BrainNorm: A Foundation Model that knows Normal via Semantic Atlas Pretraining

Madhumitha Venkatesh; Shanawaj S Madarkar; Konda Reddy Mopuri

cs.CVAug 18, 2026

BrainNorm: A Foundation Model that knows Normal via Semantic Atlas Pretraining

Authors: Madhumitha Venkatesh, Shanawaj S Madarkar, Konda Reddy Mopuri

Organizations: Indian Institute of Technology Hyderabad

Abstract

We introduce BrainNorm, a normative foundation model, trained and tested on ~66,000 T1-weighted structural MRI (T1w sMRI) scans. By leveraging language-image style contrastive pretraining on healthy cohorts across ages, BrainNorm learns a Semantic Atlas Latent space (SAL), where each scan is represented as a set of atlas-parcel embeddings. This yields parcel-specific healthy aging template trajectories that support age-consistent template matching and localized deviation scoring relative to a subject's chronological age. Across 6 downstream cohorts, BrainNorm demonstrates generalization evaluated across 25 task-setting combinations spanning age estimation, brain-age gap estimation, parcel identification, and single- & multi-disease classification tasks under direct inference, zero-shot, few-shot & full-data linear-probe settings. The resulting deviation patterns in SAL space enable zero-shot tasks for disease prediction using parcel-wise abnormalities. Fine-tuning on healthy-only cohorts of downstream datasets further improves the performance of various tasks. Across all classification tasks, linear probing on BrainNorm's frozen embeddings outperforms 9 baselines finetuned under end-to-end supervision. Furthermore, the localized deviations identified by BrainNorm across various neurodegenerative disorders closely align with established neurodegeneration pathology in clinical literature.

Figures & tables

Figure 1: BrainNorm Framework: A representation learning technique that models healthy aging as anatomically grounded representations exclusively from healthy subjects. By utilizing frozen embeddings within a Semantic Atlas Latent Space, BrainNorm enables zero-shot, few-shot, and full linear probing for downstream prediction and classification tasks on mixed clinical cohorts.

Figure 2: BrainNorm pretraining framework for a sample batch of scans. BrainNorm encodes a scan as a set of parcel embeddings in a shared semantic atlas latent (SAL) space, rather than compressing it into a single global representation (shown as a single gray vector). These embeddings are aligned to an age-indexed, parcel-specific healthy template bank initialized from structured textual priors and refined through a learnable Template Mapper.

Figure 3: BrainNorm’s template and parcel embeddings of healthy samples in the SAL space are clustered by anatomical parcel and exhibit age-ordered organization (using t-SNE plots). The visualization zooms into two representative parcels, the right hippocampus and lateral ventricle, showing age-wise ordering of parcel embeddings within each anatomical region.

Type	Metric	Mode	UKB	MCSA	ADNI	AIBL	MIRIAD	NIFD	PPMI
T $\rightarrow$ P	Acc@1	No FT	90.87	67.33	$69.50_{\pm 0.35}$	$68.71_{\pm 0.65}$	$73.93_{\pm 1.18}$	$70.62_{\pm 0.59}$	$67.18_{\pm 0.85}$
	Acc@1	FT	90.32	76.06	$79.88_{\pm 1.11}$	$80.01_{\pm 1.25}$	$83.49_{\pm 2.88}$	$82.06_{\pm 0.64}$	$79.47_{\pm 1.09}$
	Acc@5	No FT	99.02	85.19	$87.45_{\pm 0.24}$	$86.20_{\pm 0.48}$	$90.33_{\pm 0.69}$	$86.78_{\pm 0.50}$	$84.93_{\pm 0.58}$
	Acc@5	FT	99.04	94.54	$96.32_{\pm 0.39}$	$96.01_{\pm 0.50}$	$97.74_{\pm 0.83}$	$96.71_{\pm 0.29}$	$95.59_{\pm 0.56}$
P $\rightarrow$ T	Acc@1	No FT	97.17	75.02	$77.54_{\pm 0.31}$	$76.64_{\pm 0.83}$	$81.30_{\pm 1.01}$	$63.84_{\pm 3.11}$	$74.84_{\pm 0.60}$
	Acc@1	FT	96.72	86.46	$89.64_{\pm 0.78}$	$89.89_{\pm 1.18}$	$93.20_{\pm 1.76}$	$91.35_{\pm 0.37}$	$88.71_{\pm 0.54}$

