NeuroCBIR: A Fast and Accurate Image Retrieval System for Whole-Brain and Region-Specific MRI

Felix Nieto-del-Amor; Jingru Fu; J. -Sebastian Muehlboeck; Eric Westman; Daniel Ferreira; Rodrigo Moreno

cs.CVOct 5, 2026

NeuroCBIR: A Fast and Accurate Image Retrieval System for Whole-Brain and Region-Specific MRI

Authors: Felix Nieto-del-Amor, Jingru Fu, J. -Sebastian Muehlboeck, Eric Westman, Daniel Ferreira, Rodrigo Moreno

Abstract

Content-based image retrieval (CBIR) in neuroimaging enables the identification of structurally similar brain scans, supporting diagnosis, prognosis, and treatment planning; however, existing methods are often limited to small datasets, single brain regions, or coarse class labels, thereby restricting their clinical utility and generalizability. Here, we present NeuroCBIR, a framework for fast and flexible retrieval of both whole-brain and region-specific 3D T1w MRI scans. A total of 103 cortical and subcortical regions are extracted to enable both whole-brain and region-level queries. NeuroCBIR leverages latent representations learned by a variational autoencoder (VAE) combined with contrastive learning, producing scan-specific embeddings that capture anatomical patterns. These embeddings were evaluated for subject re-identification, zero-shot age prediction, and zero-shot multi-class pathology stratification. Re-identification performance was high across both whole-brain and brain-region levels (mean average precision across the top-5 retrieved images (mAP@5) >= 98.4%), with robust generalization across datasets and acquisition conditions. While NeuroCBIR is not trained for age prediction or pathology stratification, zero-shot evaluations for these two tasks demonstrate that the embeddings encode meaningful information for downstream tasks. Embedding extraction on a 4-core CPU required approximately 18.7 s per scan, whereas similarity search was effectively instantaneous (less than 0.01 s). NeuroCBIR is publicly available for brain MRI with more than 26,000 precomputed T1w MRI embeddings. It supports reproducible research, region-specific flexibility, and clinically meaningful personalized diagnostic support. The software is available at https://github.com/minnelab/NeuroCBIR.

Figures & tables

Reference	Dataset / Task	Method	Evaluation	Performance
Huang et al. (2012)	2D T1w MRI (glioma, meningioma, pituitary)	Bag-of-visual-words + metric learning	Compared tumor types Hit = retrieved same tumor class.	Prec@10 $\approx 94\%$ mAP $\approx 91\%$
Onga et al. (2019)	3D T1w MRI ADNI2 AD / EMCI / LMCI / SMCI / CN	CAE + metric learning	Clustering and classification across diagnostic groups	Compression = 4096:1 Acc $\approx 81.5\%$
Deepak and Ameer (2020)	2D T1 MRI Figshare (brain tumor dataset)	GoogLeNet CNN + Siamese network	Retrieval Hit = retrieved same tumor class	Compression = $77\,264$ :1 Prec@10 $=97.7\%$ mAP@10 $=97.6\%$
Nishimaki et al. (2022)	3D T1w MRI ADNI2 (AD vs. CN)	Localized $\beta$ -VAE	Classification across diagnostic groups	Compression = 4096:1 AUC $=0.718$
Puglisi et al. (2024)	2D T1w MRI ADNI	DeepBrainPrint	Retrieval Hit = scans from same subject	mAP@3 $=95.5\%$
Muraki et al. (2024)	3D T1w MRI ADNI (AD vs. CN)	Isometric VAE	Classification across diagnostic groups	Compression = 512:1 Acc $=91.5\%$

Table 1: Summary of recent literature on brain MRI retrieval and classification.

