README.md

---
license: cc-by-nc-4.0
task_categories:
  - image-segmentation
tags:
  - medical-imaging
  - computed-tomography
  - liver
  - lesion-segmentation
  - non-contrast-ct
  - multi-phase-ct
  - benchmark
  - miccai
size_categories:
  - n<1K
pretty_name: TriALS — Triphasic-Aided Liver Lesion Segmentation
extra_gated_heading: "Request access to TriALS"
extra_gated_description: >-
  TriALS is released for non-commercial research use only. Please complete
  the fields below to request access.
extra_gated_fields:
  Full name: text
  Institution: text
  Country: country
  Role:
    type: select
    options:
      - PhD student
      - Postdoc
      - Faculty / PI
      - Research scientist (industry)
      - Clinician
      - label: Other
        value: other
  I will use TriALS for non-commercial academic research only: checkbox
extra_gated_button_content: "Request access"
---

# TriALS: Triphasic-Aided Liver Lesion Segmentation Benchmark in Non-Contrast CT

<p align="center">
  <img src="https://raw.githubusercontent.com/xmed-lab/TriALS/main/assets/multiphase_example.png" width="900" alt="Four-phase CT acquisition with lesion annotations"/>
</p>

*Same patient across all four CT phases (non-contrast, arterial, portal venous, delayed). Top row: raw images. Bottom row: lesion annotations. Many lesions are occult on non-contrast CT and only become conspicuous after contrast administration — this is the core diagnostic challenge TriALS targets.*

TriALS is the first multi-centre benchmark for liver lesion segmentation in **non-contrast CT (NCCT)**, supported by aligned four-phase acquisitions (non-contrast, arterial, portal venous, delayed) from Egyptian and Chinese institutions. It was organised across the MICCAI 2024 and 2025 challenges and is designed to enable development of diagnostic AI for contrast-limited clinical settings.

- **Paper**: *TriALS: Triphasic-Aided Liver Lesion Segmentation Benchmark in Non-Contrast CT* (2026)
- **Code & evaluation**: https://github.com/xmed-lab/TriALS
- **License**: CC BY 4.0

## Dataset summary

| | |
|---|---|
| **Released cases** | 80 (training only) |
| **Volumes released** | 320 (4 phases × 80) |
| **Institutions** | Ain Shams University (Egypt), Sun Yat-Sen Memorial Hospital (China) |
| **Modality** | Abdominal CT, NIfTI, native clinical resolution, Hounsfield units (no resampling) |
| **Annotations** | Per-phase lesion masks (4) + combined label registered to NCCT |

**Note on the test sets.** Egypt internal, China internal, and the external (Nanfang Hospital) cohorts are **not released publicly** and are retained for ongoing validation. Researchers who wish to evaluate their method on these cohorts can contact the corresponding authors to arrange held-out evaluation via the challenge pipeline.

## Tasks

- **Task 1 — Venous-phase segmentation.** Standard contrast-enhanced lesion segmentation on the portal venous phase.
- **Task 2 — Non-contrast segmentation.** Lesion segmentation on NCCT under two paradigms:
  - *Visible*: evaluated against lesions annotated from NCCT alone.
  - *Combined*: evaluated against the multi-phase fused label (full lesion burden, including lesions occult on NCCT). This is the primary clinical target.

<p align="center">
  <img src="https://raw.githubusercontent.com/xmed-lab/TriALS/main/assets/sample_case.png" width="720" alt="Sample case: venous CT and lesion label"/>
</p>

*A single case: portal venous CT (left) and the corresponding lesion label within the liver region (right).*

## Directory structure

```
Dataset_TriALS/
├── imagesTr/
│   ├── TriALS-0/
│   │   ├── TriALS-0_nocontrast.nii.gz
│   │   ├── TriALS-0_arterial.nii.gz
│   │   ├── TriALS-0_venous.nii.gz
│   │   └── TriALS-0_delayed.nii.gz
│   ├── TriALS-1/
│   │   └── ...
│   └── TriALS-<caseID>/
└── labelsTr/
    ├── TriALS-0/
    │   ├── TriALS-0_nocontrast.nii.gz
    │   ├── TriALS-0_arterial.nii.gz
    │   ├── TriALS-0_venous.nii.gz
    │   ├── TriALS-0_delayed.nii.gz
    │   └── TriALS-0_combined.nii.gz
    ├── TriALS-1/
    │   └── ...
    └── TriALS-<caseID>/
```

Each case has its own folder under `imagesTr/` and `labelsTr/`. Filenames follow `TriALS-<caseID>_<phase>.nii.gz`.

### Case ID convention

| Centre | Case ID range | n |
|---|---|---|
| Egypt (Ain Shams University) | `TriALS-0` … `TriALS-59` | 60 |
| China (Sun Yat-Sen Memorial) | `TriALS-200` … `TriALS-219` | 20 |

China cases are numbered from 200 onwards to keep centre identity recoverable from the filename.

