Dataset

# Hepatic Thermal Ablation Planning Dataset

Dataset of 50 clinical hepatic cases (94 tumor targets) for percutaneous thermal ablation planning, sourced from the Colorectal-Liver-Metastases (CRLM) database of pre-segmented images. Each case includes CT volumes, anatomical meshes, safety map visualizations, and paired treatment plans (automatic and expert).

## Tier Structure

Cases are categorized by planning difficulty following the COLLISION classification of ablability:

| Tier | Cases | Description |
| ------ | ----- | ------------------------------------------------------ |
| Tier 1 | 25 | Single tumor, simple access, lowest risk |
| Tier 2 | 15 | Intermediate: multiple lesions or bilobar distribution |
| Tier 3 | 10 | Multi-lesion and bilobar, highest planning complexity |

## Directory Layout

```
dataset/
├── metrics.csv
├── tier_1/
│ └── case-XXXX/
├── tier_2/
│ └── case-XXXX/
└── tier_3/
└── case-XXXX/
```

Each case directory contains:

```
case-XXXX/
├── volumes/ CT scan and segmentation (NRRD)
├── meshes/ Anatomical structure meshes (VTK)
├── plans/ Treatment plans and needle markers (VTK + JSON)
└── images/ Safety map and overview visualizations (PNG)
```

## File Descriptions

### volumes/

| File | Description |
| ----------------------- | ------------------------------- |
| `ct_scan.nrrd` | CT volume |
| `segmentation.seg.nrrd` | Multi-label segmentation volume |

### meshes/

Surface meshes in VTK format, named `{case_id}_{structure}.vtk`. Standard structures include: liver, skin, vessels, bones, aorta, heart, lungs, kidneys, spleen, gallbladder, stomach, colon, small bowel, duodenum, esophagus, pancreas, adrenal gland, spinal cord, muscles. Each case also contains one or more `tumor_N.vtk` meshes.

### plans/

Treatment plans contain ablation zone ellipsoid meshes and needle trajectory markers.

**Ablation ellipsoids** (VTK): `{method}_{target}_{needle_config}_{burn}_ellipsoid.vtk`
- `method`: `auto` (optimizer) or `expert` (manual)
- `target`: `T1`, `T2` (tumor target index)
- `needle_config`: `N1`, `N2`, ... (needle index)
- `burn`: `B0`, `B1` (burn index along the needle axis, for pullback ablations)

**Needle markers** (JSON, 3D Slicer markups format):
- `T{n}-{i}.mrk.json` — expert needle trajectory
- `T{n}_{i}_AutoPlanner.mrk.json` — automatic planner needle trajectory

### images/

| File | Description |
| ------------------------------ | ------------------------------------------ |
| `overview.png` | 3D anatomical overview of the case |
| `safety_map_skin.png` | Safety map projected on the skin surface |
| `safety_map_equirect_T{n}.png` | Equirectangular safety map for target T{n} |
| `closeup_T{n}.png` | Close-up view of target T{n} |

Tier 1 cases (single target) contain T1 images only. Tier 2 and 3 cases contain both T1 and T2 images.

## Root CSV Files

### metrics.csv

Per-target paired evaluation metrics for both automatic and expert plans (188 rows = 94 targets x 2 methods). Includes coverage, efficiency, Dice coefficient, Hausdorff distance (HD95), distance to closest critical structure (DTC), insertion depth, and optimization timing.