Table 1: Parcel Identification Performance (Text(region)

\rightarrow

Parcel & Parcel

\rightarrow

Text(region)) using CN subjects across pretraining datasets and downstream mixed datasets (

mean_{\pm std}

over 5-folds). Performance improves with finetuning. [FT: Finetuned on CN-only subjects of mixed cohorts]

Figure 4: A. BrainNorm’s zero-shot performance comparison using anomaly scores and BAG across all classification tasks. These plots illustrate the performance gains of BrainNorm achieved by fine-tuning on CN subjects alone, along with superior performance compared to other supervised baselines across all tasks. Notably, BrainNorm’s disease-prior restricted anomaly technique enables zero-shot learning for multi-disease classification. B. BAG distribution improves with finetuning across all datasets and groups.

Figure 5: A. Classification performance of BrainNorm’s full-data linear probing (LP) results of M1 & M2 techniques outperforms all end-to-end supervised (finetuned) baselines. B. Few-shot linear probing of BrainNorm shows superior performance with both M1 and M2 representations compared to end-to-end finetuning of pretrained baselines, indicating stronger transferability under small-label budgets. BrainNorm’s parcel deviation-based representations ( M2 ) shows better results in multiple tasks compared to utilizing only parcel embeddings ( M2 ). Plotted values are mean across 5-folds for mixed cohorts.

Figure 6: BrainNorm’s age estimation outperforms other supervised baselines’ age prediction performance.

Figure 7: A. Zero-shot identification and visualization of the top- $k$ anomalous brain parcels across disease cohorts in sagittal and coronal planes. Higher deviation scores from the normative reference are represented by darker shades. B. KDE density plot of UMAP using the parcel residuals ( $\mathbf{x}_{M2}$ ).

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	#scans	#subjects	Age (y)	Gender	Field strength	Manufacturer
			(min,max, $\text{mean}_{\pm\text{std}}$ )	(M / F)	(1.5T / 3.0T)	(Siemens/GE/Philips)
UKB (CN)	48,018	43,631	$(44.0,83.0,64.23_{\pm 7.70})$	20,592 / 23,039	0 / 43,631	43,631 / 0 / 0
MCSA (CN)	2,662	1,611	$(49.0,89.9,71.83_{\pm 9.83})$	853 / 758	0 / 1,611	0 / 1,611 / 0
ADNI	5,627	1,703	$(52.4,93.1,72.93_{\pm 7.87})$	734 / 969	429 / 1,411	1,063 / 476 / 289
ADNI (MCI)	5,308	687	$(55.6,91.6,74.18_{\pm 7.44})$	408 / 279	275 / 512	389 / 258 / 134
AIBL	1,084	599	$(56.3,93.0,73.36_{\pm 6.42})$	255 / 343	107 / 541	599 / 0 / 0

Appendix

Table A1: Cohort statistics of all datasets used in our experiments.

Figure A1: Details of the Template Mapper module.

Dataset	Ablation	Text $\rightarrow$ Parcel		Parcel $\rightarrow$ Text		Age-MAE $\downarrow$	Pearson- $r$ $\uparrow$
Dataset	Ablation	Acc@1 $\uparrow$	Acc@5 $\uparrow$	Acc@1 $\uparrow$	Acc@5 $\uparrow$	Age-MAE $\downarrow$	Pearson- $r$ $\uparrow$
UKB	1	90.87	99.02	97.17	99.80	4.14	0.85
UKB	2	90.99	99.09	97.99	99.99	4.98	0.77
UKB	3	0.48	2.34	0.32	2.05	7.74	0.44
MCSA	1	67.33	85.19	75.02	89.54	3.94	0.88
MCSA	2	70.53	86.60	74.19	88.12	6.78	0.84

Appendix

Table A2: Ablation of the Template Mapper and AgeSoftCE objective. (1) SigLIP + Template Mapper + AgeSoftCE (ours); (2) SigLIP + Template Mapper, without AgeSoftCE; (3) SigLIP without Template Mapper and AgeSoftCE

