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

TL;DR

Curia-2 advances radiology foundation models by enhancing pre-training strategies, enabling billion-parameter Vision Transformers, and establishing comprehensive evaluation benchmarks for 2D and 3D medical imaging tasks.

Contribution

It introduces Curia-2, a significantly improved pre-training approach for radiology models, scaling to billion-parameter Vision Transformers and extending benchmarking frameworks.

Findings

Curia-2 outperforms previous models on vision tasks.

It scales models up to billion parameters for CT and MRI.

It provides new benchmarks for 2D and 3D radiological data.

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.