Dataset	$\mathbf{N_{I}}$	$\mathbf{N_{S}}$	$\mathbf{N_{R}}$	$\mathbf{N_{T}}$	Age	Class			Field strength		Manufacturer
Dataset	$\mathbf{N_{I}}$	$\mathbf{N_{S}}$	$\mathbf{N_{R}}$	$\mathbf{N_{T}}$		CN	MCI	AD	1.5T	3T	SH	GE	PH
ADNI	$20\,367$	$2389$	$2.0$	$4.1$	75 ± 7	$6260$	$11\,572$	$2524$	$7951$	$8779$	$7453$	$3563$	$1794$
OASIS3	$2642$	$1303$	$1.0$	$2.0$	71 ± 9	$2149$	$392$	$100$	$208$	$2429$	$2642$	$0$	$0$
AIBL	$1276$	$685$	$1.0$	$1.8$	74 ± 7	$938$	$179$	$152$	$266$	$1010$	$1276$	$0$	$0$
MIRIAD	$706$	$69$	$1.3$	$7.6$	70 ± 7	$243$	$0$	$463$	$706$	$0$	$0$	$706$	$0$
SLIM	$1015$	$571$	$1.0$	$1.8$	21 ± 1	$1015$	$0$	$0$	$0$	$1015$	$1015$	$0$	$0$

Table 2: Summary of the MRI datasets used in this study.

Figure 1: Pairwise FAED between MRI datasets stratified by MRI field strength (1.5T and 3.0T). Lower values indicate greater similarity in feature distributions, while higher values reflect larger domain shifts. White cells correspond to uncomputed comparisons. The comparisons were done with the fully-preprocessed brain images.

Figure 2: Overview of the NeuroCBIR framework. (A) VAE training: the encoder ( $E_{\phi}$ ) maps input images to mean ( $\boldsymbol{\mu}_{\phi}$ ) and variance ( $\boldsymbol{\sigma}_{\phi}$ ), which define the latent variable ( $\boldsymbol{u}_{\phi}$ ). The decoder ( $D_{\theta}$ ) reconstructs the image, optimized using an $L_{1}$ loss, latent regularization ( $L_{\mathrm{KL}}$ ), perceptual loss, and an adversarial loss with critic ( $C_{\psi}$ ). (B) CL training: the latent representations, i.e. the means $\boldsymbol{\mu}_{\phi}$ from the pretrained VAE encoder, are passed through an encoder–projector module $(E^{\prime}_{\omega},P_{\omega})$ , yielding compact embeddings ( $\boldsymbol{z}_{\omega}$ ). Similar (intra-subject) embeddings are pulled together, while dissimilar (inter-subject) embeddings are pushed apart by the criteria specified for the Multi-Positive Ranking Contrastive Loss ( $L_{MPRCL}$ ). We denote by $\mathrm{I}_{i,j}$ the input image corresponding to subject $S$ and index $R$ . (C) Inference: the frozen Q2E module maps a query image to its embedding, which is compared against the precomputed database { $\mathbf{z}_{\omega,1,1}$ , $\mathbf{z}_{\omega,1,2}$ , …, $\mathbf{z}_{\omega,S,R-1}$ , $\mathbf{z}_{\omega,S,R}$ } to rank and retrieve the top- $k$ most similar images.

Figure 3: Example of input and target images for whole-brain and brain-region data. We denote by $\mathrm{I}_{S,R}$ the input image corresponding to subject $S$ and index $R$ (i.e., time point). $\mathrm{I}_{1,1}$ – $\mathrm{I}_{1,4}$ show four images for the subject 1, and $\mathrm{I}_{2,1}$ – $\mathrm{I}_{2,2}$ show two images for the subject 2. The brain-region images illustrate the hippocampus.

Figure 4: Illustration of the proposed Multi-Positive Ranking Contrastive Loss (MPRCL). For a given anchor query, hard positives are defined by shared identity labels, while soft and hard negatives are selected using percentile thresholds applied to the VAE latent similarity. Soft negatives correspond to samples whose VAE similarity exceeds the upper percentile

\pi_{\mathrm{soft}}

, whereas negatives fall below the lower percentile

\pi_{\mathrm{hn}}

. In the embedding space, the loss enforces margin-based ranking constraints that encourage hard positives to be closest to the anchor, soft negatives to occupy an intermediate similarity range, and negatives to be pushed further away. The margins are adaptively scaled by the anchor-specific similarity range, using scaling factors

\alpha_{\mathrm{id}}

for identity-based constraints and

\alpha_{\mathrm{soft}}

for soft-negative versus negative constraints.

