Ircad Liver Eval

---
name: ircad-liver-eval
description: Evaluates medical image segmentation models on liver CT volumes, testing their ability to accurately delineate organ boundaries using interactive or automatic refinement techniques. The protocol measures how well models handle low-contrast boundaries and varying slice geometries in clinical imaging. Use when the user wants to benchmark on IRCAD, or asks about evaluating this task. Reports Dice coefficient.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2006.15320
  bibtex_key: kitrungrotsakul2020interactive
  confidence: high
---

# ircad-liver-eval

> Interactive Deep Refinement Network for Medical Image Segmentation — Kitrungrotsakul et al. (2020) (arXiv:2006.15320, 2020)

## What this evaluates

Evaluates medical image segmentation models on liver CT volumes, testing their ability to accurately delineate organ boundaries using interactive or automatic refinement techniques. The protocol measures how well models handle low-contrast boundaries and varying slice geometries in clinical imaging.

## Datasets

- **IRCAD** — total ?; splits: train (-1), test (-1)

## Metrics

- `Dice coefficient` **(primary)** — range: [0, 1]
  - 2 * |A ∩ B| / (|A| + |B|), measuring the overlap between the predicted segmentation and the ground truth mask.
- `Sensitivity (SEN)` — range: [0, 1]
  - True Positives / (True Positives + False Negatives), measuring the recall of the target organ pixels.
- `Positive Predictive Value (PPV)` — range: [0, 1]
  - True Positives / (True Positives + False Positives), measuring the precision of the predicted region.

## Input / output format

**Input**: 2D slices extracted from 3D CT volumes. For interactive baselines, seed points are provided (e.g., background from boundary dilation, foreground from skeletonization, or top/bottom/left/right corner points).

**Output**: 2D binary segmentation mask per slice, aggregated into a 3D volume for final evaluation.

## Scoring recipe

```python
import numpy as np
def compute_metrics(pred, gold):
    pred = pred.astype(bool)
    gold = gold.astype(bool)
    intersection = np.logical_and(pred, gold).sum()
    union = np.logical_or(pred, gold).sum()
    dice = 2 * intersection / union if union > 0 else 0.0
    tp = np.logical_and(pred, gold).sum()
    fn = np.logical_and(np.logical_not(pred), gold).sum()
    fp = np.logical_and(pred, np.logical_not(gold)).sum()
    sen = tp / (tp + fn) if (tp + fn) > 0 else 0.0
    ppv = tp / (tp + fp) if (tp + fp) > 0 else 0.0
    return dice, sen, ppv
```

## Common pitfalls

- DEXTR performs poorly on this dataset due to medical image intensity variations rather than architectural flaws, making direct comparison with natural-image baselines misleading.
- Seed point generation differs across baselines (skeletonization vs. corner points), affecting interactive method comparability.
- Evaluation is performed on 2D slices but reported as 3D volume Dice scores, which may smooth over slice-level segmentation errors.

## Evidence (verbatim from paper)

> The overlap index (Dice coefficient), Sensitivity (SEN), and Positive Predictive Value (PPV) were used as evaluation measures. The leave-one-out method was used in our experiments. We selected one CT volume as a test image, and other CT volumes were used for training.

## Citation

```bibtex
@misc{kitrungrotsakul2020interactive,
  title={Interactive Deep Refinement Network for Medical Image Segmentation},
  author={Kitrungrotsakul et al. (2020)},
  year={2020},
  note={arXiv:2006.15320}
}
```

- arXiv: 2006.15320