Glas Gland Segmentation Eval

glas-gland-segmentation-eval

Weakly Supervised Teacher-Student Framework with Progressive Pseudo-mask Refinement for Gland Segmentation — Khan et al. (2026) (arXiv:2603.08605, 2026)

What this evaluates

Evaluates a weakly supervised segmentation model's ability to accurately delineate glandular structures in colorectal histopathology images using sparse annotations. It also probes cross-domain generalization across different institutional cohorts with varying staining protocols and scanner characteristics.

Datasets

• GlaS — total ?; splits: test (-1)

Metrics

• mIoU (primary) — range: percent
  • Mean Intersection over Union: the average IoU score computed per class across all images. IoU = TP / (TP + FP + FN).
• mDice — range: percent
  • Mean Dice coefficient: the average Dice score computed per class. Dice = 2TP / (2TP + FP + FN).

Input / output format

Input: H&E stained histopathology image patches or whole-slide images.

Output: Pixel-wise segmentation mask assigning each pixel to one of four classes: malignant glands, benign glands, poorly differentiated clusters/glands, or background stroma.

Scoring recipe

def compute_metrics(pred_mask, gt_mask, num_classes=4):
    ious, dices = [], []
    for c in range(num_classes):
        pred_c = (pred_mask == c)
        gt_c = (gt_mask == c)
        intersection = np.logical_and(pred_c, gt_c).sum()
        union = np.logical_or(pred_c, gt_c).sum()
        iou = intersection / union if union > 0 else 0.0
        dice = 2 * intersection / (pred_c.sum() + gt_c.sum()) if (pred_c.sum() + gt_c.sum()) > 0 else 0.0
        ious.append(iou)
        dices.append(dice)
    return np.mean(ious) * 100, np.mean(dices) * 100

Common pitfalls

• Metrics are reported as mean ± standard deviation across five independent training runs with different random seeds; single-run results will not match the paper's values.
• Quantitative evaluation is strictly limited to the GlaS dataset; external cohorts (TCGA-COAD, TCGA-READ, SPIDER) lack pixel-level ground truth and are only assessed qualitatively.
• Performance on SPIDER degrades due to domain shift (staining heterogeneity, lower quality), which may cause fragmented boundaries and false positives if not explicitly noted.

Evidence (verbatim from paper)

Metrics are mean Intersection over Union (mIoU) and mean Dice coefficient (mDice), reported as mean ± standard deviation across five independent training runs with different random seeds. Our framework achieves competitive state-of-the-art performance on the GlaS benchmark, achieving an mIoU of 80.10% and an mDice of 89.10%.

Citation

@misc{khan2026weakly,
  title={Weakly Supervised Teacher-Student Framework with Progressive Pseudo-mask Refinement for Gland Segmentation},
  author={Khan et al. (2026)},
  year={2026},
  note={arXiv:2603.08605}
}

• arXiv: 2603.08605