$\mathbf{N_{f}}$	$\boldsymbol{\alpha}_{\mathbf{hn}}$	$\boldsymbol{\lambda}_{\mathbf{sn}}$	$\boldsymbol{\alpha}_{\mathbf{sn}}$	$\boldsymbol{\pi}_{\mathbf{hn}}$	$\boldsymbol{\lambda_{\mathrm{hn}}}$	mAP@5 $\uparrow$	$\boldsymbol{\rho}_{\mathbf{MS-SSIM}}$ $\downarrow$
$16$	$0.5$	$1.0$	$0.05$	$30$	$0.10$	$94.7$	$-0.72$
$16$	$0.2$	$1.0$	$0.05$	$30$	$0.10$	$95.3$	$-0.71$
$32$	$0.5$	$1.0$	$0.05$	$30$	$0.10$	$95.2$	$-0.70$
$32$	$0.2$	$1.0$	$0.05$	$30$	$0.10$	$96.9$	$-0.74$
$32$	$0.2$	$0.2$	$0.05$	$30$	$0.10$	$97.0$	$\mathbf{-0.74}$
$32$	$0.2$	$0.2$	$0.10$	$30$	$0.10$	$97.6$	$-0.73$

Table 3: Hyperparameter optimization results for the whole-brain approach. For the validation partition: retrieval quality is assessed using mAP@5, while embedding quality is evaluated via Spearman’s rank correlation between embedding similarities and MS-SSIM (

\rho_{\mathrm{MS-SSIM}}

). The last three rows correspond to ablations of the selected model (

\star

Metric	All	Train	Val	Test	ADNI	OASIS3	AIBL	MIRIAD	SLIM 1
mAP@5 ( $\uparrow$ )	$99.2$	$99.6$	$98.7$	$98.4$	$99.3$	$92.5$	$100.0$	$99.9$	-
mS@1 ( $\uparrow$ )	$98.8$	$99.4$	$99.0$	$97.2$	$99.9$	$93.6$	$94.9$	$100.0$	$89.1$
mS@5 ( $\uparrow$ )	$99.9$	$100.0$	$99.8$	$99.8$	$99.9$	$99.4$	$100.0$	$100.0$	-
$\rho_{\mathrm{MS-SSIM}}$ ( $\downarrow$ )	$-0.71$	$-0.64$	$-0.71$	$-0.74$	$-0.71$	$-0.65$	$-0.75$	$-0.77$	$-0.79$

Table 4: Retrieval re-identification performance (mAP@5, mS@1, mS@5) and Spearman’s rank correlation between embedding similarities and MS-SSIM (

\rho_{\mathrm{MS-SSIM}}

) for the whole-brain NeuroCBIR framework across multiple datasets.

Figure 5: Retrieval examples of T1w brain MRI scans. A.1 belongs to a subject with 7 MRI scans. B.1 are from a subject with only 3 scans. C.1 belongs to a patient with only two scans. The blue frame denotes the queries, the green frames are for the "hits" (same subject scans), and the red frames are for the different subject scans that are deemed to be similar to the query by NeuroCBIR. Except for A.12, images from the same subject are ranked higher by NeuroCBIR.

Figure 6: mAP@5 and mS@5 for the testing partition of the top 20 best / worst brain structures sorted by mAP@5. Abbreviations: CC = Corpus Callosum; Mid = Middle; lh/rh = left/right hemisphere; ctx = cortex.

Metric	Structure	All	Train	Val	Test	ADNI	OASIS3	AIBL	MIRIAD	SLIM 1
mAP@5 ( $\uparrow$ )	Left-Hippocampus	$98.9$	$99.3$	$98.1$	$98.3$	$99.0$	$92.7$	$100.0$	$100.0$	-
	Left-Thalamus	$96.8$	$98.5$	$93.2$	$93.6$	$97.0$	$81.5$	$100.0$	$99.5$	-
	Left-Amygdala	$86.0$	$91.4$	$73.8$	$75.5$	$85.8$	$72.2$	$96.7$	$96.3$	-
	Left-Lateral-Ventricle	$99.6$	$99.7$	$99.4$	$99.3$	$99.6$	$96.0$	$100.0$	$100.0$	-
mS@1 ( $\uparrow$ )	Left-Hippocampus	$98.8$	$99.1$	$99.1$	$97.7$	$99.7$	$92.5$	$97.5$	$100.0$	$91.6$
	Left-Thalamus	$95.9$	$97.5$	$96.7$	$91.5$	$98.7$	$80.6$	$88.0$	$100.0$	$69.3$

