Curia-2: Scaling Self-Supervised Learning for Radiology Foundation Models

Antoine Saporta; Baptiste Callard; Corentin Dancette; Julien Khlaut; Charles Corbière; Leo Butsanets; Amaury Prat; Pierre Manceron

cs.CVApr 2, 2026

Curia-2: Scaling Self-Supervised Learning for Radiology Foundation Models

Authors: Antoine Saporta, Baptiste Callard, Corentin Dancette, Julien Khlaut, Charles Corbière, Leo Butsanets, Amaury Prat, Pierre Manceron

Organizations: Raidium, 27 rue du Faubourg Saint-Jacques, 75014 Paris, France. · Department of Vascular and Oncological Interventional Radiology, Hôpital Européen Georges Pompidou, AP-HP, Paris, France · Faculté de Santé, Université Paris-Cité, Paris, France.

Abstract

The rapid growth of medical imaging has fueled the development of Foundation Models (FMs) to reduce the growing, unsustainable workload on radiologists. While recent FMs have shown the power of large-scale pre-training to CT and MRI analysis, there remains significant room to optimize how these models learn from complex radiological volumes. Building upon the Curia framework, this work introduces Curia-2, which significantly improves the original pre-training strategy and representation quality to better capture the specificities of radiological data. The proposed methodology enables scaling the architecture up to billion-parameter Vision Transformers, marking a first for multi-modal CT and MRI FMs. Furthermore, we formalize the evaluation of these models by extending and restructuring CuriaBench into two distinct tracks: a 2D track tailored for slice-based vision models and a 3D track for volumetric benchmarking. Our results demonstrate that Curia-2 outperforms all FMs on vision-focused tasks and fairs competitively to vision-language models on clinically complex tasks such as finding detection. Weights will be made publicly available to foster further research.

Figures & tables

Panel of Figure 1 — Figure 1: (a) Scaling Laws. The training dynamic steadily improves after each modification (cf. Section 3.2 ). ✩: fine-tuning in 512 resolution. (b) Few-shot Learning (A1). Curia-2 g demonstrates quicker convergence and requires significantly fewer samples to achieve the same performance as other models.

Panel of Figure 2 — Figure 2: (Left) Content-Aware Cropping. Comparison of valid (green) and filtered out (red) crop regions. (Right) Anatomically-Guided Masking. Each pair shows the average of 2000 mask samplings of our anatomically-guided masking against the blockwise masking with uniform prior strategy (DINOv2 [ 17 ] ).

A1	${\dagger}$	2D/3D	CT Organ Recognition in CT scans from TotalSegmentator [ 23 ] dataset. Note that we used 104 organ labels compared to 54 in [ 8 ] .
A2	${\dagger}$	2D/3D	MRI Organ Recognition from TotalSegmentator MRI [ 10 ] dataset.
A3	${\dagger}$	2D/3D	Neuroimaging Age Estimation from T1-weighted MRI (IXI dataset [ 13 ] ).
O1	${\dagger}$	2D/3D	Kidney Lesion Malignancy classification in CT (KITS23 [ 4 ] dataset).
O2		2D/3D	Lung Nodule Malignancy classification in CT (LUNA16 dataset [ 20 ] ).
O3		2D	Tumor Anatomical Localization prediction in CT (DeepLesion [ 24 ] ).
O4		3D	Kidney Cancer Survival in contrast-enhanced CT from TCIA [ 5 ] cohort.

Table 1: Evaluation Benchmark. The tasks are split into 7 categories: Anatomy (A), Oncology (O), Musculoskeletal (M), Emergency (E), Infectious (I), Neurodegenerative (N), Multi-Finding Screening (F).

{\dagger}

: Development task.

Strategy	Average	A1 acc	A2 acc	A3 r2	O1 AUC	M1 AUC	E1 AUC	E2 AUC	E3 AUC	E4 AUC	I bacc
Curia B (512)	84.8	85.4	82.3	75.8	74.5	87.8	93.7	82.6	84.7	89.5	91.5
ViT-B (256)	46.9 $\downarrow$ 37.9	22.3	17.7	0.7	49.2	68.8	69.9	64.6	64.6	74.0	37.2
$+$ Cropping	81.5 $\uparrow$ 34.6	85.4	81.4	64.8	71.0	87.2	92.1	78.8	86.8	87.1	80.5
$-$ KoLeo	83.1 $\uparrow$ 1.6	84.6	81.4	74.8	70.3	87.2	93.3	82.7	82.4	87.3	87.1
$+$ SigReg	83.1 $=$	84.7	81.3	75.2	70.4	87.3	92.9	83.2	83.4	85.9	86.9
$+$ Masking	83.3 $\uparrow$ 0.2	84.2	78.2	75.3	67.9	87.1	93.1	84.1	88.7	87.5	87.2

Table 2: Ablation Study. Training collapses when lowering the resolution to 256. Our refinements stabilize convergence on radiological data and consistently improve performance while facilitating the expansion to larger ViTs.

Panel of Figure 3 — Figure 3: Dense Features. Cosine similarity maps between a query patch $\bm{+}$ and all patches in multiple images. Curia-2 g demonstrates strong semantic understanding of anatomical structures and cross-modality alignment capabilities, mapping structures between CT and MRI domains, even under rotations.

