Puma Challenge Eval

puma-challenge-eval

Leveraging Pathology Foundation Models for Panoptic Segmentation of Melanoma in H&E Images — Jiaqi Lv et al. (2025) (arXiv:2507.13974, 2025)

What this evaluates

Evaluates pixel-level segmentation capability for distinguishing five histopathological tissue classes (tumour, stroma, necrosis, blood vessels, epidermis) in melanoma H&E images.

Datasets

• PUMA Challenge dataset — total ?; splits: train (-1), preliminary_test (10), final_test (-1)

Metrics

• Dice score (primary) — range: [0, 1]
  • Computes the overlap between predicted and ground truth segmentation masks. Calculated as 2 * |A ∩ B| / (|A| + |B|), reported as micro-average across all classes and per-class.

Input / output format

Input: RGB H&E histopathology image patches.

Output: 5-channel pixel-level segmentation mask corresponding to tumour, stroma, necrosis, blood vessels, and epidermis.

Scoring recipe

def compute_dice(pred, gt, num_classes=5):
    dice_scores = []
    for c in range(num_classes):
        p = pred[:, c]
        g = gt[:, c]
        intersection = np.sum(p * g)
        dice = (2.0 * intersection) / (np.sum(p) + np.sum(g) + 1e-6)
        dice_scores.append(dice)
    micro_dice = np.mean(dice_scores)
    return micro_dice, dice_scores

Common pitfalls

• Class imbalance causes high variance in Dice scores for rare classes like necrosis and blood vessels.
• The preliminary test set contains only 10 images, making it unrepresentative of overall performance.
• Micro-average Dice can mask poor performance on minority classes if not examined per-class.

Evidence (verbatim from paper)

Our method achieved the highest average Dice score $(68.23%)$ . We found that incorporating the dual-stage loss strategy led to an improvement in segmentation performance, from $66.84%$ to $68.23%$ . However, we also observed a large standard deviation in Dice scores for the rarer tissue classes, notably epidermis, blood vessels, and especially necrosis.

Citation

@misc{lv2025leveraging,
  title={Leveraging Pathology Foundation Models for Panoptic Segmentation of Melanoma in H&E Images},
  author={Jiaqi Lv et al. (2025)},
  year={2025},
  note={arXiv:2507.13974}
}

• arXiv: 2507.13974