Dataset	Initialization	Text $\rightarrow$ Parcel		Parcel $\rightarrow$ Text		Age-MAE $\downarrow$	Pearson- $r$ $\uparrow$
Dataset	Initialization	Acc@1 $\uparrow$	Acc@5 $\uparrow$	Acc@1 $\uparrow$	Acc@5 $\uparrow$	Age-MAE $\downarrow$	Pearson- $r$ $\uparrow$
UKB	Random	55.93	58.34	56.02	58.19	10.32	0.68
UKB	Categorical	99.99	99.99	99.98	99.99	10.05	0.73
UKB	Learnable	91.20	99.21	96.91	97.71	7.29	0.79
UKB	Qwen3	90.87	99.02	97.17	99.80	4.14	0.85
MCSA	Random	15.37	17.80	14.97	17.65	17.40	0.67

Appendix

Table A3: Ablation of text-prior initialization strategies. We compare Qwen3-based template initialization against randomly initialized embeddings, fixed categorical embeddings, and learnable parcel embeddings with FiLM-based age conditioning.

Dataset	Template	Text $\rightarrow$ Parcel		Parcel $\rightarrow$ Text		Age-MAE $\downarrow$	Pearson- $r$ $\uparrow$
Dataset	Template	Acc@1 $\uparrow$	Acc@5 $\uparrow$	Acc@1 $\uparrow$	Acc@5 $\uparrow$	Age-MAE $\downarrow$	Pearson- $r$ $\uparrow$
UKB	1	90.87	99.02	97.17	99.80	4.14	0.85
UKB	2	86.10	97.87	94.49	99.66	4.47	0.85
UKB	3	85.96	97.89	93.43	99.58	3.99	0.85
UKB	4	80.39	96.57	93.48	99.45	3.91	0.85
MCSA	1	67.33	85.19	75.02	89.54	3.94	0.88

Appendix

Table A4: Robustness to semantic rephrasing of text queries. We evaluate four semantically equivalent prompt formulations that vary in syntax and word order.

	Parcel Clustering			Age Ordering Within Parcel
Encoder	Silhouette $\uparrow$	Hit@1 $\uparrow$	NMI $\uparrow$	Spearman $\uparrow$	Pair Order Acc $\uparrow$	kNN Age MAE $\downarrow$
Qwen3-Embedding-8B Yang et al. (2025)	0.4637	0.9999	0.9476	0.9682	0.9311	0.5114
MedEmbed-small Balachandran (2024)	0.6667	1.0000	0.9962	0.5483	0.6922	1.9061
MedSigLIP Sellergren et al. (2025)	0.1224	0.7284	0.8790	0.2906	0.5971	6.0694
BiomedCLIP Zhang et al. (2025b)	0.6133	0.9999	0.9863	0.3816	0.6323	2.4691
Hashing+SVD	0.4440	0.9998	0.9939	0.0024	0.5010	3.4503

Appendix

Table A5: Choice of Text Encoder. Comparison of text encoders for the desired embedding geometry. The representation should (1) cluster samples by parcel identity and (2) exhibit a smooth age trajectory within each parcel . Metrics on the left measure parcel clustering, while metrics on the right measure age ordering within parcels. Best values are highlighted in bold.

Dataset	Mode	CN data	Text $\rightarrow$ Parcel		Parcel $\rightarrow$ Text
Dataset	Mode	CN data	Acc@1 $\uparrow$	Acc@5 $\uparrow$	Acc@1 $\uparrow$	Acc@5 $\uparrow$
UKB	No FT	UKB	$90.87$	$99.02$	$97.17$	$99.80$
	FT	UKB	$90.32$	$99.04$	$96.72$	$99.76$
MCSA	No FT	UKB	$67.33$	$85.19$	$75.02$	$89.54$
MCSA	FT	All	$76.06$	$94.54$	$86.46$	$97.81$
ADNI	No FT	UKB	$69.50_{\pm 0.35}$	$87.45_{\pm 0.24}$	$77.54_{\pm 0.31}$	$91.17_{\pm 0.23}$

Appendix

Table A6: Parcel identification on held-out test sets and downstream datasets using only CN subjects (5-fold mean ±std where available). PT: pretrained encoder used directly for inference. FT: finetuned only on the CN subset of the corresponding dataset or finetuned on pooled CN data from ADNI+AIBL+MIRIAD+NIFD+PPMI (All).

