity level 0 10 20 30 40 50 RankMe Gibbs ringing (RankMe) BrainIACBrainFMNeuroSimCLRNeuroVFM3DINO Fig. 8: Representation robustness to frequency-domain artifacts on T2 MRI. Linear CKA and RankMe are shown across five corruption settings for each encoder. 24J. A. Mielcarz et al. L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Rician noise (CKA) L1L2L3L4L5 Severity level 20 40 60 RankMe Rician noise (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Blur (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 RankMe Blur (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Bias field (CKA) L1L2L3L4L5 Severity level 20 40 60 RankMe Bias field (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Gamma contrast (CKA) L1L2L3L4L5 Severity level 20 40 60 RankMe Gamma contrast (RankMe) BrainIACBrainFMNeuroSimCLRNeuroVFM3DINO Fig. 9: Representation robustness to image-domain artifacts on FLAIR MRI. Linear CKA and RankMe are shown across five corruption settings for each encoder. 3D Foundation Models Under MRI Artifacts25 L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA K-space spike (CKA) L1L2L3L4L5 Severity level 0 20 40 60 RankMe K-space spike (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Motion ghosting (CKA) L1L2L3L4L5 Severity level 0 10 20 30 40 50 RankMe Motion ghosting (RankMe) L1L2L3L4L5 Severity level 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Linear CKA Gibbs ringing (CKA) L1L2L3L4L5 Severity level 10 20 30 40 50 RankMe Gibbs ringing (RankMe) BrainIACBrainFMNeuroSimCLRNeuroVFM3DINO Fig. 10: Representation robustness to frequency-domain artifacts on FLAIR MRI. Lin- ear CKA and RankMe are shown across five corruption settings for each encoder. BrainIACBrainFMNeuroSimCLR NeuroVFM3DINO Rician noise Gaussian blur Bias field Gamma contrast K-space spike Motion ghosting Gibbs ringing Clean Fig. 11: UMAP projections of T1ce representations from five encoders under seven artifacts. Clean and artifact-corrupted samples are embedded jointly for each encoder. 26J. A. Mielcarz et al. BrainIACBrainFMNeuroSimCLR NeuroVFM3DINO Rician noise Gaussian blur Bias field Gamma contrast K-space spike Motion ghosting Gibbs ringing Clean Fig. 12: UMAP projections of T2 representations from five encoders under seven ar- tifacts. Clean and artifact-corrupted samples are embedded jointly for each encoder. BrainIACBrainFMNeuroSimCLR NeuroVFM3DINO Rician noise Gaussian blur Bias field Gamma contrast K-space spike Motion ghosting Gibbs ringing Clean Fig. 13: UMAP projections of FLAIR representations from five encoders under seven artifacts. Clean and artifact-corrupted samples are embedded jointly for each encoder. 3D Foundation Models Under MRI Artifacts27 Rician noise cleanL1L2L3L4L5 Gaussian blur Bias field Gamma contrast Gibbs ringing Motion ghosting K-space spike Fig. 14: Middle slices of representative T1 MRI under the seven simulated artifacts. Each row corresponds to one artifact type, and columns show the clean image followed by the five predefined corruption settings. 28J. A. Mielcarz et al. Rician noise cleanL1L2L3L4L5 Gaussian blur Bias field Gamma contrast K-space spike Motion ghosting Gibbs ringing Fig. 15: Middle slices of representative TumorSynth predictions under artifact corrup- tion. Each row corresponds to one artifact type, and columns show the 18 predicted anatomical and tumor classes for the clean image and five corruption settings.