Table 5: Retrieval re-identification performance (mAP@5, mS@1, mS@5) and Spearman’s rank correlation between embedding similarities and MS-SSIM (

\rho_{\mathrm{MS-SSIM}}

) for different brain-regions of the NeuroCBIR framework across multiple datasets. The full detailed results are provided in A .

Figure 7: Retrieval of similar left hippocampus segmentations. A.1 belongs to a subject with at least 6 MRI scans. B.1 are from a subject with only 5 scans (B1.1–B1.5). C.1 belongs to a patient with no additional MRI scans. The blue frame denotes the queries, the green frames are for the "hits", and the red frames are for the "fails".

Figure 8: Coefficient of Variation (CV) of Spearman correlations between whole-brain similarity and region-based similarity across top-

k

retrieved cases, excluding scans from the query subject. Left panel shows variability across regions for each query, and right panel shows variability across queries for each region. Top-

k

groups of 50, 200, and 1000 are shown. The CV is plotted on a logarithmic scale to highlight differences across a wide range of values. Points indicate individual observations, while horizontal markers show the mean CV for each top-

k

group. Higher CV indicates greater heterogeneity in the contribution of regions or queries to whole-brain similarity.

Comparison	p -value	ssd	Cliff’s $\boldsymbol{\delta}$	Magnitude
Test vs. Train	$0.000\,000$		$-0.027\,000$	negligible
CN vs. MCI	$0.003\,080$		$-0.006\,000$	negligible
CN vs. AD	$0.685\,000$		$-0.004\,000$	negligible
MCI vs. AD	$1.000\,000$		$0.002\,000$	negligible
ADNI vs. AIBL	$1.000\,000$		$-0.009\,000$	negligible
ADNI vs. MIRIAD	$0.338\,000$		$-0.006\,000$	negligible

Table 6: Statistical comparison of retrieval performance across datasets, scanner field strengths, and vendors. Reported values include p -values (Mann–Whitney U test), significance indicators (ssd), Cliff’s

\delta

effect sizes, and their qualitative magnitudes. Asterisks (*) denote significant differences at p -value<0.05.

Method	Comp.	$\mathbf{N_{f}}$ $\downarrow$	mAP@5 $\uparrow$	mS@5 $\uparrow$	$\boldsymbol{\rho}_{\mathbf{MS-SSIM}}$ $\downarrow$
ResNet-10	$2048$ :1	512	$7.2$	$32.6$	$-0.55$
ResNet-50	$512$ :1	2048	$10.8$	$43.9$	$-0.70$
ResNet-101	$512$ :1	2048	$5.7$	$30.0$	$-0.51$
CNN-AD	$2628$ :1	256	$0.1$	$0.7$	$-0.01$
VoComni-L	$576$ :1	1536	$14.3$	$50.9$	$-0.34$
BrainIAC	$1152$ :1	768	$72.4$	$95.8$	$-0.27$

Table 7: Comparison of different feature extractors and dimensionality reduction approaches on the retrieval task using the top-5 results from the whole-brain testing set. Reported metrics include the compression rate (Comp.), number of features (

\mathrm{N_{f}}

), mean precision, and success rate.

Step	RAM (MB) $\boldsymbol{(\downarrow)}$		Time (s) $\boldsymbol{(\downarrow)}$
Step	baseline	peak	1 core	4 cores
Whole-brain
Load Q2E	$862$	$1087$	$8.7$	$6.5$
Run Q2E	$1087$	$6804$	$66.4$	$18.7$
Top 5 retrieval	-	-	<0.01	<0.01
Brain-regions

Table 8: Execution time and memory usage for whole-brain and brain-region processing. Baseline refers to the RAM usage before running, peak refers to the peak of RAM measured during the execution.

