Semigda Medical Seg Eval

---
name: semigda-medical-seg-eval
description: Evaluates semi-supervised medical image segmentation performance under limited labeled data ratios (10% and 30%). It measures segmentation accuracy and boundary precision across multiple medical domains including colonoscopy, dermoscopy, pathology, and ultrasound. Use when the user wants to benchmark on Colonoscopy (CVC-ClinicDB, Kvasir, CVC-300), ISIC-2018, BCSS, BUSI, or asks about evaluating this task. Reports Dice coefficient (Dice).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.23274
  bibtex_key: [REDACTED]
  confidence: high
---

# semigda-medical-seg-eval

> SemiGDA: Generative Dual-distribution Alignment for Semi-Supervised Medical Image Segmentation — Huang et al. (2026) (arXiv:2604.23274, 2026)

## What this evaluates

Evaluates semi-supervised medical image segmentation performance under limited labeled data ratios (10% and 30%). It measures segmentation accuracy and boundary precision across multiple medical domains including colonoscopy, dermoscopy, pathology, and ultrasound.

## Datasets

- **Colonoscopy (CVC-ClinicDB, Kvasir, CVC-300)** — total ?; splits: train (1450), val (145), test (-1)
- **ISIC-2018** — total 3594; splits: train (2075), val (519), test (1000)
- **BCSS** — total 3888; splits: train (-1), val (-1), test (-1)
- **BUSI** — total 647; splits: train (-1), test (-1)

## Metrics

- `Dice coefficient (Dice)` **(primary)** — range: percent
  - Dice = 2 * |A ∩ B| / (|A| + |B|), where A and B are the predicted and ground truth segmentation masks. Measures overlap between prediction and target.
- `Intersection over Union (IoU)` — range: percent
  - IoU = |A ∩ B| / |A ∪ B|. Measures the ratio of intersection area to union area between prediction and ground truth.
- `95% Hausdorff Distance (95HD)` — range: other
  - 95th percentile of the maximum distance from any point on the predicted boundary to the closest point on the ground truth boundary. Measures boundary precision.

## Input / output format

**Input**: 224×224 medical images (colonoscopy, dermoscopy, pathology patches, ultrasound) with corresponding segmentation masks for labeled samples; unlabeled samples provided without masks.

**Output**: Binary segmentation mask of the same spatial dimensions as the input image.

## Scoring recipe

```python
def evaluate(preds, gts):
    dice_scores, iou_scores, hd95_scores = [], [], []
    for pred, gt in zip(preds, gts):
        inter = np.sum(pred & gt)
        dice = 2.0 * inter / (np.sum(pred) + np.sum(gt))
        iou = inter / np.sum(pred | gt)
        pred_pts = np.argwhere(pred)
        gt_pts = np.argwhere(gt)
        dists = cdist(pred_pts, gt_pts)
        hd95 = np.percentile(np.max(dists, axis=1), 95)
        dice_scores.append(dice)
        iou_scores.append(iou)
        hd95_scores.append(hd95)
    return np.mean(dice_scores), np.mean(iou_scores), np.mean(hd95_scores)
```

## Common pitfalls

- Semi-supervised setting uses only 10% or 30% labeled data for training, with the rest unlabeled; evaluation is strictly on the test split.
- Inference averages two predictions (likely from dual encoders or teacher-student setup), which must be replicated to match reported scores.
- All input images are resized to 224×224 before processing, which may impact boundary metrics like 95HD compared to full-resolution evaluation.

## Evidence (verbatim from paper)

> We employ three commonly used metrics, namely the Dice coefficient (Dice), Intersection over Union (IoU), and 95% Hausdorff Distance (95HD).

## Citation

```bibtex
@misc{huang2026semigda,
  title={SemiGDA: Generative Dual-distribution Alignment for Semi-Supervised Medical Image Segmentation},
  author={Huang et al. (2026)},
  year={2026},
  note={arXiv:2604.23274}
}
```

- arXiv: 2604.23274