Mednext V2 Segmentation Eval

mednext-v2-segmentation-eval

MedNeXt-v2: Scaling 3D ConvNeXts for Large-Scale Supervised Representation Learning in Medical Image Segmentation — Saikat Roy et al. (2025) (arXiv:2512.17774, 2025)

What this evaluates

Evaluates 3D medical image segmentation backbones on diverse anatomical structures across CT and MR modalities. Probes the model's ability to learn robust spatial representations and generalize to fine-tuning on small and large-scale clinical datasets.

Datasets

• Pediatric CT-Seg — total ?; splits: 5-fold cross-validation (-1)
• Stanford Knee MR — total ?; splits: 5-fold cross-validation (-1)
• Toothfairy — total ?; splits: 5-fold cross-validation (-1)
• Stanford Brain Mets — total ?; splits: 5-fold cross-validation (-1)
• PANTHER Pancreatic Tumor — total ?; splits: 5-fold cross-validation (-1)
• CTSpine1k — total ?; splits: 5-fold cross-validation (-1)

Metrics

• Dice Similarity Coefficient (DSC) (primary) — range: [0, 1]
  • Measures voxel-wise overlap between prediction and ground truth: 2|A∩B|/(|A|+|B|).
• Normalized Surface Distance (NSD) at 1mm tolerance — range: [0, 1]
  • Computes the fraction of surface voxels within 1mm of the ground truth surface, normalized by the total surface area.

Input / output format

Input: 3D medical image volumes (CT or MR modalities) with isotropic spacing after z-score normalization and resampling.

Output: 3D segmentation masks predicting anatomical structures or pathologies.

Scoring recipe

def compute_metrics(pred, gt):
    intersection = np.sum(pred & gt)
    union = np.sum(pred) + np.sum(gt)
    dsc = 2 * intersection / union if union > 0 else 0.0
    pred_surf = get_surface(pred)
    gt_surf = get_surface(gt)
    dists = cdist(pred_surf, gt_surf)
    nsd = np.mean(np.any(dists <= 1.0, axis=1))
    return dsc, nsd

Common pitfalls

• Excludes Liver Tumor Seg samples from CTSpine1k to prevent data leakage with pretraining cohorts.
• Uses dataset-specific patch sizes (e.g., nnUNet's large patch for CADS) for fair baseline comparison.
• Initial backbone validation uses an 80-20 train/val split, while final evaluation uses 5-fold cross-validation.

Evidence (verbatim from paper)

We use 5-fold cross validation for our final evaluation providing Dice Similarity Coefficient (DSC) and Normalized Surface Distance (NSD) at 1mm tolerance on a diverse and challenging set of six public datasets across CT and MR modalities to perform our evaluation.

Citation

@misc{roy2025mednextv2,
  title={MedNeXt-v2: Scaling 3D ConvNeXts for Large-Scale Supervised Representation Learning in Medical Image Segmentation},
  author={Saikat Roy et al. (2025)},
  year={2025},
  note={arXiv:2512.17774}
}

• arXiv: 2512.17774