A foundation model for brain tumor MRI analysis: WHO grading and subtype classification.

OBJECTIVES

This study aims to develop a self-supervised foundational model based on routine MRI and to evaluate its performance in tasks of brain tumor grading and pathological subtype classification.

METHODS

We developed the Unified Multimodal Brain Imaging Foundation (UMBIF) model by performing self-supervised learning on 51,029 MRI images from multiple institutions. The model was first trained using a contrastive masked image modeling task to extract robust feature representations. Next, UMBIF was fine-tuned for downstream tasks-glioma grading and histological classification-using multi-center cohorts. Finally, we compared UMBIF model with mainstream convolutional neural networks and machine learning algorithms, evaluating accuracy, sensitivity, specificity, and area under the curve (AUC).

RESULTS

Compared to self-supervised pretraining methods applied to natural images or single large tumor region images, the UMBIF architecture effectively extracted more comprehensive feature representations, leading to superior model performance. The optimal classifier with pretrained weights demonstrated outstanding results on independent test datasets, achieving accuracies of 0.840 (AUC: 0.723) for grade II, 0.684 (AUC: 0.854) for grade III, 0.775 (AUC: 0.743) for grade IV gliomas and 0.903 (AUC: 0.966) for histological classification, respectively, highlighting its potential in clinical decision-making.

CONCLUSIONS

The UMBIF model demonstrated robust applicability across clinically relevant glioma-grading formulations and LGG/HGG subtype classification. By enhancing classification performance with pretrained weights and reducing reliance on annotated data, it holds strong clinical potential for improving diagnostic efficiency and decision-making.