NeurIPS 2026

REFORM-3DA Representation-Centric Evaluation Framework for 3D Medical Vision Foundation Models

¹ Department of Computer Science, Khalifa University, Abu Dhabi, UAE

² University of Western Australia

* Corresponding author: Naoufel Werghi

Abstract

Evaluation of 3D medical foundation models is typically dominated by downstream task scores, which entangle representation quality with adaptation choices. This is especially critical in volumetric imaging, where spacing and anisotropy can substantially alter image statistics while the anatomy itself remains unchanged.

We therefore introduce REFORM-3D, a Representation-centric Evaluation FramewORk for benchmarking frozen 3D medical vision foundation Models across three questions: Acquisition Robustness, asking whether features remain stable when spacing varies while anatomy is fixed; Cross-Modal Anatomical Alignment, asking whether labeled anatomical structures remain recoverable across modalities under balanced bidirectional retrieval; and Low-Shot Anatomical Transfer, asking whether organ identity remains recoverable from limited labeled support under case-disjoint evaluation.

These are also pretraining questions, because what a 3D encoder learns depends on how volumes are geometrically presented during self-supervised pretraining. We further introduce Med3DINO encoder family, pretrained on 62.6K unlabeled MR, CT, PET, CTA, and CBCT volumes under three explicit geometric assumptions: isotropic normalization, voxel-relative sampling, and spacing-aware resampling.

Across Med3DINO and heterogeneous external encoders, REFORM-3D reveals module-specific strengths rather than a single model ordering. Within the controlled Med3DINO comparison, voxel-relative sampling yields lower embedding drift, whereas spacing-aware resampling favors anatomical transfer; semantic-transfer advantages depend on the dataset and feature type.

The REFORM-3D framework

The encoder remains frozen across all three modules. Each protocol examines a different property, with metrics reported separately.

Acquisition robustness

Hold scan identity fixed and vary voxel spacing. Measure feature displacement, preservation of feature-space structure, and semantic transfer.

Cosine drift · CKA · Matched semantic transfer

Cross-modal alignment

Retrieve labeled anatomical structures between CT and MR in balanced, bidirectional retrieval pools.

mAP · Top-1 · Top-5

Low-shot anatomical transfer

Recover organ identity from limited labeled support while keeping support and query cases disjoint.

Centroid BA · Probe BA · Adaptation gain

Med3DINO: pretraining geometry and corpus

Med3DINO is pretrained under three geometric policies on a quality-curated multimodal corpus. Individual modality and sequence images count as separate pretraining volumes.

Retained 3D volumes: 62,638
Imaging modalities: 5
Source collections: 64
Mapped organ cores: 182

CT and MR form the core.

PET, CTA, and CBCT extend the modality spectrum. The anatomical mapping consolidates 238 fine-grained labels into 182 organ cores across nine coarse regions.

The 64 dataset (source) collections describe the source inventory before preprocessing and validation. Corpus exposure and taxonomy are distinct from benchmark eligibility and evaluation labels. These statistics describe Med3DINO pretraining, not a dataset redistributed by the public evaluation package.

CT34,052 volumes · 54.4%

MR23,245 volumes · 37.1%

PET3,368 volumes · 5.4%

CTA1,070 volumes · 1.7%

CBCT903 volumes · 1.4%

Obtain provider-controlled scans and annotations from their original sources under the applicable terms.

Pretraining geometry: ISO, REL and SA

ISO

Isotropic normalization

Resample to 1.5 × 1.5 × 1.5 mm, then use voxel-relative crops.

REL

Voxel-relative sampling

Retain native voxel spacing and define crops relative to voxel dimensions.

Spacing-aware sampling

Retain native spacing, combine relative in-plane and physical through-plane sampling, and gate cross-plane rotations by anisotropy.

The corpus, initialization, architecture, objective, non-geometric augmentations, and optimization budget are held fixed. Geometry, crop definition, and rotation handling vary together: this is a complete-policy comparison.

Selected paper findings

A representative encoder panel illustrates the three evaluation axes. Choose a dataset in each table and expand the alternative CLS results.

Spacing sensitivity

Does similarity preserve organ identity?

Uniformly high similarity can coexist with weak organ-specific separation. Compare the same-organ diagonal with different-organ off-diagonal entries.

