README.md

August 25, 2026 · View on GitHub

SimSID: Unsupervised Anomaly Detection in Chest Radiography

SimSID is an unsupervised anomaly detection model for chest X-ray images. SimSID learns the recurrent anatomical patterns that normal chest radiographs share, then flags a test image as anomalous when its patterns do not fit — using no anomaly labels during training.

Anomaly detection in radiography is both easier and harder than in photographic images. It is easier because radiography is spatially structured: consistent imaging protocols mean the same anatomy lands in roughly the same place every time. It is harder because the anomalies are subtle, and annotating them takes medical expertise. SimSID exploits the first fact to work around the second.

This task is described in the literature as unsupervised anomaly detection, out-of-distribution detection, one-class learning, novelty detection, and normality modeling. SimSID applies it to two-dimensional chest radiography.

Note

SimSID is the journal extension of SQUID. SQUID (CVPR 2023, code at tiangexiang/SQUID) introduced the method. SimSID (IEEE TPAMI 2024) is the significant technical improvement over it, and is the version to use. If you are looking for the CVPR paper's code, it is in the SQUID repository; this repository supersedes it.

SimSID overview

Results

SimSID formulates anomaly detection as an image reconstruction task, using a space-aware memory matrix and an in-painting block in feature space. During training it taxonomizes the ingrained anatomical structures into recurrent visual patterns; at inference, patterns it has not seen read as anomalies.

Against the previous state of the art in unsupervised anomaly detection, SimSID improves AUC by:

benchmark	modality	SimSID AUC gain over prior state of the art
ZhangLab Chest X-ray	chest radiography	+8.0%
COVIDx	chest radiography	+5.0%
Stanford CheXpert	chest radiography	+9.9%

The earlier SQUID model surpassed 13 state-of-the-art unsupervised anomaly detection methods by at least 5 AUC points on two chest X-ray benchmarks. SimSID improves on SQUID.

Paper

Exploiting Structural Consistency of Chest Anatomy for Unsupervised Anomaly Detection in Radiography Images
Tiange Xiang¹, Yixiao Zhang², Yongyi Lu², Alan L. Yuille², Chaoyi Zhang¹, Weidong Cai¹, and Zongwei Zhou²
¹University of Sydney, ²Johns Hopkins University
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 46, no. 9, pp. 6070–6081, 2024
doi:10.1109/TPAMI.2024.3382009 | arXiv:2403.08689 | paper

SQUID: Deep Feature In-Painting for Unsupervised Anomaly Detection
Tiange Xiang¹, Yixiao Zhang², Yongyi Lu², Alan L. Yuille², Chaoyi Zhang¹, Weidong Cai¹, and Zongwei Zhou²
¹University of Sydney, ²Johns Hopkins University
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 23890–23901
paper | arXiv:2111.13495 | code

Installation

SimSID needs PyTorch, a CUDA-capable GPU, and a small set of standard scientific packages.

git clone https://github.com/MrGiovanni/SimSID.git
cd SimSID
conda create -n simsid python=3.10 -y
conda activate simsid
pip install -r requirements.txt

Reproducing the original environment (Python 3.6, CUDA 10.0)

environment.yml pins the exact 2020-era stack the paper was developed on: Python 3.6.10, CUDA 10.0, torchvision 0.5.0. Python 3.6 reached end of life in December 2021 and CUDA 10.0 does not support GPUs newer than Turing, so this will not resolve on most current machines. It is kept for the record.

conda env create -f environment.yml
conda activate simsid

Data

SimSID is trained and evaluated on three public chest X-ray benchmarks. Download each, then set self.data_root in configs/base.py to the directory holding them.

dataset	what to download	link
ZhangLab Chest X-ray	official train/test split plus our validation split	Google Drive
Stanford CheXpert	official train/validation split plus our test split	Google Drive
COVIDx	see `dataloader/dataloader_covidx.py`	—

configs/base.py ships a developer's local path as the default data_root. Change it before running anything.

Training SimSID

Experiments are driven by config files in configs/. Every config inherits from configs/base.py, so read that one first.

config	benchmark
`configs/zhang_dev.py`	ZhangLab Chest X-ray
`configs/chexpert_best.py`	Stanford CheXpert
`configs/covidx_dev.py`	COVIDx

python main.py --config zhang_dev.py --exp experiment_name

Checkpoints, TensorBoard logs, and sample test images are written to checkpoints/<exp>/.

Evaluating SimSID

python eval.py --exp experiment_name

eval.py reads the checkpoint written by main.py at checkpoints/<exp>/.

Important

No pre-trained SimSID weights are released yet. checkpoints/ is created by main.py during training; there is nothing to download into it. Train a model first, or open an issue if you need released weights.

Repository layout

path	contents
`models/`	SimSID and its components: the space-aware memory matrix (`memory.py`), the feature-space in-painting block (`inpaint.py`), the autoencoder backbone (`squid.py`), and the discriminator
`dataloader/`	Loaders for ZhangLab (`dataloader_zhang.py`), CheXpert (`dataloader_chexpert.py`), and COVIDx (`dataloader_covidx.py`)
`configs/`	Experiment configs, all inheriting `configs/base.py`
`main.py`	Train SimSID
`eval.py`	Evaluate a trained SimSID checkpoint

Citation

If SimSID is useful in your research, please cite the TPAMI paper. If you use the CVPR version, please also cite SQUID.

@article{xiang2024exploiting,
  title={Exploiting Structural Consistency of Chest Anatomy for Unsupervised Anomaly Detection in Radiography Images},
  author={Xiang, Tiange and Zhang, Yixiao and Lu, Yongyi and Yuille, Alan L. and Zhang, Chaoyi and Cai, Weidong and Zhou, Zongwei},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume={46},
  number={9},
  pages={6070--6081},
  year={2024},
  doi={10.1109/TPAMI.2024.3382009},
  url={https://github.com/MrGiovanni/SimSID}
}

@inproceedings{xiang2023squid,
  title={SQUID: Deep Feature In-Painting for Unsupervised Anomaly Detection},
  author={Xiang, Tiange and Zhang, Yixiao and Lu, Yongyi and Yuille, Alan L. and Zhang, Chaoyi and Cai, Weidong and Zhou, Zongwei},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={23890--23901},
  year={2023},
  url={https://github.com/tiangexiang/SQUID}
}