Revisiting Reconstruction-based Anomaly Detection for Whole Slide Image (IEEE TMI 2026)

Dataset Preparation

We use the Camelyon16-BMAD dataset as an example to reproduce the experimental results.

Download the dataset from the following link:

https://drive.google.com/drive/folders/1AC-wWZl_K18CWL2eIxUScoSOoxT4IBuw

The file Histopathology_AD.zip corresponds to the Camelyon16-BMAD dataset.

After downloading, extract the dataset:

unzip Histopathology_AD.zip

After extraction, the directory structure should look like the following:

camelyon16_256/
├── test/
│   ├── good/
│   └── Ungood/
├── train/
│   └── good/
└── valid/
    ├── good/
    └── Ungood/

Configuration

Modify the dataset path and output path in the configuration file.

configs/conch/conch_bmad_config_ecr4ad.yaml

Update the following fields:

data_root: /path/to/dataset
save_path: /path/to/save/results

Foundation Model Weights

Before running training, please update the path of the foundation model weights in:

networks/vit_encoder.py

For example, when using CONCH, modify the following code:

if "conch" == name.lower():
    # your pytorch_model.bin file path here
    model, preprocess = create_model_from_pretrained(
        'conch_ViT-B-16',
        "your pytorch_model.bin file path here"
    )

Replace "your pytorch_model.bin file path here" with the actual path to the downloaded CONCH pretrained weights.

Weights & Biases (Wandb)

Training logs are recorded using Weights & Biases.

Before running the training script, add your Wandb API key in patch_train.py:

wandb.login(key="your_wandb_key_here")

Training

Run the training script:

python patch_train.py

During training, the script will automatically:

• Load the dataset
• Train the model
• Evaluate the model performance
• Log training metrics to Wandb

Other Datasets

The Camelyon16 and GleasonArvaniti datasets can be found at https://camelyon17.grand-challenge.org/, and https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/OCYCMP, respectively.

Citation

If you find this work useful, please consider citing:

@article{xiao2026revisiting,
  title={Revisiting Reconstruction-based Anomaly Detection for Whole Slide Image},
  author={Xiao, Bin and Wangulu, Collins and van der Kwast, Theodorus and Yousef, George M and Zabihollahy, Fatemeh},
  journal={IEEE Transactions on Medical Imaging},
  year={2026},
  publisher={IEEE}
}

Acknowledgements

This repository is built upon the implementation of Dinomaly. We sincerely thank the authors for making their code publicly available.