AbdomenAtlas

AbdomenAtlas Mean ± SD · three evaluation seeds

Encoder	Feature	Drift ↓	CKA ↑	Transfer ↑
External encoders · contextual comparisons
VoCo	Multiscale GAP	0.80 ± 0.09	86.70 ± 0.10	74.60 ± 0.15
3DINO	GAP	3.60 ± 0.00	96.20 ± 0.11	77.80 ± 0.59
SAM-Med3D	Post-neck GAP	3.35 ± 0.04	86.88 ± 0.37	74.66 ± 0.32
Hi-End-MAE	CLS	0.94 ± 0.01	97.18 ± 0.17	72.68 ± 0.67
MiM	GAP	1.44 ± 0.01	99.59 ± 0.02	73.69 ± 0.49
GL-MAE	CLS	0.78 ± 0.00	99.29 ± 0.01	74.65 ± 0.94
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	4.10 ± 0.02	95.90 ± 0.11	72.50 ± 0.15
Med3DINO-REL	GAP	1.90 ± 0.03	96.70 ± 0.06	78.90 ± 0.78
Med3DINO-SA	GAP	5.40 ± 0.02	97.00 ± 0.02	79.70 ± 0.43

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	Drift ↓	CKA ↑	Transfer ↑
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	16.30 ± 0.01	87.50 ± 0.07	75.10 ± 0.18
Med3DINO-ISO	CLS	19.00 ± 0.09	78.50 ± 0.14	73.60 ± 0.78
Med3DINO-REL	CLS	10.90 ± 0.07	89.10 ± 0.13	77.20 ± 0.61
Med3DINO-SA	CLS	18.40 ± 0.06	85.30 ± 0.05	81.00 ± 0.16

KiTS23

KiTS23 Mean ± SD · three evaluation seeds

Encoder	Feature	Drift ↓	CKA ↑	Transfer ↑
External encoders · contextual comparisons
VoCo	Multiscale GAP	0.70 ± 0.00	81.30 ± 0.00	46.10 ± 4.07
3DINO	GAP	3.70 ± 0.00	91.90 ± 0.00	39.90 ± 2.23
SAM-Med3D	Post-neck GAP	2.83 ± 0.00	95.27 ± 0.00	52.82 ± 5.00
Hi-End-MAE	CLS	0.78 ± 0.00	97.44 ± 0.00	47.99 ± 3.40
MiM	GAP	0.75 ± 0.00	99.86 ± 0.00	48.94 ± 2.62
GL-MAE	CLS	0.50 ± 0.00	99.88 ± 0.00	48.51 ± 4.69
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	3.90 ± 0.00	91.50 ± 0.00	43.80 ± 2.41
Med3DINO-REL	GAP	4.40 ± 0.00	92.60 ± 0.00	41.60 ± 3.55
Med3DINO-SA	GAP	5.80 ± 0.00	94.30 ± 0.00	49.20 ± 3.32

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	Drift ↓	CKA ↑	Transfer ↑
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	15.40 ± 0.00	77.30 ± 0.00	45.50 ± 1.08
Med3DINO-ISO	CLS	20.00 ± 0.00	73.60 ± 0.00	34.80 ± 4.02
Med3DINO-REL	CLS	12.00 ± 0.00	86.10 ± 0.00	44.10 ± 2.22
Med3DINO-SA	CLS	19.20 ± 0.00	84.30 ± 0.00	57.80 ± 3.25

ImageCAS

ImageCAS Mean ± SD · three evaluation seeds

Encoder	Feature	Drift ↓	CKA ↑	Transfer ↑
External encoders · contextual comparisons
VoCo	Multiscale GAP	2.30 ± 0.00	42.60 ± 0.00	52.00 ± 0.13
3DINO	GAP	11.30 ± 0.00	88.30 ± 0.00	53.90 ± 0.35
SAM-Med3D	Post-neck GAP	0.87 ± 0.00	96.13 ± 0.00	54.23 ± 0.49
Hi-End-MAE	CLS	1.15 ± 0.00	93.61 ± 0.00	51.87 ± 0.14
MiM	GAP	2.06 ± 0.00	99.20 ± 0.00	54.01 ± 0.38
GL-MAE	CLS	1.06 ± 0.00	99.35 ± 0.00	52.79 ± 1.02
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	13.20 ± 0.00	92.40 ± 0.00	54.00 ± 0.23
Med3DINO-REL	GAP	12.30 ± 0.00	80.90 ± 0.00	53.40 ± 0.43
Med3DINO-SA	GAP	9.70 ± 0.00	85.80 ± 0.00	53.70 ± 0.24