Class	Parcel IDs
AD/MCI	38, 39, 40, 41, 36, 37, 68, 69, 66, 67, 62, 63, 84, 85, 18, 19, 216, 217
PD	156, 157, 158, 159, 74, 75, 72, 73, 76, 77, 154, 155, 160, 161, 162, 163, 164, 165, 0, 1, 14, 15, 58, 59, 172, 173
FTD	2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 82, 83, 86, 87, 80, 81, 84, 85, 32, 33, 146, 147, 148, 149, 150, 151, 24, 25, 26, 27, 28, 29, 30, 31, 20, 21, 152, 153

Appendix

Table A7: Disease-specific prior parcel IDs used for region-restricted analysis.

Figure A2: Zero-shot performance with increasing top-K parcels considered for both with and without fine-tuning on the full dataset.

Task	HBR	HBR	BrainNorm	BrainNorm
	Direct Transfer	CN-only Adaptation	Direct Transfer	CN-only Fine-tuning
AD/CN [ADNI]	75.20	72.01	54.78	72.37
AD/CN [MIRIAD]	90.81	88.97	71.00	93.51
AD/CN [AIBL]	71.04	72.51	49.55	70.70
sMCI/pMCI [ADNI]	58.81	59.89	47.33	62.79
PD/CN [PPMI]	51.86	49.96	49.60	53.02

Appendix

Table A8: Zero-shot disease classification using subject-level anomaly scores. Comparison between HBR and BrainNorm under direct transfer from UKB and target-cohort adaptation using only healthy controls. Results are reported as AUC (%). Best and second-best results for each task may be highlighted in bold and italics, respectively.

Task	HBR BAcc	HBR AUC	BrainNorm BAcc	BrainNorm AUC	$\Delta$ BAcc	$\Delta$ AUC
AD/CN [ADNI]	82.29	90.25	86.85	94.27	+4.56	+4.02
AD/CN [AIBL]	83.53	91.06	85.91	93.39	+2.38	+2.33
AD/CN [MIRIAD]	88.47	94.57	92.72	98.26	+4.25	+3.69
sMCI/pMCI [ADNI]	67.47	74.55	71.96	80.75	+4.49	+6.20
PD/CN [PPMI]	55.87	60.42	59.08	66.12	+3.21	+5.70
FTD/CN [NIFD]	70.14	75.42	81.82	89.15	+11.68	+13.73

Appendix

Table A9: Linear probing of frozen normative-deviation representations. No CN-only adaptation or downstream representation fine-tuning is performed. BrainNorm M2 uses frozen parcel-deviation representations, while HBR uses concatenated regional morphometric deviation scores.

\Delta

denotes BrainNorm minus HBR. Results are reported in %.

Figure A3: t-SNE visualizations of text priors ( $\mathbf{e}^{\mathrm{raw}}_{a,p}$ ), learned templates ( $\mathbf{t}_{a,p}$ ), and parcel embeddings ( $\mathbf{p}_{i,p}$ ) on held-out UKB Sudlow et al. (2015) test data. The learned SAL space exhibits visible anatomical clustering and ribbon-like derived structure of age-ordering within each parcel.

Figure A4: UMAP scatter plots and KDE density maps visualize the distribution of scan-level residual embeddings across neurodegenerative cohorts. The diagnosis-based scatter plot is colored by clinical group with an age gradient, revealing an apparent age-related organization in the embedding space, where relatively younger scans tend to lie toward the right and older scans toward the left. The diagnosis-specific KDE maps further show structured group-wise density patterns, suggesting that the residual embedding space captures disease-associated deviations from the normative representation. In contrast, the manufacturer-, sex-, and field-strength-based scatter plots and KDE maps show that scans from different vendors, sex categories, and field strengths are broadly dispersed across the embedding space rather than forming distinct confounder-specific clusters. This suggests that the observed diagnostic organization is unlikely to be primarily driven by scanner manufacturer, sex, or field-strength effects, supporting the interpretation that the learned SAL space captures biologically meaningful anomaly deviations with limited evidence of these confounding effects.

