Medclipseg Eval

medclipseg-eval

MedCLIPSeg: Probabilistic Vision-Language Adaptation for Data-Efficient and Generalizable Medical Image Segmentation — Koleilat et al. (2026) (arXiv:2602.20423, 2026)

What this evaluates

Evaluates data-efficient and domain-generalizable medical image segmentation using vision-language adaptation. It probes a model's ability to segment anatomical structures and lesions across diverse imaging modalities with limited supervision, while maintaining robustness to out-of-distribution domain shifts and providing calibrated uncertainty estimates.

Datasets

• BUSI — total ?; splits: train (-1), test (-1)
• BTMRI — total ?; splits: train (-1), test (-1)
• ISIC — total ?; splits: train (-1), test (-1)
• Kvasir-SEG — total ?; splits: train (-1), test (-1)
• QaTa-COV19 — total ?; splits: train (-1), test (-1)
• EUS — total ?; splits: train (-1), test (-1)
• BUSUC — total ?; splits: train (-1), test (-1)
• BUSBRA — total ?; splits: train (-1), test (-1)
• BUID — total ?; splits: train (-1), test (-1)
• UDIAT — total ?; splits: train (-1), test (-1)
• CVC-ColonDB — total ?; splits: train (-1), test (-1)
• CVC-ClinicDB — total ?; splits: train (-1), test (-1)
• CVC-300 — total ?; splits: train (-1), test (-1)
• BKAI — total ?; splits: train (-1), test (-1)
• BRISC — total ?; splits: train (-1), test (-1)
• UWaterlooSkinCancer — total ?; splits: train (-1), test (-1)

Metrics

• DSC (primary) — range: percent
  • Dice Similarity Coefficient: 2 * |A ∩ B| / (|A| + |B|), measuring voxel-wise overlap between prediction and ground truth.
• NSD — range: percent
  • Normalized Surface Distance: Average symmetric surface distance between predicted and ground truth contours, normalized by image dimensions.
• HM DSC — range: percent
  • Harmonic Mean of In-Distribution (ID) and Out-of-Distribution (OOD) DSC scores, used to balance performance across domain shifts.
• Brier score — range: [0, 1]
  • Mean squared difference between predicted uncertainty probabilities and actual binary outcomes, measuring calibration quality.
• Spearman correlation — range: [-1, 1]
  • Rank-based correlation coefficient between predicted uncertainty maps and actual segmentation errors.

Input / output format

Input: Medical images paired with textual prompts describing the target anatomy or pathology. If prompts are missing, they are synthetically generated using GPT-5.

Output: Pixel-level segmentation masks and continuous per-pixel uncertainty scores.

Scoring recipe

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

def compute_nsd(pred_mask, gt_mask):
    # Compute symmetric surface distance between contours
    return np.mean(symmetric_surface_distance(pred_mask, gt_mask))

def compute_hm_dsc(id_dsc, ood_dsc):
    return 2 * (id_dsc * ood_dsc) / (id_dsc + ood_dsc)

Common pitfalls

• Domain generalization evaluation strictly forbids fine-tuning on target datasets; models are trained fully supervised on a single source and tested directly on OOD targets without adaptation.
• Text prompts for datasets lacking clinical annotations are synthetically generated via GPT-5, which may introduce distribution shifts compared to human-written reports.
• Uncertainty calibration metrics (Brier score, Spearman correlation) are computed exclusively over foreground regions, not the full image background.

Evidence (verbatim from paper)

We use Dice Similarity Coefficient (DSC) and normalized surface distance (NSD) to compare segmentation accuracy.

Citation

@misc{koleilat2026medclipseg,
  title={MedCLIPSeg: Probabilistic Vision-Language Adaptation for Data-Efficient and Generalizable Medical Image Segmentation},
  author={Koleilat et al. (2026)},
  year={2026},
  note={arXiv:2602.20423}
}

• arXiv: 2602.20423