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	Drift ↓	CKA ↑	Transfer ↑
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	32.10 ± 0.00	83.40 ± 0.00	55.90 ± 0.01
Med3DINO-ISO	CLS	22.00 ± 0.00	93.10 ± 0.00	52.90 ± 0.57
Med3DINO-REL	CLS	24.20 ± 0.00	83.30 ± 0.00	54.60 ± 0.45
Med3DINO-SA	CLS	35.60 ± 0.00	83.80 ± 0.00	54.20 ± 0.37

TotalSegmentator-MR

TotalSegmentator-MR Mean ± SD · three evaluation seeds

Encoder	Feature	Drift ↓	CKA ↑	Transfer ↑
External encoders · contextual comparisons
VoCo	Multiscale GAP	0.10 ± 0.00	95.94 ± 0.00	71.10 ± 0.82
3DINO	GAP	1.30 ± 0.00	98.90 ± 0.00	80.70 ± 0.23
SAM-Med3D	Post-neck GAP	0.51 ± 0.00	97.53 ± 0.00	72.61 ± 1.40
Hi-End-MAE	CLS	0.23 ± 0.00	99.18 ± 0.00	70.28 ± 1.85
MiM	GAP	0.01 ± 0.00	99.94 ± 0.00	56.35 ± 1.53
GL-MAE	CLS	0.04 ± 0.00	99.95 ± 0.00	58.70 ± 0.32
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	2.00 ± 0.00	97.90 ± 0.00	78.00 ± 0.87
Med3DINO-REL	GAP	1.80 ± 0.00	97.40 ± 0.00	81.70 ± 1.05
Med3DINO-SA	GAP	1.70 ± 0.00	99.20 ± 0.00	83.50 ± 0.57

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	Drift ↓	CKA ↑	Transfer ↑
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	6.00 ± 0.00	92.40 ± 0.00	74.60 ± 0.73
Med3DINO-ISO	CLS	7.00 ± 0.00	92.40 ± 0.00	59.30 ± 0.34
Med3DINO-REL	CLS	5.60 ± 0.00	93.50 ± 0.00	77.00 ± 0.91
Med3DINO-SA	CLS	5.70 ± 0.00	94.20 ± 0.00	80.60 ± 1.18

Selected results from the controlled same-scan acquisition table. All three metrics are scaled by 100; lower drift and higher CKA / transfer are preferred. Mean ± SD over three evaluation seeds for fixed checkpoints.

This is a geometric spacing stress test. It does not simulate the full variability of clinical scanner physics.

CT–MR retrieval

Does anatomy remain recoverable across modalities?

Retrieval strengths depend on the encoder and dataset. Within the controlled Med3DINO triplet, SA has the highest reported mean mAP on these three datasets with both GAP and CLS. The selected external rows provide additional context.

TotalSegmentator CT–MR

TotalSegmentator CT–MR Mean ± SD · three evaluation seeds

Encoder	Feature	mAP ↑	Top-1 ↑	Top-5 ↑
External encoders · contextual comparisons
VoCo	Multiscale GAP	11.40 ± 0.12	18.90 ± 1.79	49.90 ± 2.28
3DINO	GAP	20.10 ± 0.16	36.90 ± 1.59	64.50 ± 1.77
SAM-Med3D	Post-neck GAP	13.61 ± 0.09	21.01 ± 0.21	51.59 ± 0.59
Hi-End-MAE	CLS	11.03 ± 0.07	20.71 ± 0.70	49.96 ± 2.19
MiM	GAP	9.00 ± 0.05	9.56 ± 1.00	31.49 ± 2.44
GL-MAE	CLS	9.34 ± 0.04	9.69 ± 0.12	29.76 ± 0.97
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	16.20 ± 0.11	26.80 ± 0.50	54.60 ± 1.33
Med3DINO-REL	GAP	21.00 ± 0.17	38.50 ± 1.07	65.80 ± 0.59
Med3DINO-SA	GAP	27.20 ± 0.24	57.00 ± 0.68	78.10 ± 0.42

