Official code and resources repository for "Normal and Abnormal Pathology Knowledge-Augmented Vision-Language Model for Anomaly Detection in Pathology Images" [paper] (ICCV 2025) in PyTorch. If you use any content of this repo for your work, please cite the following bib entry:

@inproceedings{song2025normal,
  title={Normal and abnormal pathology knowledge-augmented vision-language model for anomaly detection in pathology images},
  author={Song, Jinsol and Wang, Jiamu and Nguyen, Anh Tien and Byeon, Keunho and Ahn, Sangjeong and Lee, Sung Hak and Kwak, Jin Tae},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={22066--22076},
  year={2025}
}

The lymph node WSI dataset used for training cannot be publicly released at this time.
To enable reproducibility, we provide the trained model weights as follows:

• Original trained weights: Used to obtain the experimental results reported in the paper's tables (100 normal words and 50 abnormal words).
• Re-trained weights: Re-trained using a cleaned vocabulary with duplicates removed, corresponding to the word counts reported in the paper (92 normal words and 48 abnormal words).

To load a pretrained model without further training, set "train": 0 and "update_centroids": 0, and specify the path to the pretrained weights in "trained_model_pth".

Note on vocabulary sizes:
In the paper, we report 92 normal words and 48 abnormal words. However, the table results were generated using 100 normal words and 50 abnormal words (with duplicates). The cleaned-word model was trained later using a deduplicated vocabulary to match the reported counts. The normal and abnormal pathology terms used in training can be found in the words folder.

The paths specified by "train_imgs_path" and "val_imgs_path" should contain processed patch images, not raw WSI files. All images in these directories and their subfolders must be normal images only. The folder structure can be arbitrary.

Ideally, validation normal images should not overlap with the training data, as this provides a better estimate of model performance (as done in the paper). However, if sufficient new validation images are not available, it is acceptable to use a subset of the training dataset for validation.

For our experiments, we used a quite large training dataset, performing 1 epoch with a batch size of 100 and accumulating 100 batches before each weight update; you may adjust these settings depending on the size of your dataset.

To compute patch-level anomaly scores, set "test": 1 and "test_patches_WSIs_path". The top path should contain subfolders, each named after a WSI's ID ({WSI_id}). Inside each WSI folder, include only the patch images corresponding to that WSI. The folder structure should look like:

your_test_WSI_patch_folders_top_path/
├── WSI_001/
│   ├── WSI_001_0_0_50.png
│   ├── WSI_001_512_0_40.png
│   └── ...
├── WSI_002/
│   ├── WSI_002_0_0_10.png
│   ├── WSI_002_1024_512_0.png
│   └── ...
└── ...

To compute WSI-level anomaly scores after obtaining patch-level scores, set "WSI_level": 1 and "test_raw_WSIs_path". The top path should contain the raw WSI files named as {WSI_id}.svs, corresponding to the WSI IDs used for patch generation. The folder structure should look like:

your_test_raw_WSI_files_top_path/
├── WSI_001.svs
├── WSI_002.svs
├── WSI_003.svs
└── ...

To apply the 3×3 erosion described in the paper, the coordinates of each patch are required. Therefore, patch filenames extracted from each WSI should follow the format: {WSI_id}_{w_coord}_{h_coord}_{abnormal_percent(optional)}.png. If you are only computing patch-level anomaly scores or WSI-level anomaly scores without applying erosion, following this filename convention is not required. When computing WSI-level anomaly scores with the 3×3 erosion applied, make sure the "WSI_downsampling_factor" is set to the same value as the patch size specified in "test_patch_size" (both width and height of the patch). This setting matches the configuration used in the paper.