Brisc Segmentation Eval

brisc-segmentation-eval

BRISC: Annotated Dataset for Brain Tumor Segmentation and Classification — Fateh et al. (2025) (arXiv:2506.14318, 2025)

What this evaluates

Evaluates deep learning models on multi-planar brain tumor segmentation from contrast-enhanced T1-weighted MRI scans. It probes multi-scale feature integration, cross-view generalization, and robustness to class imbalance across glioma, meningioma, and pituitary tumor types.

Datasets

• BRISC — total 6000; splits: full (6000)

Metrics

• mIoU (primary) — range: [0, 1]
  • Mean Intersection over Union averaged across three tumor classes (Glioma, Meningioma, Pituitary). IoU per class is computed as the intersection of predicted and ground truth masks divided by their union, with a small epsilon for numerical stability.
• Weighted mIoU — range: [0, 1]
  • Weighted average of per-class mIoU scores, where weights are proportional to the number of samples in each tumor class to account for dataset imbalance.

Input / output format

Input: Multi-planar (axial, sagittal, coronal) contrast-enhanced T1-weighted MRI brain scans.

Output: Binary segmentation masks indicating tumor regions for each scan.

Scoring recipe

def compute_iou(y_true, y_pred):
    y_pred_bin = (y_pred > 0.5).astype(int)
    inter = np.sum(y_true * y_pred_bin)
    union = np.sum(y_true) + np.sum(y_pred_bin) - inter + 1e-6
    return inter / union

def compute_mIoU(y_trues, y_preds, classes):
    return np.mean([compute_iou(y_trues[c], y_preds[c]) for c in classes])

def compute_weighted_mIoU(y_trues, y_preds, class_counts):
    ious = [compute_iou(y_trues[c], y_preds[c]) for c in class_counts]
    weights = np.array([class_counts[c] for c in class_counts])
    return np.average(ious, weights=weights / weights.sum())

Common pitfalls

• Predictions must be thresholded (e.g., >0.5) before computing IoU, as the formula explicitly uses binary predicted labels.
• The dataset is imbalanced across tumor types; using a simple arithmetic mean mIoU instead of the specified weighted mIoU will misrepresent overall performance.
• Multi-planar masks (axial, sagittal, coronal) are provided, but the paper does not specify whether evaluation is performed per-plane or aggregated across planes.

Evidence (verbatim from paper)

Intersection over Union (IoU), also known as the Jaccard Index, is a fundamental metric for evaluating binary segmentation tasks. It quantifies the overlap between the predicted tumor regions and the ground truth, normalized by their union [76]. ... Furthermore, we report a weighted mIoU , which is calculated based on the proportion of samples belonging to each tumor type, providing a more representative performance indicator across the dataset.

Citation

@misc{fateh2025brisc,
  title={BRISC: Annotated Dataset for Brain Tumor Segmentation and Classification},
  author={Fateh et al. (2025)},
  year={2025},
  note={arXiv:2506.14318}
}

• arXiv: 2506.14318