Region	MAE $\boldsymbol{(\downarrow)}$	$\boldsymbol{r~(\uparrow)}$	Slope
Whole-brain
Random baseline	5.2 ± 3.5	0.00	0.00
ResNet-10	4.5 ± 3.5	0.46	0.32
ResNet-50	4.8 ± 3.6	0.35	0.16
ResNet-101	5.0 ± 3.6	0.27	0.13
CNN-AD	5.8 ± 4.0	0.00	0.00

Table 9: Zero-shot similarity-based brain age estimation across selected brain regions. For each region, the predicted age was computed using the similarity-based weighted average of the top-

500

retrieved embeddings. Median absolute error (MAE) in years is reported as mean ± standard deviation. Pearson correlation

r

and regression slope indicate the correspondence between predicted and chronological age.

Figure 9: Zero-shot similarity-based brain age estimation using NeuroCBIR for the left hippocampus. Predicted age is obtained by aggregating the ages of the top- $500$ most similar CN subjects. Panel (A) shows predicted versus chronological age for CN queries, panel (B) for AD queries, and panel (C) compares the age-gap distributions for CN and AD.

Method	CN/MCI/AD		CN/AD
Method	BAcc $\uparrow$	MF1 $\uparrow$	BAcc $\uparrow$	MF1 $\uparrow$
Whole-brain
Random baseline	33.3	33.3	50.0	50.0
ResNet-10	49.8	48.1	69.5	59.9
ResNet-50	50.2	48.0	67.8	60.2
ResNet-101	47.4	44.7	65.0	56.4

Table 10: Zero-shot similarity-based disease risk estimation across selected brain regions. For each query, the predicted class was obtained from the similarity-weighted labels of the top-

500

retrieved embeddings. Performance is reported for both the three-class (CN/MCI/AD) and binary (CN/AD) classification tasks using balanced accuracy (BAcc) and macro F1-score (MF1).

Figure 10: Row-normalized confusion matrices for zero-shot similarity-based disease risk estimation across brain regions (Top-k = 300). Rows are normalized by true class, showing per-class recall (chance-level performance corresponds to 0.33.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Structure	mAP@5	mS@1	mS@5	Structure	mAP@5	mS@1	mS@5
Right-Cerebral-White-Matter	99.6	99.4	100.0	Left-Cerebellum-White-Matter	98.3	96.8	99.9
Left-Cerebral-White-Matter	99.5	99.4	99.9	ctx-lh-lateralorbitofrontal	98.2	98.3	99.6
Right-Cerebellum-Cortex	99.5	98.6	99.9	ctx-lh-pericalcarine	98.2	97.8	99.4
ctx-lh-superiortemporal	99.5	99.3	99.9	ctx-rh-parstriangularis	98.1	98.0	99.6
ctx-rh-fusiform	99.5	99.0	99.9	ctx-lh-bankssts	98.1	99.0	99.6
ctx-lh-fusiform	99.5	98.8	99.9	ctx-rh-bankssts	98.0	99.2	99.7

Appendix

Table 11: mAP@5, mS@1 and mS@5 for different brain regions of the NeuroCBIR framework across the testing dataset.

Explore similar work

Jul 12, 2026cs.CV

Contrastive Joint-Embedding Prediction for Representation Learning in Structural MRI

Self-supervised learning offers a compelling approach for medical imaging, where labeled data are scarce and acquisition costs are high. We present COJEPA, a self-supervised framework for volumetric brain MRI that combines a joint-embedding predictive architecture (JEPA) with a contrastive loss (CO), targeting two complementary properties: local predictivity and global discriminability. The model is trained without labels on T1-weighted structural MRI from two cohorts (HCP-YA and AABC,

N{=}2286

, ages 22 to 90), extending I-JEPA to 3D with foreground-aware block masking, a hierarchical convolutional patch embedding, and world-space sinusoidal positional encodings. We evaluate all three objectives across zero-shot twin retrieval, brain tumor segmentation (BraTS 2024), and age regression (OpenBHB). COJEPA achieves the best monozygotic twin recall at rank@1 (0.84), the best finetuning age MAE (2.55 years on OpenBHB 3.0T), and matches CO on BraTS whole-tumor Dice, demonstrating that the combined objective yields representations that are simultaneously discriminative and locally structured.

