Crossmoda Eval

crossmoda-eval

crossMoDA Challenge: Evolution of Cross-Modality Domain Adaptation Techniques for Vestibular Schwannoma and Cochlea Segmentation from 2021 to 2023 — Wijethilake et al. (2025) (arXiv:2506.12006, 2025)

What this evaluates

Evaluates unsupervised cross-modality domain adaptation for medical image segmentation (Vestibular Schwannoma and Cochlea) and tumour grading (Koos classification) from ceT1 to T2 MRI.

Datasets

• crossMoDA — total ?; splits: test (-1)

Metrics

• DSC (primary) — range: [0, 1]
  • Dice Similarity Coefficient: 2|A∩B| / (|A|+|B|), measuring overlap between predicted and ground truth masks.
• ASSD — range: mm
  • Average Symmetric Surface Distance: mean distance between boundary points of predicted and ground truth masks.
• MA-MAE — range: other
  • Mean Absolute Error for Koos grade classification, measuring average deviation from true tumour grade.

Input / output format

Input: ceT1 and T2 MRI scans for segmentation; ceT1 MRI for Koos classification.

Output: Binary segmentation masks for Vestibular Schwannoma (and intra/extra-meatal subregions) and Cochlea; or Koos grade label.

Scoring recipe

def compute_dsc(pred, gt):
    intersection = np.sum(pred & gt)
    return 2.0 * intersection / (np.sum(pred) + np.sum(gt))

def compute_assd(pred, gt):
    pred_boundary = get_boundary(pred)
    gt_boundary = get_boundary(gt)
    dists = [min(np.linalg.norm(p - g) for g in gt_boundary) for p in pred_boundary] + \
            [min(np.linalg.norm(g - p) for p in pred_boundary) for g in gt_boundary]
    return np.mean(dists)

Common pitfalls

• Outliers in ASSD heavily skew aggregate-then-rank aggregation, making rank-then-aggregate more robust.
• Cochlea segmentation shows higher stability than VS due to anatomical uniformity, affecting cross-method comparisons.
• Challenge rankings use a rank-then-aggregate strategy with mean aggregation, not simple metric averaging across test cases.

Evidence (verbatim from paper)

Both teams achieved a median DSC above 86% for both cochlea and VS segmentation.

Citation

@misc{wijethilake2025crossmoda,
  title={crossMoDA Challenge: Evolution of Cross-Modality Domain Adaptation Techniques for Vestibular Schwannoma and Cochlea Segmentation from 2021 to 2023},
  author={Wijethilake et al. (2025)},
  year={2025},
  note={arXiv:2506.12006}
}

• arXiv: 2506.12006