Glioblastoma Subtype Prediction Eval

glioblastoma-subtype-prediction-eval

Multimodal Sheaf-based Network for Glioblastoma Molecular Subtype Prediction — Idrissova et al. (2025) (arXiv:2508.09717, 2025)

What this evaluates

Evaluates a model's ability to predict glioblastoma molecular subtypes using paired MRI and histopathology data. It probes the model's capacity to fuse heterogeneous imaging modalities, preserve topological structures, and handle missing data scenarios.

Datasets

• Ivy GAP + Cancer Stem Cells ISH Survey — total 30; splits: 3-fold cross-validation (-1)

Metrics

• accuracy (primary) — range: [0, 1]
  • Proportion of correctly classified patients out of the total number of patients. Calculated as (TP+TN) divided by total samples.
• sensitivity — range: [0, 1]
  • True positive rate; proportion of actual positive cases correctly identified by the model.
• specificity — range: [0, 1]
  • True negative rate; proportion of actual negative cases correctly identified by the model.
• macro-F1 — range: [0, 1]
  • Unweighted mean of F1 scores across all molecular subtypes, treating each class equally regardless of support.
• micro-F1 — range: [0, 1]
  • F1 score calculated globally by counting total true positives, false negatives, and false positives across all classes.

Input / output format

Input: Per-patient paired MRI (T1, T2, FLAIR, T1-weighted) and histopathology WSI data, preprocessed into modality-specific graphs where nodes represent 3D MRI volume blocks or 512x512 histopathology tiles, with edges defined by k-nearest neighbors and local/annotated labels. Node features are radiomic statistics for MRI and ResNet-extracted features for histopathology.

Output: Predicted glioblastoma molecular subtype label per patient.

Scoring recipe

def compute_metrics(y_true, y_pred):
    accuracy = (y_true == y_pred).mean()
    tp = ((y_true == 1) & (y_pred == 1)).sum()
    tn = ((y_true == 0) & (y_pred == 0)).sum()
    fp = ((y_true == 0) & (y_pred == 1)).sum()
    fn = ((y_true == 1) & (y_pred == 0)).sum()
    sensitivity = tp / (tp + fn)
    specificity = tn / (tn + fp)
    macro_f1 = f1_score(y_true, y_pred, average='macro')
    micro_f1 = f1_score(y_true, y_pred, average='micro')
    return accuracy, sensitivity, specificity, macro_f1, micro_f1

Common pitfalls

• 3-fold cross-validation is used instead of a fixed train/val/test split, making direct comparison with single-split benchmarks difficult.
• Missing modality scenario uses a different benchmark model (AAVGA) rather than the proposed method, limiting direct comparison for incomplete data.
• Dataset size is small (30 patients), which may lead to high variance in metric scores across folds.

Evidence (verbatim from paper)

The evaluation metrics included accuracy, sensitivity, specificity, macro- and micro-F1 scores. All benchmark models were trained and tested on the same dataset, with 3-fold cross validation.

Citation

@misc{idrissova2025multimodal,
  title={Multimodal Sheaf-based Network for Glioblastoma Molecular Subtype Prediction},
  author={Idrissova et al. (2025)},
  year={2025},
  note={arXiv:2508.09717}
}

• arXiv: 2508.09717