Model	Avg.	A1 acc	A2 acc	A3 r2	O1 AUC	O2 AUC	O3 bacc	M1 AUC	M2 AUC	M3 AUC	E1 AUC	E2 AUC	E3 AUC	E4 AUC	I bacc
MI2	84.4	75.8	63.2	72.4	68.1	92.9	88.9	85.6	86.3	93.0	90.1	93.2	94.0	88.5	90.0
BMCLIP	79.1	66.7	63.2	69.4	62.7	84.8	86.5	83.9	84.5	92.4	87.8	79.2	71.5	85.1	89.3
MGemma	80.1	67.6	63.4	57.8	63.8	91.2	89.9	83.6	84.7	91.0	87.7	85.1	77.0	88.7	90.4
Curia B	86.2	85.4	82.3	75.8	74.4	86.6	91.9	87.8	86.4	94.6	93.7	82.6	84.7	89.5	91.5
Curia L	87.9	88.7	89.1	75.6	80.0	84.8	91.7	87.4	86.3	93.9	93.7	87.1	89.1	89.6	93.4
Curia-2 B	86.8	87.2	85.2	75.4	75.3	89.6	93.5	87.3	86.9	94.5	92.6	82.7	89.0	88.8	87.3

Table 3: 2D Track Results. Curia-2 outperforms 2D FMs, including MedImageInsight (MI2), BioMedCLIP (BMCLIP), MedGemma (MGemma) and Curia.

Model	Avg .	A1 acc	A2 acc	A3 r2	O1 AUC	O2 AUC	O4 cindex	M4 AUC	E1 AUC	E2 AUC	E3 AUC	N bacc	F acc
CT-CLIP	51.5	38.1	20.9	22.0	63.8	56.8	58.1	43.5	69.3	42.6	69.7	69.9	63.4
Merlin	65.4	45.4	38.5	53.3	55.7	64.9	66.3	50.4	85.1	92.5	75.6	76.0	81.2
Pillar-0	71.5	50.7	40.0	53.5	70.2	84.9	68.1	69.2	92.4	98.2	56.9	91.7	82.4
MI2	83.6	86.3	83.0	80.4	79.8	89.8	59.3	71.2	91.5	94.4	94.6	94.4	79.0
BMCLIP	78.3	77.8	71.9	85.0	64.1	81.3	62.3	72.8	89.2	90.1	79.3	91.7	74.1
MGemma	81.8	76.2	79.5	77.7	77.9	88.7	75.3	70.1	87.9	94.0	89.7	91.2	72.7

Table 4: 3D Track results. Curia-2 scores the best results on this generalist benchmark against both 3D volumetric FMs and 2D slice-based FMs.

Explore similar work

Vision Foundation Models in Radiology: A Scoping Review of Data, Methodology, Evaluation and Clinical Translation
Jul 8, 2026Alejandro Vergara-Richart, Xavier Rafael-Palou, Almudena Fuster-Matanzo +3Medical Imaging Foundation Models Vision Foundation Models
Quantitative Imaging Biomarkers in Medicine, Quibim S.L., València, Spain · Universitat Politècnica de València, València, Spain · Quantitative Imaging Biomarkers in Medicine (Quibim S.L.), Valencia, Spain +1
GreenRFM: Learning a resource-efficient radiology vision-language foundation model via supervision-centric pre-training
Mar 6, 2026Yingtai Li, Shuai Ming, Qiuli Wang +13Medical Imaging Foundation Models Radiology VLMs
School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC), Hefei, Anhui, 230026, China · Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advanced Research, USTC, Suzhou, Jiangsu, 215123, China · Department of Radiology, The First Affiliated Hospital of USTC, Division of Life Sciences and Medicine, USTC, Hefei, Anhui, 230001, China +7
nnFoundation: 3D Foundation Models for Radiology
Sep 22, 2026Constantin Ulrich Harsy, Tassilo Wald, Karol Gotkowski +803D Foundation Models Medical Imaging Foundation Models

Jul 8, 2026J/K move · Enter open · S save

Vision Foundation Models in Radiology: A Scoping Review of Data, Methodology, Evaluation and Clinical Translation

Alejandro Vergara-Richart, Xavier Rafael-Palou, Almudena Fuster-Matanzo, Ignacio Iborra Roncales+2

Quantitative Imaging Biomarkers in Medicine, Quibim S.L., València, Spain · Universitat Politècnica de València, València, Spain · Quantitative Imaging Biomarkers in Medicine (Quibim S.L.), Valencia, Spain+1

Medical Imaging Foundation Models Vision Foundation Models Medical Image Benchmarks

Open paper AskPDF arXiv

Vision foundation models (VFMs) are increasingly being developed for radiological imaging, yet their definition, development and evaluation remain heterogeneous. We conducted a PRISMAScR scoping review of peer-reviewed studies published between January 2017 and March 2026 describing foundation models trained exclusively on radiological imaging data. Sixty-seven studies were included and mapped across three pillars: data scale and heterogeneity, architectural and pretraining scalability, and downstream transferability and generalization. Datasets primarily covered brain MRI, thoracoabdominal CT, and chest X-ray, ranging from fewer than 100,000 samples to multi-million-image cohorts. Transformer-based architectures and self-supervised pretraining predominated, particularly masked image modeling, contrastive learning and multi-stage approaches. Evaluation focused mainly on segmentation and classification, whereas cross-center, cross-scanner, anatomical and modality-shift validation was inconsistently reported. Alignment with FUTURE-AI principles was uneven. Overall, radiology-specific VFMs show promising transferability, but clinical translation remains constrained by limited data representativeness, heterogeneous benchmarks, incomplete reporting and insufficient deployment-oriented evaluation.

Abstract

Explore similar work

Vision Foundation Models in Radiology: A Scoping Review of Data, Methodology, Evaluation and Clinical Translation

GreenRFM: Learning a resource-efficient radiology vision-language foundation model via supervision-centric pre-training

nnFoundation: 3D Foundation Models for Radiology