FoundBioNet

The 2021 WHO classification of central nervous system tumors makes isocitrate dehydrogenase (IDH) mutation status a defining feature of glioma diagnosis and prognosis, but establishing that status normally requires invasive tissue sampling. Biopsy carries surgical risk and samples a single region of a tumor that is often molecularly heterogeneous. MRI is acquired routinely in every glioma workup and encodes suggestive cues — IDH-mutant tumors tend to have well-defined margins and the characteristic T2-FLAIR mismatch sign — yet these signals are subtle enough that conventional deep-learning classifiers, trained on the small annotated cohorts available in neuro-oncology, generalize poorly across scanners and centers.

FoundBioNet, developed at Macquarie University with Tehran University of Medical Sciences and presented at MICCAI 2025, addresses this by treating IDH genotyping as a transfer-learning problem on top of a pretrained 3D imaging foundation model rather than as a task to be learned from scratch. It starts from the 62M-parameter Tiny variant of BrainSegFounder, a SWIN-UNETR backbone self-supervised on tens of thousands of brain MRI volumes and fine-tuned for tumor segmentation, and adds two task-specific modules that steer the representation toward the imaging phenotype of IDH mutation.

The result is a single classifier trained once and then applied, with fixed weights, to four independent multi-center test cohorts spanning different scanners, acquisition protocols, and patient populations — an evaluation design that separates genuine cross-cohort generalization from the per-dataset tuning common in radiogenomics.

#Key Features

• Tumor-aware feature encoding: The TAFE module uses the segmentation decoder's tumor logits to gate multi-scale encoder features, so classification is driven by voxels inside the lesion rather than by incidental anatomy or scanner signature.
• Cross-modality differential: The CMD module explicitly contrasts the T2 and FLAIR channels to amplify the T2-FLAIR mismatch sign, a radiological marker of IDH mutation that is specific but easily missed.
• Joint segmentation and classification: A dual-stream fusion module merges the two feature streams, and the network trains end-to-end with a combined Dice segmentation loss and IDH classification loss, using localization as an auxiliary supervisory signal.
• Fixed-checkpoint external validation: One trained model is evaluated without retraining on four held-out cohorts drawn from six public glioma datasets, rather than being refit per site.
• Voxel-level interpretability: Occlusion sensitivity maps computed with MONAI show that predictions concentrate on tumor tissue, giving radiologists a check on where the evidence comes from.

#Technical Details

The full cohort comprises 1,705 glioma patients (354 IDH-mutant, 1,351 IDH-wildtype) assembled from TCGA-LGG/GBM, Ivy GAP, RHUH-GBM, UPenn-GBM, UCSF-PDGM, and the Erasmus Glioma Database, each with T1, contrast-enhanced T1, T2, and FLAIR sequences registered to 1 mm isotropic space, bias-corrected, skull-stripped, z-scored, and cropped to 96 x 96 x 96 voxels. BrainSegFounder's stage-3 weights are first refined on UCSF-PDGM under 5-fold cross-validation, with cases overlapping BraTS 2021 excluded, then adapted for classification and trained with Adam (learning rate 1e-4, batch size 2, early stopping). On internal cross-validation the model reaches 93.31% AUC; on the external test sets it achieves 90.58% ± 1.25 AUC on EGD, 88.08% ± 3.08 on TCGA, 80.31% ± 1.09 on UPenn-GBM, and 65.41% ± 3.35 on the combined Ivy GAP and RHUH cohort, outperforming ResNet10, ResNet50, SENet101, DenseNet121, and self-supervised ViT baselines with most comparisons significant at p ≤ 0.05. Ablations on EGD show TAFE alone at 84.38% AUC and the fused TAFE+CMD configuration at 90.58%.

#Applications

The intended use is presurgical molecular characterization in neuro-oncology: flagging likely IDH-mutant gliomas from imaging already acquired for diagnosis, informing surgical and radiotherapy planning, and supporting cases where biopsy is risky, nondiagnostic, or unavailable. Because the pipeline consumes standard clinical sequences and returns both a tumor segmentation and a genotype probability with an interpretability map, it fits into existing neuroradiology workflows, and the same recipe offers researchers a template for probing other imaging biomarkers with a pretrained 3D backbone.

#Impact

FoundBioNet demonstrates that a self-supervised segmentation backbone can be repurposed into a competitive molecular classifier, and that task-specific inductive bias — tumor gating and a modality-difference channel — matters more than raw feature volume when labels are scarce. Its cross-cohort protocol also sets a useful bar for radiogenomics work, where single-center results have historically overstated generalization. Limitations are visible in that same protocol: accuracy drops sharply on the small, heavily imbalanced UPenn-GBM cohort (MCC 26.30%) and on Ivy GAP/RHUH, both dominated by IDH-wildtype glioblastoma, so the model is far more reliable on cohorts with balanced molecular classes. The repository releases the TAFE, CMD, and DSF module code under GPL-3.0 but no trained checkpoint, so reproducing the reported results requires retraining from the public datasets.