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	mAP ↑	Top-1 ↑	Top-5 ↑
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	17.60 ± 0.06	32.20 ± 0.51	60.60 ± 0.61
Med3DINO-ISO	CLS	14.60 ± 0.13	23.20 ± 1.02	54.00 ± 1.35
Med3DINO-REL	CLS	18.10 ± 0.06	33.80 ± 0.77	64.80 ± 0.24
Med3DINO-SA	CLS	19.90 ± 0.06	41.10 ± 0.20	72.20 ± 0.23

MMWHS

MMWHS Mean ± SD · three evaluation seeds

Encoder	Feature	mAP ↑	Top-1 ↑	Top-5 ↑
External encoders · contextual comparisons
VoCo	Multiscale GAP	19.30 ± 0.00	13.20 ± 0.00	25.40 ± 0.00
3DINO	GAP	30.20 ± 0.00	35.70 ± 0.00	76.40 ± 0.00
SAM-Med3D	Post-neck GAP	18.30 ± 0.00	15.36 ± 0.00	36.43 ± 0.00
Hi-End-MAE	CLS	19.10 ± 0.00	12.50 ± 0.00	42.86 ± 0.00
MiM	GAP	18.26 ± 0.00	15.36 ± 0.00	42.50 ± 0.00
GL-MAE	CLS	17.40 ± 0.00	13.93 ± 0.00	44.29 ± 0.00
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	25.90 ± 0.00	26.10 ± 0.00	56.10 ± 0.00
Med3DINO-REL	GAP	39.20 ± 0.00	43.20 ± 0.00	72.50 ± 0.00
Med3DINO-SA	GAP	41.90 ± 0.00	48.60 ± 0.00	83.20 ± 0.00

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	mAP ↑	Top-1 ↑	Top-5 ↑
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	29.40 ± 0.00	40.70 ± 0.00	77.10 ± 0.00
Med3DINO-ISO	CLS	28.10 ± 0.00	31.40 ± 0.00	60.00 ± 0.00
Med3DINO-REL	CLS	25.40 ± 0.00	32.10 ± 0.00	69.60 ± 0.00
Med3DINO-SA	CLS	31.10 ± 0.00	43.90 ± 0.00	82.90 ± 0.00

AMOS22

AMOS22 Mean ± SD · three evaluation seeds

Encoder	Feature	mAP ↑	Top-1 ↑	Top-5 ↑
External encoders · contextual comparisons
VoCo	Multiscale GAP	16.30 ± 0.11	25.70 ± 0.42	50.30 ± 0.70
3DINO	GAP	28.00 ± 0.11	44.60 ± 0.40	77.60 ± 0.46
SAM-Med3D	Post-neck GAP	15.32 ± 0.13	24.48 ± 0.53	45.75 ± 0.92
Hi-End-MAE	CLS	13.28 ± 0.05	15.78 ± 0.27	42.64 ± 1.02
MiM	GAP	9.58 ± 0.04	11.37 ± 0.46	28.32 ± 0.58
GL-MAE	CLS	9.76 ± 0.01	8.13 ± 0.27	25.61 ± 0.82
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	21.60 ± 0.04	27.80 ± 0.13	59.60 ± 0.42
Med3DINO-REL	GAP	31.80 ± 0.04	54.00 ± 0.70	78.00 ± 0.37
Med3DINO-SA	GAP	44.80 ± 0.09	74.80 ± 0.66	91.20 ± 0.05

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	mAP ↑	Top-1 ↑	Top-5 ↑
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	21.10 ± 0.09	27.40 ± 1.21	57.10 ± 0.59
Med3DINO-ISO	CLS	20.00 ± 0.12	32.70 ± 0.22	61.20 ± 0.53
Med3DINO-REL	CLS	23.50 ± 0.10	38.00 ± 0.66	69.60 ± 0.16
Med3DINO-SA	CLS	28.30 ± 0.14	56.70 ± 0.73	84.30 ± 0.24

Selected rows from the cross-modal alignment table. Scores are on a 0–100 scale; higher is better. Results average both balanced retrieval directions and report mean ± SD over three evaluation seeds for fixed checkpoints.

