Speaker
Description
Accurate non-invasive prediction of isocitrate dehydrogenase (IDH) mutation status remains a major challenge in glioma imaging. Although magnetic resonance imaging (MRI) provides valuable structural and functional information, conventional radiological features lack consistent predictive power across datasets. Recent advances in foundation models trained on large-scale MRI datasets offer a new opportunity to extract transferable imaging representations that may generalize across modalities. In this study, we investigated whether latent embeddings derived from an MRI foundation model could be applied to both MRI and positron emission tomography (PET) data for IDH mutation classification.
A total of 487 subjects from the publicly available BraTS dataset were used to train and evaluate MRI-based classification models after age- and sex-matching (102 IDH-mutant and 102 IDH–wild-type cases). External validation was performed in an independent clinical cohort of 10 glioma patients who underwent hybrid dynamic ¹⁸F-fluoropivalate (FPIA) PET/MRI. Latent feature representations were extracted using BrainIAC, a pretrained MRI foundation model. Principal component analysis was applied to reduce feature dimensionality before classification using XGBoost. Two experimental settings were investigated: (1) MRI-only models trained on BraTS and externally validated on the clinical PET/MRI cohort, and (2) multimodal PET–MRI models integrating static ¹⁸F-FPIA PET embeddings with MRI embeddings. A conventional radiomics pipeline was implemented for comparison.
In the MRI-only analysis, the combination of T1, FLAIR, and arterial spin labeling (ASL) achieved the strongest external validation performance on the mpFPIA dataset, reaching an accuracy of 0.80 and an AUC of 0.90. When PET embeddings were incorporated, embedding-based models consistently outperformed radiomics across modalities. PET-derived embeddings alone achieved an accuracy of 0.75 and an AUC of 0.86, while multimodal PET–MRI embeddings reached an AUC of 0.78 and an accuracy of 0.68. Across all configurations, embedding-based models demonstrated higher accuracy, more balanced precision–recall performance, and substantially lower feature dimensionality compared with radiomics-based approaches. Saliency analysis further confirmed that the learned representations focused predominantly on tumor regions.
These results provide the first evidence that MRI foundation model embeddings can generalize to PET imaging, enabling unified multimodal pipelines for molecular glioma characterization. Cross-modal representation learning may therefore represent a promising strategy for developing robust imaging biomarkers for non-invasive tumor genotyping.
| Track | PSMR |
|---|---|
| Presentation type | Oral |