Pediatric Brain Tumor Seg Eval

pediatric-brain-tumor-seg-eval

A New Logic For Pediatric Brain Tumor Segmentation — Bengtsson et al. (2024) (arXiv:2411.01390, 2024)

What this evaluates

Evaluates deep learning architectures for multi-class segmentation of pediatric brain tumors on MRI scans. It probes the model's ability to accurately delineate tumor sub-regions (whole tumor, enhanced tumor, cystic component, edema) and assesses cross-domain generalizability to adult glioma data.

Datasets

• PED BraTS 2024 — total 26; splits: test (26)
• CBTN — total ?; splits: test (-1)
• BraTS Adult Glioma 2023 — total ?; splits: test (-1)

Metrics

• lesion-wise Dice (primary) — range: [0, 1]
  • Dice similarity coefficient calculated per lesion instance: 2|A∩B|/(|A|+|B|), where A and B are predicted and ground truth masks.
• lesion-wise HD95 — range: mm
  • 95th percentile of the Hausdorff distance between the surfaces of predicted and ground truth lesion masks.
• precision — range: [0, 1]
  • Ratio of true positive voxels to all predicted positive voxels.
• recall — range: [0, 1]
  • Ratio of true positive voxels to all ground truth positive voxels.

Input / output format

Input: 3D MRI volumes of brain tumors.

Output: Multi-class segmentation masks for tumor sub-regions (WT, ET, CC, ED, TC).

Scoring recipe

def compute_metrics(pred_mask, gt_mask):
    intersection = np.sum(pred_mask & gt_mask)
    dice = 2 * intersection / (np.sum(pred_mask) + np.sum(gt_mask))
    dists = hausdorff_distance(pred_mask, gt_mask)
    hd95 = np.percentile(dists, 95)
    precision = intersection / np.sum(pred_mask)
    recall = intersection / np.sum(gt_mask)
    return dice, hd95, precision, recall

Common pitfalls

• HD95 is highly sensitive to small surface outliers or noise in ground truth masks.
• Lesion-wise metrics require instance matching, which can be unstable for very small or fragmented tumors.
• Cross-dataset evaluation (pediatric to adult) reveals domain shift that standard in-distribution metrics may mask.

Evidence (verbatim from paper)

We evaluated four models’ performance on our 26-patient testing set from PED BraTS 2024. Using both the nnU-Net and SegMamba frameworks, we trained one baseline model each on all 4 labels and another model using the 3L/WT segmentation approach for each framework. For both frameworks we see improvements in lesion-wise Dice, lesion-wise HD95, precision and recall scores almost across the board, as seen in Table 1.

Citation

@misc{bengtsson2024newlogic,
  title={A New Logic For Pediatric Brain Tumor Segmentation},
  author={Bengtsson et al. (2024)},
  year={2024},
  note={arXiv:2411.01390}
}

• arXiv: 2411.01390