Crops retain surrounding context. These scores measure organ recoverability in that setting, rather than establishing context-independent anatomical alignment. External encoders differ in pretraining, architecture, and feature interface; geometric-policy attribution is restricted to ISO/REL/SA.

Low-shot transfer

What anatomy can a few labels recover?

Low-shot performance separates centroid organization from recoverability after fitting a readout. Within Med3DINO, SA has the highest reported mean probe BA on these datasets with both feature readouts; external encoders show different starting points and adaptation gains.

TotalSegmentator CT–MR

TotalSegmentator CT–MR Mean ± SD · three evaluation seeds

Encoder	Feature	Centroid BA ↑	Probe BA ↑	Adaptation gain
External encoders · contextual comparisons
VoCo	Multiscale GAP	25.80 ± 0.60	36.10 ± 0.97	10.20 ± 0.72
3DINO	GAP	47.00 ± 4.79	52.50 ± 3.20	5.50 ± 2.30
SAM-Med3D	Post-neck GAP	34.56 ± 3.99	36.54 ± 1.81	1.98 ± 2.18
Hi-End-MAE	CLS	31.80 ± 1.28	44.39 ± 1.76	12.60 ± 3.01
MiM	GAP	14.83 ± 2.92	19.15 ± 0.78	4.32 ± 2.94
GL-MAE	CLS	14.35 ± 1.09	18.34 ± 2.88	3.98 ± 2.55
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	37.50 ± 2.09	44.30 ± 4.17	6.80 ± 2.14
Med3DINO-REL	GAP	52.30 ± 1.81	54.80 ± 1.97	2.50 ± 1.33
Med3DINO-SA	GAP	55.30 ± 1.57	58.80 ± 1.75	3.50 ± 0.76

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	Centroid BA ↑	Probe BA ↑	Adaptation gain
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	47.20 ± 2.24	47.70 ± 2.19	0.50 ± 0.75
Med3DINO-ISO	CLS	43.10 ± 2.66	43.10 ± 1.80	0.00 ± 0.95
Med3DINO-REL	CLS	45.80 ± 3.04	46.10 ± 3.09	0.30 ± 0.10
Med3DINO-SA	CLS	48.20 ± 4.80	48.60 ± 3.39	0.40 ± 2.31

MMWHS

MMWHS Mean ± SD · three evaluation seeds

Encoder	Feature	Centroid BA ↑	Probe BA ↑	Adaptation gain
External encoders · contextual comparisons
VoCo	Multiscale GAP	29.20 ± 3.03	47.60 ± 7.34	18.40 ± 8.78
3DINO	GAP	65.90 ± 5.51	74.30 ± 4.24	8.50 ± 2.33
SAM-Med3D	Post-neck GAP	34.79 ± 1.88	55.29 ± 8.41	20.50 ± 9.22
Hi-End-MAE	CLS	44.05 ± 10.80	61.11 ± 17.66	17.06 ± 7.15
MiM	GAP	23.81 ± 3.83	46.16 ± 12.41	22.35 ± 8.73
GL-MAE	CLS	33.33 ± 14.59	42.33 ± 12.02	8.99 ± 4.42
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	54.20 ± 6.35	67.10 ± 4.96	12.80 ± 3.99
Med3DINO-REL	GAP	75.10 ± 1.00	79.00 ± 3.08	3.80 ± 2.58
Med3DINO-SA	GAP	74.50 ± 2.33	80.00 ± 0.83	5.60 ± 1.60

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	Centroid BA ↑	Probe BA ↑	Adaptation gain
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	57.40 ± 2.10	61.60 ± 0.40	4.20 ± 1.82
Med3DINO-ISO	CLS	43.10 ± 8.36	53.30 ± 7.20	10.20 ± 2.00
Med3DINO-REL	CLS	50.40 ± 6.97	58.10 ± 5.57	7.70 ± 1.39
Med3DINO-SA	CLS	57.30 ± 5.50	62.80 ± 4.80	5.50 ± 1.28

