European researchers release 741-exam public breast MRI dataset for AI

European researchers from six institutions across five countries have released a public multi-center breast MRI dataset, EuBreastMRI, containing 741 exams that include malignant, benign, and non-lesion cases. Unlike previous public repositories dominated by single-center, uniform-protocol data, this dataset was collected from heterogeneous scanners and imaging protocols to reflect the real-world variability that AI models must handle in clinical practice. To establish a baseline for comparison, the team also released benchmark results using a transformer model, setting a standard for future work.

The landscape of public breast MRI data has long been shaped by two pillars: the Duke Breast Cancer MRI Dataset (about 1,000 cases from one institution) and various collections on The Cancer Imaging Archive (TCIA) such as ACRIN and RIDER. The Duke dataset is larger but suffers from homogeneity — all exams follow a single center’s protocol, which can lead to overfitting when models are tested on data from different scanners or patient populations. TCIA aggregates many datasets but lacks a unified curation process, making it difficult to use for apples-to-apples benchmarking. In contrast, the new EuBreastMRI dataset is purpose-built for AI research: it is carefully curated, includes ground-truth labels, and provides a transformer-based benchmark to standardize evaluation. This mirrors the trajectory of other medical imaging fields, such as the CheXpert dataset for chest X-rays, which pushed the community toward more realistic and generalizable models.

The implications of this release extend beyond academic curiosity. The global breast cancer imaging market, including MRI, is projected to reach approximately $5 billion by 2030, and AI is expected to play a critical role in improving diagnostic accuracy and workflow efficiency. However, the hidden risks here are substantial. First, 741 exams is still a relatively small sample for deep learning — many state-of-the-art models require tens of thousands of images to saturate performance. Second, the dataset is drawn exclusively from European centers, raising concerns about generalizability to other ethnicities and breast density distributions, which are known to affect MRI interpretation. Third, while the transformer model provides a strong baseline, it may not be optimal for all scanner manufacturers or clinical protocols, and the study does not address how incidental findings or non-lesion cases contribute to false positives. Finally, real-world clinical workflows involve integrating mammography and ultrasound, not just MRI, so models trained on this single modality alone are not yet ready for deployment.

For AI researchers and radiology startups, the EuBreastMRI dataset is a valuable step forward — it offers a more realistic training ground than its predecessors. But the small size and lack of ethnic diversity remind us that public datasets are only as good as their limitations. The dataset’s true value lies in enabling reproducible comparisons and fostering collaboration to collect even larger, more diverse, multi-modal cohorts in the future. The field should treat this as a foundation, not a final answer.