Figure A5: Scan-level interpretability across neurodegenerative disease. Representative correct and incorrect predictions are shown for AD, pMCI, sMCI, FTD, and PD. Correct predictions correspond to AD, pMCI, sMCI, FTD, and PD subjects assigned to their respective ground-truth classes, whereas wrong predictions correspond to AD, FTD, and PD subjects misclassified as CN, pMCI subjects misclassified as sMCI, and sMCI subjects misclassified as pMCI. For each subject, the spatial map shows the highest-ranking parcel-wise deviations from the age-matched normative representation, with the image colors normalized independently within that subject to emphasize the spatial distribution of its regional abnormalities. The adjacent bar plot reports the corresponding raw parcel anomaly scores, with bar colors mapped using a common anomaly-score scale across all scans to enable direct comparison of deviation magnitude between subjects. Across correct predictions, the highlighted parcels show strong disease-relevant deviations. In contrast, wrong predictions generally retain anatomically relevant regional patterns but exhibit anomaly magnitudes that are shifted in the direction of the predicted class: AD, FTD, PD, and pMCI false negatives show comparatively weaker deviations, whereas sMCI subjects misclassified as pMCI show comparatively stronger deviations.

Dataset	Method	Mode	CN Train/FT data	MAE $\downarrow$	RMSE $\downarrow$	Pearson $r\uparrow$	$R^{2}\uparrow$
UKB	3D-ViT	Full Train	UKB	$4.23$	$5.65$	$0.77$	$0.58$
	BrainIAC	Full FT	UKB	$6.76$	$8.30$	$0.16$	$0.16$
	BrainMVP	Full FT	UKB	$4.15$	$\underline{5.21}$	$0.83$	$\mathbf{0.79}$
	BrainHarmony	Full FT	UKB	$\underline{4.15}$	$\mathbf{5.20}$	$0.83$	$0.76$
	RadFM	Full FT	UKB	$4.83$	$5.62$	$0.82$	$\underline{0.77}$
	NeuroVFM	Full FT	UKB	$4.25$	$5.11$	$0.81$	$0.72$

Appendix

Table A10: BrainNorm’s Age estimation on held-out test sets and downstream datasets using only CN subjects vs Prior works age prediction with full supervision. FT denotes finetuning on the CN subset of the corresponding dataset or pooled CN data from ADNI+AIBL+MIRIAD+NIFD+PPMI (All). Values are reported as mean ±std over 5 folds where available.

Dataset/Task	Method	Full FT	No FT	FT (own)	FT (all CN)
ADNI [AD/CN]	A	-	$54.78_{\pm 4.04}$	$72.80_{\pm 4.39}$	$72.37_{\pm 5.07}$
	P+A	-	$64.61_{\pm 3.00}$	$79.27_{\pm 3.94}$	$79.16_{\pm 3.79}$
	BAG	-	$82.80_{\pm 1.95}$	$86.19_{\pm 2.62}$	$86.02_{\pm 2.12}$
	BrainIAC+BAG	$46.68_{\pm 3.97}$	-	-	-
	BrainMVP+BAG	$57.58_{\pm 2.20}$	-	-	-
	BrainHarmony+BAG	$66.65_{\pm 3.78}$	-	-	-

Appendix

Table A11: Zero-shot disease classification performance (AUC in %) across binary and multi-class tasks using top-k@20. Results are reported as mean ±std . FT (own): finetuning on the corresponding dataset CN only. FT (all CN): finetuning on pooled CN data from ADNI+AIBL+MIRIAD+NIFD+PPMI. A: Anomaly. P+A: Disease Prior restricted Anomaly. BAG: Brain Age Gap

Method	Full FT		No FT+LP		CN-only FT (own)+LP		CN-only FT (all)+LP
Method	BACC	AUC	BACC	AUC	BACC	AUC	BACC	AUC
ADNI \| AD/CN
3D-ViT	$78.06_{\pm 3.56}$	$83.63_{\pm 2.61}$	–	–	–	–	–	–
BrainHarmony	$66.51_{\pm 7.98}$	$78.78_{\pm 5.70}$	–	–	–	–	–	–
BrainMVP	$71.94_{\pm 2.15}$	$83.11_{\pm 4.65}$	–	–	–	–	–	–
BrainIAC	$66.70_{\pm 2.56}$	$74.66_{\pm 3.70}$	–	–	–	–	–	–

Appendix

Table A12: Full-train linear probing performance (in %) across datasets and binary tasks for AD and MCI classes. Prior work rows are reported with full finetuning (end-to-end). BrainNorm methods are shown under No FT, FT (own), and FT (all). Values are reported as

mean_{\pm std}

. Within each task block, the best value is in bold and second-best is underlined .