AMOS22

AMOS22 Mean ± SD · three evaluation seeds

Encoder	Feature	Centroid BA ↑	Probe BA ↑	Adaptation gain
External encoders · contextual comparisons
VoCo	Multiscale GAP	37.10 ± 1.01	52.70 ± 0.84	15.60 ± 1.78
3DINO	GAP	64.40 ± 1.96	70.10 ± 0.36	5.70 ± 1.65
SAM-Med3D	Post-neck GAP	35.68 ± 0.27	43.90 ± 3.92	8.22 ± 4.15
Hi-End-MAE	CLS	40.74 ± 2.84	57.98 ± 1.66	17.25 ± 4.43
MiM	GAP	15.67 ± 0.62	33.26 ± 5.22	17.59 ± 5.36
GL-MAE	CLS	16.57 ± 2.31	24.50 ± 3.82	7.93 ± 1.77
Med3DINO · controlled geometric policies
Med3DINO-ISO	GAP	55.90 ± 2.78	61.50 ± 5.44	5.60 ± 3.46
Med3DINO-REL	GAP	67.40 ± 2.14	71.10 ± 3.38	3.70 ± 1.66
Med3DINO-SA	GAP	71.70 ± 3.12	76.20 ± 2.39	4.50 ± 0.91

Compare CLS features for 3DINO and Med3DINO

Encoder	Feature	Centroid BA ↑	Probe BA ↑	Adaptation gain
Alternative CLS readout · 3DINO and Med3DINO
3DINO	CLS	57.60 ± 2.66	60.40 ± 1.64	2.80 ± 1.33
Med3DINO-ISO	CLS	51.70 ± 0.95	55.10 ± 1.17	3.40 ± 1.40
Med3DINO-REL	CLS	58.20 ± 3.65	63.00 ± 3.42	4.80 ± 0.25
Med3DINO-SA	CLS	60.80 ± 3.32	64.90 ± 2.81	4.10 ± 1.72

Selected results from the low-shot transfer table. Balanced accuracy is on a 0–100 scale; adaptation gain is the reported change from centroid to probe accuracy. Two labeled support samples per organ and modality; case-disjoint queries; mean ± SD over three evaluation seeds for fixed checkpoints. A gain is not a standalone ranking criterion.

Only the readout adapts; the encoder stays frozen. A larger gain can arise from a weaker centroid baseline.

Interpretation and scope

What the framework supports

Separate measurements of geometric sensitivity, CT–MR organ recoverability, and low-shot anatomical transfer can support an encoder shortlist for a specified task.

Where conclusions stop

The findings do not establish general invariance or guarantee downstream performance. Geometric-policy conclusions are established within the studied Med3DINO architecture and self-supervised recipe.

Use REFORM-3D

The standalone public release provides the three evaluation modules with KiTS23 and MMWHS examples.

Public evaluation workflow
Dataset	Module	Metrics
KiTS23	Controlled spacing sensitivity	Embedding drift, CKA, matched semantic transfer
MMWHS	CT–MR retrieval	mAP, Top-1, Top-5
MMWHS	Low-shot organ recognition	Centroid and probe balanced accuracy, adaptation gain

Install

Use Python 3.11 on Linux and a PyTorch build compatible with your CUDA installation.

Install · Linux / bash

git clone https://github.com/Endarzboy/REFORM3D.git
cd REFORM3D
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install -c requirements.txt .

Full installation instructions ↗

Configure and run

Obtain data and encoder weights separately, then edit their paths in the YAML. This example runs MMWHS.

Run · after editing YAML paths

reform3d check --config configs/mmwhs-demo.yaml
reform3d run --config configs/mmwhs-demo.yaml

Use configs/kits23-demo.yaml for controlled spacing evaluation. A completed run writes an HTML report, score CSV, and run metadata.

Citation

Use the following BibTeX entry to cite REFORM-3D (or Med3DINO).

Paper citation · BibTeX

Download .bib

@inproceedings{assefa_reform3d,
  title = {REFORM-3D: A Representation-Centric Evaluation Framework for 3D Medical Vision Foundation Models},
  author = {Assefa, Maregu and Naseer, Muzammal and Velayudhan, Divya and Kumie, Gedamu Alemu and Werghi, Naoufel},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2026},
  pages ={1--10}
}

Acknowledgments

This research was funded by the Center for Autonomous Robotic Systems (CARS), Khalifa University of Science and Technology (KU-CARS).