Fabian Mager, Lars Kai Hansen

Department of Applied Mathematics and Computer Science Technical University of Denmark

Self-Supervised Learning Joint-Embedding Predictive Architectures Contrastive Learning Representation Learning

Ask PDF

May 19, 2026cs.CV

NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding

We present NeuroQA, a large-scale benchmark for visual question answering in 3D brain magnetic resonance imaging (MRI), with 56,953 QA pairs from 12,977 subjects across 12 datasets. It spans ages 5-104 and five clinical domains: Alzheimer's, Parkinson's, tumors, white matter disease, and neurodevelopment. Unlike prior medical Visual Question Answering (VQA) efforts that operate on 2D slices or rely on narrow diagnostic labels, NeuroQA pairs every item with a full 3D volume. It evaluates 11 clinically grounded reasoning skills across Yes/No, multiple-choice, and open-ended formats. Of the 203 templates, 131 are image-grounded (answerable from a 3-plane viewer) and 72 are image-informed (ground truth from quantitative volumetry or clinical instruments). To remove text-only shortcuts, we apply answer-distribution refinement, reducing closed-format text-only accuracy from

>

80% to 44.6%; image necessity is assessed separately through an image-grounding protocol released with the benchmark. A 38-rule deterministic pipeline and two rounds of expert review verify every QA pair against FreeSurfer measurements, metadata, or radiology report fields, with zero same-subject contradictions across templates. We conduct a clinician evaluation in which two clinicians independently assess 100 frozen test items on a three-plane viewer. On closed-format (Yes/No + multiple-choice) test-public items, the best zero-shot vision-language model and a supervised 3D CNN baseline reach 47.5% and 43.7% accuracy respectively, both below the 49.4% text-only majority-template floor. NeuroQA adopts a two-tier release with public QA pairs for open-access datasets and reproducible generation scripts for datasets restricted by data use agreements (DUAs), plus subject-level splits, a held-out private test set, and an online leaderboard.

Mohammad H. Abbasi, Favour Nerrise, Shaurnav Ghosh +12

Stanford University · 1Stanford University, Stanford, CA, USA

Medical Visual Question Answering Magnetic Resonance Imaging Question-Answer Pairs Neuroflex

Ask PDF

Sep 21, 2026cs.CV

Evaluating the Generalization of Neuroimaging Foundation Models on African Brain MRI

Neuroimaging foundation models pretrained on large, predominantly western cohorts are increasingly proposed as general-purpose backbones for brain MRI analysis. Yet, their ability to generalize to underrepresented clinical populations remains largely untested. We evaluate four recent foundation models (BrainIAC, Neuro-JEPA, NeuroVFM, and Primus) on a three-way diagnostic classification task (Control, Dementia, Parkinson's disease) using a cohort of 88 subjects from a Nigerian clinical brain MRI dataset, across four modality configurations (T1w, T2w, T1w+T2w, FLAIR), and compare against an end-to-end trained ViT3D baseline. The frozen backbones collapse to majority-class predictions, while Neuro-JEPA on FLAIR shows modest but still limited discrimination. In contrast, the end-to-end trained ViT3D achieves higher accuracy and MCC on every task (up to 53.4% accuracy, MCC=0.27) and is the only model with non-trivial recall. Our findings suggest that these frozen neuroimaging foundation models are insufficient for fine-grained diagnostic classification in small, non-western clinical cohorts, motivating parameter-efficient adaptation and broader multi-site external validation for equitable deployment in global health settings.

Oluwatobi Iyanuoluwa Akinmuleya, Olatokun Shamsudeen Akano, Samuel Danquah Ankapong +2

Department of Medicine, Shenyang Medical College, Shenyang, China · College of Medicine, University of Ibadan, Ibadan, Nigeria · University of Ghana, Accra, Ghana +2

Multimodal Neuroimaging Magnetic Resonance Imaging Electroencephalography Foundation Models Cohort

Ask PDF