purdueviperlab/Surgical_Instruments_Unsupervised_Defect_Detection

Dataset Card for Surgical Instruments Unsupervised Defect Detection

This dataset is designed for unsupervised anomaly/defect detection of surgical instruments. It contains images of defect-free and defective surgical instruments and is intended to support research on visual anomaly detection, defect localization, and automated inspection of surgical instruments.

Dataset Details

Dataset Description

This dataset is introduced in the paper "Unsupervised Defect Detection for Surgical Instruments" by Joseph Huang, Yichi Zhang, Xiaoyu Ji, Jingxi Yu, Wei Chen, Seunghyun Hwang, Qiang Qiu, Amy R. Reibman, Edward J. Delp, and Fengqing Zhu.

The dataset is designed for research on unsupervised defect detection and anomaly localization in surgical instrument images. It contains defect-free surgical instrument images for training and defective instrument images for evaluation.

The dataset was constructed to study the challenges of applying unsupervised anomaly detection methods to surgical instruments, including background interference, subtle defects, and domain shift from natural images to surgical instrument imagery.

• Curated by: Joseph Huang, Yichi Zhang, Xiaoyu Ji, Jingxi Yu, Wei Chen, Carter Gibbs, Qiang Qiu, Amy R. Reibman, Edward J. Delp, Fengqing Zhu

• Funded by: Indiana University Health and the Office of Research at Purdue University

• License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC-BY-NC-ND-4.0)

Dataset Sources

• Paper: https://www.doi.org/10.1109/ICIP61757.2026.11630477

Uses

Direct Use

The dataset is intended for research and development of Anomaly/Defect detection. It can be used to train or evaluate methods that learn the distribution of defect-free surgical instruments and identify potential defects.

Out-of-Scope Use

The dataset is not intended for direct clinical deployment, medical diagnosis, or autonomous decisions regarding the safety or sterility of surgical instruments.

The dataset should never be used for:

• Determining whether a surgical instrument is safe for clinical use.
• Determining whether an instrument is sterile or free from biological contamination.
• Making clinical decisions.
• Replacing qualified personnel responsible for surgical instrument inspection and quality assurance.
• Training or evaluating systems without considering the limitations of simulated defects and controlled image acquisition conditions.

The dataset is intended for research and benchmarking purposes. Performance on this dataset should not be interpreted as evidence of clinical effectiveness or readiness for deployment.

Dataset Structure

The dataset contains:

• 20 classes of surgical instruments
• 340 defect-free images
• 300 defective images used for evaluation

Defect categories including:

• Blood
• Pores
• Cuts
• Corrosion
• Cracks

Blood defects are created using dyes/food coloring, physically simulating biohazard. All the other defect categories are simulated defects generated using TF-IDG (Training-Free Industrial Defect Generation).

The directory structure is:

instrument1/
|-- train/
|   |- good/...
|-- test/
    |- good/...
    |- defect_type1/...
    |- defect_type2/...
    |-...
|- ground_truth/
    |- defect_type1/...
    |- defect_type2/...
    |-...
instrument2/
|- ...

Dataset Creation

Curation Rationale

Existing unsupervised anomaly detection methods are commonly evaluated on industrial datasets. However, surgical instruments present domain-specific challenges, including metallic surfaces, complex shapes, textured backgrounds, reflective regions, and small defects.

The dataset was created to provide a domain-specific testbed for studying unsupervised defect detection in surgical instrument imagery.

Source Data

Data Collection and Processing

The surgical instrument images were acquired under controlled imaging conditions.

According to the paper, images were captured inside a lightbox using an iPhone 16 Pro with the ProCamera application.

The reported acquisition settings include:

• Resolution: 3024 * 4032 pixels
• Focal length: 24 mm
• Aperture: f/1.78
• ISO: 50
• Shutter speed: 1/1005 s
• Exposure compensation: 0 EV
• White balance: 5600 K
• Tint: 0
• Autofocus: Enabled

The dataset contains images from 20 surgical instrument classes.

Bias, Risks, and Limitations

Several limitations should be considered when using this dataset:

• Simulated defects.
  The defective samples include physically simulated blood-like defects and defects generated using TF-IDG. These defects may not fully represent the appearance, geometry, and variability of naturally occurring defects.
• Controlled imaging conditions.
  Images were captured using a controlled imaging setup. Performance may change under different cameras, illumination conditions, viewing angles, backgrounds, or acquisition environments.
• Limited instrument diversity.
  The dataset contains 20 surgical instrument classes and may not represent the full diversity of surgical instruments used in hospitals and clinical environments.

Recommendations

Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.

Citation

BibTeX:

If you use this dataset, please cite:

@article{huang2026unsupervised,
  title={Unsupervised Defect Detection for Surgical Instruments},
  author={Huang, Joseph and Zhang, Yichi and Ji, Xiaoyu and Yu, Jingxi and Chen, Wei and Hwang, Seunghyun and Qiu, Qiang and Reibman, Amy R. and Delp, Edward J. and Zhu, Fengqing},
  journal={2026 IEEE International Conference on Image Processing (ICIP)},
  year={2026},
  pages={1-6},
  doi={10.1109/ICIP61757.2026.11630477}
}

More Information

Here is the citation for the TF-IDG paper we used for synthetic data generation:

@article{Xu2025defectgeneration,
  author={Xu, Ruyi and Chiu, Yen-Tzu and Chen, Tai-I and Chew, Oscar and Chuang, Yung-Yu and Cheng, Wen-Huang},
  journal={2025 IEEE/CVF International Conference on Computer Vision (ICCV)}, 
  title={Training-Free Industrial Defect Generation with Diffusion Models}, 
  year={2025},
  pages={24214-24223},
  doi={10.1109/ICCV51701.2025.02244}}

Dataset Card Contact

Fengqing Zhu (zhu0@purdue.edu)