Method	Full FT		No FT+LP		CN-only FT (own)+LP		CN-only FT (all)+LP
Method	BACC	AUC	BACC	AUC	BACC	AUC	BACC	AUC
NIFD \| FTD/CN
3D-ViT	$52.96_{\pm 1.98}$	$60.68_{\pm 7.12}$	–	–	–	–	–	–
BrainHarmony	$52.59_{\pm 3.56}$	$69.42_{\pm 3.36}$	–	–	–	–	–	–
BrainMVP	$56.60_{\pm 4.34}$	$82.26_{\pm 6.45}$	–	–	–	–	–	–
BrainIAC	$53.86_{\pm 5.68}$	$66.12_{\pm 5.12}$	–	–	–	–	–	–

Appendix

Table A13: Full-train linear probing performance (in %) across datasets and binary tasks for FTD and PD classes. Prior work rows are reported with full finetuning (end-to-end). BrainNorm methods are shown under No FT, FT (own), and FT (all). Values are reported as

mean_{\pm std}

. Within each task block, the best value is in bold and second-best is underlined .

Method	Full FT		No FT+LP		CN-only FT (own)+LP		CN-only FT (all)+LP
Method	BACC	AUC	BACC	AUC	BACC	AUC	BACC	AUC
ADNI+NIFD+PPMI \| AD/FTD/PD
3D-ViT	$58.86_{\pm 1.65}$	$79.72_{\pm 1.08}$	–	–	–	–	–	–
BrainHarmony	$74.49_{\pm 3.83}$	$90.57_{\pm 1.85}$	–	–	–	–	–	–
BrainMVP	$82.79_{\pm 2.63}$	$96.78_{\pm 0.92}$	–	–	–	–	–	–
BrainIAC	$55.65_{\pm 1.50}$	$79.17_{\pm 1.44}$	–	–	–	–	–	–

Appendix

Table A14: Full-train linear probing performance across datasets and multi-class tasks. Prior work rows are reported with full finetuning (end-to-end). BrainNorm methods are shown under No FT, FT (own), and FT (all). Values are reported as

mean_{\pm std}

. Within each task block, the best value is in bold and second-best is underlined .

Method	Mode	$\mathbf{k=1}$	$\mathbf{k=2}$	$\mathbf{k=4}$	$\mathbf{k=8}$	$\mathbf{k=16}$
ADNI \| [AD/CN]
BrainNorm_M1	No FT + LP	$55.31_{\pm 25.18}$	$57.69_{\pm 25.01}$	$62.54_{\pm 19.82}$	$75.21_{\pm 7.97}$	$80.60_{\pm 2.06}$
	CN-only FT(own) + LP	$68.86_{\pm 23.78}$	$67.60_{\pm 22.95}$	$59.59_{\pm 27.55}$	$80.23_{\pm 1.77}$	$80.84_{\pm 2.49}$
	CN-only FT (all) + LP	$69.22_{\pm 23.65}$	$69.47_{\pm 21.01}$	$59.60_{\pm 27.49}$	$80.33_{\pm 2.61}$	$81.09_{\pm 2.21}$
BrainNorm_M2	No FT + LP	$55.30_{\pm 13.75}$	$50.70_{\pm 11.70}$	$70.57_{\pm 9.48}$	$78.28_{\pm 6.23}$	$83.16_{\pm 4.02}$
	CN-only FT(own) + LP	$79.51_{\pm 10.37}$	$69.28_{\pm 21.02}$	$77.13_{\pm 12.10}$	$84.95_{\pm 1.88}$	$86.08_{\pm 3.34}$

Appendix

Table A15: Few-shot performance across binary tasks and datasets for BrainNorm variants. Each method is shown under three modes: No FT, CN-only FT on the corresponding dataset only, and CN-only FT on pooled CN data from ADNI+AIBL+MIRIAD+PPMI+NIFD. Values are reported as mean

\pm

standard deviation.