### Phase convention

| Phase tag | Meaning | In images | In labels |
|---|---|---|---|
| `nocontrast` | Non-contrast CT | ✓ | ✓ |
| `arterial` | Arterial phase | ✓ | ✓ |
| `venous` | Portal venous phase | ✓ | ✓ |
| `delayed` | Delayed phase | ✓ | ✓ |
| `combined` | Multi-phase fused label, registered to NCCT | — | ✓ |

## Download

### Full dataset

```python
from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="marwankefah/TriALS",
    repo_type="dataset",
)
```

### Selective download by task and centre

Choose `TASK` ∈ {`"task1"`, `"task2"`, `"all"`} and `CENTRE` ∈ {`"egypt"`, `"china"`, `"both"`}:

```python
from huggingface_hub import snapshot_download

# ---- Configure ----
TASK   = "task1"   # "task1" (venous), "task2" (NCCT + combined label), or "all"
CENTRE = "both"    # "egypt", "china", or "both"
# -------------------

centre_ranges = {
    "egypt": list(range(0, 60)),
    "china": list(range(200, 220)),
    "both":  list(range(0, 60)) + list(range(200, 220)),
}

task_phases = {
    "task1": {"images": ["venous"],
              "labels": ["venous"]},
    "task2": {"images": ["nocontrast"],
              "labels": ["nocontrast", "combined"]},
    "all":   {"images": ["nocontrast", "arterial", "venous", "delayed"],
              "labels": ["nocontrast", "arterial", "venous", "delayed", "combined"]},
}

phases = task_phases[TASK]
patterns = []
for cid in centre_ranges[CENTRE]:
    case = f"TriALS-{cid}"
    for p in phases["images"]:
        patterns.append(f"Dataset_TriALS/imagesTr/{case}/{case}_{p}.nii.gz")
    for p in phases["labels"]:
        patterns.append(f"Dataset_TriALS/labelsTr/{case}/{case}_{p}.nii.gz")

path = snapshot_download(
    repo_id="marwankefah/TriALS",
    repo_type="dataset",
    allow_patterns=patterns,
)
print(f"Downloaded to: {path}")
```

A few common recipes:

```python
# Task 1, Egyptian cohort only
TASK, CENTRE = "task1", "egypt"

# Task 2 with multi-phase training data (all 4 image phases + NCCT/combined labels)
TASK, CENTRE = "all", "both"
# then restrict labels in your dataloader to ["nocontrast", "combined"]
```

### Load a volume

```python
import os, nibabel as nib

root = f"{path}/Dataset_TriALS"
case  = "TriALS-0"
phase = "venous"

img = nib.load(os.path.join(root, "imagesTr", case, f"{case}_{phase}.nii.gz"))
lbl = nib.load(os.path.join(root, "labelsTr", case, f"{case}_{phase}.nii.gz"))

# Combined multi-phase label (labels only, registered to NCCT)
combined = nib.load(os.path.join(root, "labelsTr", case, f"{case}_combined.nii.gz"))
```

## Annotation protocol

Two-stage hybrid human–algorithm protocol. An automated model trained on LiTS and pilot in-house cases produced preliminary segmentations, which were propagated across phases via non-rigid registration. These were then manually corrected by 10 trained annotators (radiology residents/fellows) in 3D Slicer; each case was reviewed by a second annotator, with disagreements escalated to a senior abdominal radiology consultant.

### Combined label

<p align="center">
  <img src="https://raw.githubusercontent.com/xmed-lab/TriALS/main/assets/combined_label.png" width="720" alt="Combined multi-phase label construction"/>
</p>

*Lesions annotated on each contrast phase (non-contrast, arterial, venous, delayed) are registered to the non-contrast frame and fused into a single combined label that captures the full lesion burden — including lesions occult on NCCT alone.*

The combined label was constructed by rigid-then-non-rigid registration of the contrast phases to NCCT via elastix, followed by multi-phase label fusion and consultant-level QA. All raw per-phase volumes and per-phase labels are released alongside the combined labels to enable independent verification of the registration and fusion pipeline.

## Ethics and data governance

- **Egyptian cohort** (Ain Shams University Hospitals, Cairo): collected under local Research Ethics Committee approval (FWA 000017585), with waiver of individual informed consent for retrospective de-identified use.
- **Chinese cohort** (Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University): approved by the institutional review committee of Sun Yat-sen Memorial Hospital, Sun Yat-sen University.

All volumes were fully anonymised prior to transfer; direct patient identifiers, dates, and institution-identifying metadata were removed or shifted.

## Citation

If you use TriALS, please cite:

```bibtex
```

## Acknowledgements

Supported by the Research Grants Council of Hong Kong (T45-401/22-N), the National Natural Science Foundation of China (62306254; 82001768), and the Guangdong Basic and Applied Basic Research Foundation (2021A1515010226). TriALS is endorsed by SIG-AFRICAI (MICCAI Society).

## Contact

For questions, collaboration, or requests to evaluate on the held-out test cohorts, contact the corresponding authors listed on the paper.