Cads Whole Body Ct Eval

---
name: cads-whole-body-ct-eval
description: Evaluates AI models on voxel-level segmentation of 167 whole-body anatomical structures from CT scans. It probes generalization across diverse imaging protocols, patient demographics, and pathological conditions, with a focus on clinical utility in radiation oncology. Use when the user wants to benchmark on CADS-dataset, 18 Public Benchmark Datasets, or asks about evaluating this task. Reports Dice coefficient.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2507.22953
  bibtex_key: xu2025cads
  confidence: high
---

# cads-whole-body-ct-eval

> CADS: A Comprehensive Anatomical Dataset and Segmentation for Whole-Body Anatomy in Computed Tomography — Murong Xu et al. (2025) (arXiv:2507.22953, 2025)

## What this evaluates

Evaluates AI models on voxel-level segmentation of 167 whole-body anatomical structures from CT scans. It probes generalization across diverse imaging protocols, patient demographics, and pathological conditions, with a focus on clinical utility in radiation oncology.

## Datasets

- **CADS-dataset** — total 22022; splits: train (-1), test (-1); repo https://github.com/murong-xu/CADS
- **18 Public Benchmark Datasets** — total ?; splits: test (-1)

## Metrics

- `Dice coefficient` **(primary)** — range: [0, 1]
  - Measures voxel-wise overlap between predicted and ground truth masks: 2 * |A ∩ B| / (|A| + |B|). Reported as a decimal or percentage.
- `95% Hausdorff Distance (HD95)` — range: mm
  - Computes the 95th percentile of the maximum distance between corresponding points on the predicted and ground truth segmentation boundaries. Lower values indicate better boundary alignment.

## Input / output format

**Input**: 3D CT scan volumes (typically aligned to 1.5mm isotropic resolution) covering anatomical regions from head to knees.

**Output**: Voxel-level segmentation masks for 167 predefined anatomical structures.

## Scoring recipe

```python
def compute_metrics(pred, gold):
    intersection = np.sum(pred & gold)
    dice = 2.0 * intersection / (np.sum(pred) + np.sum(gold))
    pred_bd = np.argwhere(pred)
    gold_bd = np.argwhere(gold)
    dists = cdist(pred_bd, gold_bd)
    hd95 = np.percentile(dists, 95) if len(dists) > 0 else 0.0
    return dice, hd95
```

## Common pitfalls

- Relying on algorithm-generated ground truth labels (e.g., original TotalSegmentator annotations) which contain known inaccuracies for structures like ribs and vertebrae.
- Failing to stratify evaluation by annotation completeness (primary vs. secondary cohorts), which skews performance estimates in real-world clinical settings.
- Comparing models with different architectures without isolating dataset effects, leading to confounded generalization claims.

## Evidence (verbatim from paper)

> We use two primary metrics: Dice coefficient for volume overlap and 95% Hausdorff Distance (HD95) for boundary precision, while additional performance measures are detailed in Online Method Section 5.4.

## Citation

```bibtex
@misc{xu2025cads,
  title={CADS: A Comprehensive Anatomical Dataset and Segmentation for Whole-Body Anatomy in Computed Tomography},
  author={Murong Xu et al. (2025)},
  year={2025},
  note={arXiv:2507.22953}
}
```

- arXiv: 2507.22953