Method	Mode	$\mathbf{k=1}$	$\mathbf{k=2}$	$\mathbf{k=4}$	$\mathbf{k=8}$	$\mathbf{k=16}$
ADNI+NIFD+PPMI \| [AD/FTD/PD]
BrainNorm_M1	No FT + LP	$57.02_{\pm 9.59}$	$64.05_{\pm 9.45}$	$80.38_{\pm 3.50}$	$82.35_{\pm 2.93}$	$85.18_{\pm 3.57}$
	CN-only FT(own) + LP	$58.15_{\pm 9.92}$	$62.82_{\pm 9.94}$	$81.12_{\pm 3.45}$	$83.36_{\pm 3.50}$	$85.13_{\pm 2.73}$
	CN-only FT (all) + LP	$57.74_{\pm 9.61}$	$62.38_{\pm 9.24}$	$81.08_{\pm 3.52}$	$83.39_{\pm 3.95}$	$84.92_{\pm 3.55}$
BrainNorm_M2	No FT + LP	$58.97_{\pm 8.17}$	$66.91_{\pm 8.10}$	$78.65_{\pm 5.27}$	$82.12_{\pm 3.50}$	$84.34_{\pm 3.06}$
	CN-only FT(own) + LP	$58.67_{\pm 8.78}$	$67.13_{\pm 8.68}$	$79.83_{\pm 5.17}$	$81.17_{\pm 4.52}$	$83.08_{\pm 1.11}$

Appendix

Table A16: Few-shot performance across multi-class tasks for BrainNorm variants. Each method is shown under three modes: No FT, CN-only FT on the corresponding dataset only, and CN-only FT on pooled CN data from ADNI+AIBL+MIRIAD+PPMI+NIFD. Values are reported as mean

\pm

standard deviation.

Parcel ID	Atlas	Tissue_Class	Label ID	Region
0	AAL3	Gray_Matter	1	Left Precentral Gyrus
1	AAL3	Gray_Matter	2	Right Precentral Gyrus
2	AAL3	Gray_Matter	3	Left Superior Frontal Gyrus
3	AAL3	Gray_Matter	4	Right Superior Frontal Gyrus
4	AAL3	Gray_Matter	5	Left Middle Frontal Gyrus
5	AAL3	Gray_Matter	6	Right Middle Frontal Gyrus

Appendix

Table A17: Atlas parcel index mapping used in BrainNorm.

Medical Imaging Foundation Models

Medical imaging foundation models learn reusable representations for tasks across clinical images, modalities, and institutions.

See all 234 papers

Sep 4, 2026

A Generalizable Feature Extractor for Alzheimer's-Related Brain MRI Tasks

When there is not enough labeled data to properly train deep learning models, transfer learning can help. We still do not fully understand how effective it is in neuroimaging, especially for Alzheimer's disease research. It is also not clear if these transferred models can work on new datasets without being retrained for each specific task. We evaluate whether a compact, supervised pretrained model can serve as a reusable foundation model for downstream neuroimaging tasks. We freeze the 7.18 million weights of a 3D CNN previously trained for brain-age prediction, and adapt it to each task using Low-Rank Adaptation (LoRA), requiring only ~1% additional trainable parameters. We evaluate generalizability in six experiments. Adapting the model to classify cognitively normal versus Dementia on ADNI gave an AUC of 0.964 on held-out folds (Experiment #1). Applying that adapted model unchanged to OASIS-3, with no retraining, gave an AUC of 0.871 (Experiment #2). Reusing its output logit together with age and a cognitive score distinguished stable from progressing MCI with an AUC of 0.828 (Experiment #3). Adapting the same backbone to predict amyloid positivity from structural MRI gave an AUC of 0.804 (Experiment #4). Finally, the same approach estimated ICV-normalized hippocampal and white matter hypointensity volumes directly from the T1w image, with R^2 of 0.80 and 0.91 respectively, tasks normally addressed with much larger U-Net networks (Experiments #5 and #6). A compact model supervised on brain age can therefore serve as a reusable backbone, adapting to each task with ~1% additional parameters and transferring to an unseen cohort without any training. Our findings suggest that a carefully trained brain age model can serve as an effective foundation model for Alzheimer's related tasks, even under strict data constraints.

