Medical Segmentation Iac Eval

---
name: medical-segmentation-iac-eval
description: This protocol evaluates medical image segmentation models by measuring how accurately they predict anatomical structures across multiple modalities and organs. It specifically probes the ability of adaptive skip connections and differentiable search strategies to maintain or improve segmentation quality under different training regimes and backbone architectures. Use when the user wants to benchmark on ACDC, AMOS, BraTS, KiTS, or asks about evaluating this task. Reports macro-average soft Dice.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.14849
  bibtex_key: benedykciuk2026efficient
  confidence: high
---

# medical-segmentation-iac-eval

> Efficient Search of Implantable Adaptive Cells for Medical Image Segmentation — Benedykciuk et al. (2026) (arXiv:2604.14849, 2026)

## What this evaluates

This protocol evaluates medical image segmentation models by measuring how accurately they predict anatomical structures across multiple modalities and organs. It specifically probes the ability of adaptive skip connections and differentiable search strategies to maintain or improve segmentation quality under different training regimes and backbone architectures.

## Datasets

- **ACDC** — total ?; splits: train (-1), val (-1), test (60)
- **AMOS** — total ?; splits: train (-1), val (-1), test (72)
- **BraTS** — total ?; splits: train (-1), val (-1), test (250)
- **KiTS** — total ?; splits: train (-1), val (-1), test (98)

## Metrics

- `macro-average soft Dice` **(primary)** — range: [0, 1]
  - Macro-average over foreground classes of the soft Dice coefficient: (1/|F|) * sum_{c in F} (2*sum_x(y_c*p_c) + eps) / (sum_x(y_c) + sum_x(p_c) + eps), with eps=1e-5. Background is explicitly excluded from the average.

## Input / output format

**Input**: 2-D axial slices (128×128 centered crop) extracted from 3-D volumes. MRI inputs use per-modality z-score normalization over non-zero voxels; CT inputs use standard windowing followed by z-score normalization. Paired with one-hot ground-truth masks.

**Output**: Softmax probabilities p_c per class per pixel, or discrete segmentation masks.

## Scoring recipe

```python
patient_dices = []
for patient in test_set:
    tp, fp, fn = 0, 0, 0
    for slice in patient.slices:
        pred = model.predict(slice)
        gt = slice.gt
        tp += np.sum((pred == gt) & (gt > 0))
        fp += np.sum((pred == 1) & (gt == 0))
        fn += np.sum((pred == 0) & (gt > 0))
    dice = 2 * tp / (2 * tp + fp + fn + 1e-5)
    patient_dices.append(dice)
return np.mean(patient_dices)
```

## Common pitfalls

- Evaluation aggregates TP/FP/FN across all slices per patient before computing Dice, rather than averaging per-slice Dice scores.
- The macro-average explicitly excludes the background class, which can inflate scores if foreground regions are small.
- Statistical significance testing (Wilcoxon signed-rank) is applied to per-patient Dice scores, not per-slice predictions.

## Evidence (verbatim from paper)

> For evaluation, we compute Dice per test case (patient) for each foreground class and report per-class means and the mean foreground Dice (macro-average over foreground classes, background excluded). Dataset scores are averages of these per-case/patient-level values. Although all models operate on 2-D slices, evaluation is performed at the case level by aggregating true positives (TP), false positives (FP) and false negatives (FN) over all slices of a patient before computing Dice.

## Citation

```bibtex
@misc{benedykciuk2026efficient,
  title={Efficient Search of Implantable Adaptive Cells for Medical Image Segmentation},
  author={Benedykciuk et al. (2026)},
  year={2026},
  note={arXiv:2604.14849}
}
```

- arXiv: 2604.14849