README.md

---
license: apache-2.0
---
# Dataset Card for MedIAnomaly

## Dataset Description

**MedIAnomaly** is a benchmark designed to evaluate anomaly detection methods in the medical imaging domain. It provides a standardized evaluation protocol across **seven real-world medical image datasets**, including both **image-level anomaly classification (AnoCls)** and **pixel-level anomaly segmentation (AnoSeg)** tasks.  

All datasets follow a **one-class training setting**, where **only normal (non-anomalous) images are available in the training set**, and the **test set includes both normal and abnormal cases**. This reflects real-world scenarios where anomalies are rare and not annotated during training.

The benchmark includes a total of **seven datasets**, spanning across various imaging modalities (X-ray, MRI, fundus, dermatoscopy, histopathology), and ensures unified data format and preprocessing to support fair and reproducible comparison of anomaly detection methods.

![dataset](https://huggingface.co/datasets/randall-lab/medianomaly/resolve/main/example.png)

## Dataset Source
- **Homepage**: [https://github.com/caiyu6666/MedIAnomaly](https://github.com/caiyu6666/MedIAnomaly)
- **License**: [Apache License 2.0](http://www.apache.org/licenses/LICENSE-2.0)
- **Paper**: Yu Cai et al. _MedIAnomaly: A Comparative Study of Anomaly Detection in Medical Images_, arXiv 2024.

## Dataset Structure

| Dataset      | Modality              | Task              | 𝒟<sub>train</sub> | 𝒟<sub>test</sub> (Normal+Abnormal) |
|--------------|-----------------------|-------------------|------------------|----------------------------|
| RSNA         | Chest X-ray           | AnoCls            | 3851             | 1000 + 1000                |
| VinDr-CXR    | Chest X-ray           | AnoCls            | 4000             | 1000 + 1000                |
| Brain Tumor  | Brain MRI             | AnoCls            | 1000             | 600 + 600                  |
| LAG          | Retinal fundus image  | AnoCls            | 1500             | 811 + 811                  |
| ISIC2018     | Dermatoscopic image   | AnoCls            | 6705             | 909 + 603                  |
| Camelyon16   | Histopathology image  | AnoCls            | 5088             | 1120 + 1113                |
| BraTS2021    | Brain MRI             | AnoCls & AnoSeg   | 4211             | 828 + 1948                 |

### Notes on Dataset-Specific Definitions

- **RSNA**: Training images are all normal chest X-rays. Test set contains a balanced mix of normal and pneumonia images.
- **VinDr-CXR**: Training set consists only of normal chest X-rays. Test set includes both normal and abnormal findings.
- **Brain Tumor**: MRI scans. All training samples are healthy brains; test set contains normal and tumor cases.
- **LAG**: Retinal fundus images. Training set includes only normal cases; glaucomatous images appear in test set.
- **ISIC2018**: One-hot multi-label data. Only images with `NV = 1` and all other labels = 0 are considered **normal**. All others (with any other disease present) are considered **abnormal**.
- **Camelyon16**: Histopathological whole-slide patches. Training includes only benign tissue. Abnormal cancerous regions are tested.
- **BraTS2021**: Brain MRI for both classification and segmentation. Training includes only normal images. Test set includes tumor cases with segmentation masks.

## Example Usage

### RSNA
```python
from datasets import load_dataset

dataset = load_dataset("randall-lab/medianomaly", name="rsna", split="train", trust_remote_code=True)
# dataset = load_dataset("randall-lab/medianomaly", name="rsna", split="test", trust_remote_code=True)

# View a sample
example = dataset[0]
image = example["image"]
label = example["label"]  # "normal" or "abnormal"

image.show()
print(f"Label: {label}")
```

### Vin-CXR
```python
from datasets import load_dataset

dataset = load_dataset("randall-lab/medianomaly", name="vincxr", split="train", trust_remote_code=True)
# dataset = load_dataset("randall-lab/medianomaly", name="vincxr", split="test", trust_remote_code=True)

# View a sample
example = dataset[0]
image = example["image"]
label = example["label"]  # "normal" or "abnormal"

image.show()
print(f"Label: {label}")
```

### Brain Tumor
```python
from datasets import load_dataset

dataset = load_dataset("randall-lab/medianomaly", name="braintumor", split="train", trust_remote_code=True)
# dataset = load_dataset("randall-lab/medianomaly", name="braintumor", split="test", trust_remote_code=True)

# View a sample
example = dataset[0]
image = example["image"]
label = example["label"]  # "normal" or "abnormal"

image.show()
print(f"Label: {label}")
```

### LAG
```python
from datasets import load_dataset

dataset = load_dataset("randall-lab/medianomaly", name="lag", split="train", trust_remote_code=True)
# dataset = load_dataset("randall-lab/medianomaly", name="lag", split="test", trust_remote_code=True)

# View a sample
example = dataset[0]
image = example["image"]
label = example["label"]  # "normal" or "abnormal"

image.show()
print(f"Label: {label}")
```

### Camelyon16
```python
from datasets import load_dataset

dataset = load_dataset("randall-lab/medianomaly", name="camelyon16", split="train", trust_remote_code=True)
# dataset = load_dataset("randall-lab/medianomaly", name="camelyon16", split="test", trust_remote_code=True)

# View a sample
example = dataset[0]
image = example["image"]
label = example["label"]  # "normal" or "abnormal"

image.show()
print(f"Label: {label}")
```

### BraTS2021
```python
from datasets import load_dataset

# Train
dataset = load_dataset("randall-lab/medianomaly", name="brats2021", split="train", trust_remote_code=True)

example = dataset[0]
image = example["image"]
label = example["label"]  # "normal" or "abnormal"

image.show()
print(f"Label: {label}")

# Test
dataset = load_dataset("randall-lab/medianomaly", name="brats2021", split="test", trust_remote_code=True)

example = dataset[828] # >= 828 is abnormal images with seg mask
image = example["image"]
label = example["label"]  # "normal" or "abnormal"
anno = example["annotation"] # None if label is 0, seg mask if label is 1

image.show()
anno.show()
print(f"Label: {label}")
```

### ISIC2018
```python
from datasets import load_dataset

dataset = load_dataset("randall-lab/medianomaly", name="isic2018_task3", split="train", trust_remote_code=True)
# dataset = load_dataset("randall-lab/medianomaly", name="isic2018_task3", split="test", trust_remote_code=True)

# View a sample
example = dataset[0]
image = example["image"]
label = example["label"]  # "normal" or "abnormal"
labels = example["labels"] # one-hot multi label for different disease [MEL, NV, BCC, AKIEC, BKL, DF, VASC]

# Individual binary class labels (0 or 1)
mel_label = example["MEL"]
nv_label = example["NV"]
bcc_label = example["BCC"]
akiec_label = example["AKIEC"]
bkl_label = example["BKL"]
df_label = example["DF"]
vasc_label = example["VASC"]

image.show()
print(f"Label: {label}")
```

If you are using colab, you should update datasets to avoid errors
```
pip install -U datasets
```
## Citation
```
@article{cai2024medianomaly,
  title={MedIAnomaly: A comparative study of anomaly detection in medical images},
  author={Cai, Yu and Zhang, Weiwen and Chen, Hao and Cheng, Kwang-Ting},
  journal={arXiv preprint arXiv:2404.04518},
  year={2024}
}
```