Reza Rajabli, D. Louis Collins

McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University · Department of Biomedical Engineering, McGill University

Medical Imaging Foundation Models Transfer Learning Cross-Domain Transfer Learning Alzheimer's Disease Classification Alzheimer's Disease Progression Prediction

Ask PDF

Oct 20, 2026

Brain2Vec: A Self-Supervised 3D Foundation Model for Structural Brain MRI

Deep learning models have become essential tools in medical imaging, yet training them end-to-end for diverse tasks remains challenging due to limited labeled data and computational constraints. To address this, we introduce Brain2Vec, a foundation model pre-trained on 74,425 T1-weighted Magnetic Resonance Imaging (MRI) scans pooled from multiple studies. Brain2Vec can serve both as a weight initialization backbone for fine-tuning downstream tasks and as a standalone feature extractor. We evaluated Brain2Vec in three tasks: downstream predictive performance, convergence speed, and embedding quality. Using Brain2Vec's pretrained weights as initialization and fine-tuning on labeled data for a brain age prediction task, results demonstrate improved final accuracy and faster convergence compared to random initialization. Additionally, its frozen embeddings outperform anatomically defined volumetric features on four of five downstream classification tasks and perform competitively with a state-of-the-art foundation model pre-trained on substantially more data. Lastly, we show that Brain2Vec encodes information primarily from a spatially localized and neuroanatomically meaningful set of brain regions, and that manipulating a subset of latent features produces monotonic, region-specific effects on the reconstructed anatomy. Brain2Vec code and trained models are available as an open-source package on https://github.com/CBICA/Brain2Vec. Users can apply Brain2Vec locally by installing the package, or directly through the NiChart cloud platform https://cloud.neuroimagingchart.com, where Brain2Vec features can be derived without requiring local installation or specialized infrastructure.

Spiros Maggioros, Guray Erus, Gareth Harman +4

University of Pennsylvania

Medical Imaging Foundation Models Medical Image Representation Learning Neuroimaging Self-Supervised Pre-Training

Ask PDF

Sep 6, 2026

Comparative Study of Anatomical and Learned Features in AI Models for Structural Brain MRI

In this work, we comprehensively evaluate three popular feature-extraction paradigms in AI-based neuroimaging modeling: (1) computation of anatomical surfaces and volumes, (2) supervised learning with convolutional neural networks (CNNs), and (3) unsupervised pretraining of vision transformer (ViT) foundation models, followed by supervised finetuning. Our study is based on 18 publicly available datasets containing 3D structural T1-weighted MRI scans from approximately 80,000 participants across seven distinct clinical tasks. We observe that a linear model based on anatomical features matches the diagnostic performance of complex nonlinear features learned by sophisticated AI frameworks, including foundation models trained on thousands of scans. Conversely, CNNs and pretrained ViTs learn features that implicitly capture relevant anatomical information, bypassing the need for explicit feature extraction. Building upon these insights, we propose Anatomy Segmentation Pretraining (ASP), a novel method to incorporate anatomical information during foundation-model pretraining, which outperforms existing models in biological age estimation.

Boyang Yu, Miquel Lopez Escoriza, Long Chen +3

Center for Data Science, New York University, NYU, 60 Fifth Ave, New York, 10011, New York, United States. · Department of Neurology, NYU Grossman School of Medicine, 550 1st Ave, New York, 10016, New York, United States. · Courant Institute of Mathematical Sciences, NYU, 251 Mercer St, New York, 10012, New York, United States.

Neuroimaging Medical Imaging Foundation Models Brain MRI

Ask PDF

Abstract

Medical Imaging Foundation Models

Explore similar work

A Generalizable Feature Extractor for Alzheimer's-Related Brain MRI Tasks

Brain2Vec: A Self-Supervised 3D Foundation Model for Structural Brain MRI

Comparative Study of Anatomical and Learned Features in AI Models for Structural Brain MRI