Related Experiment Videos
Unsupervised multi-modal variational autoencoder for anomaly detection in coronary angiography
Emanuele Frassini1, Rick M Butler1, Benno H W Hendriks1,2
1Delft University of Technology, Delft, The Netherlands.
Summary
This study introduces a new unsupervised method using multi-modal data to detect workflow disruptions during coronary angiography (CAG). The approach effectively identifies anomalies, improving patient safety in catheterization labs.
Area of Science:
- Cardiology
- Medical Imaging
- Artificial Intelligence
Background:
- Coronary angiography (CAG) is a common cardiac procedure.
- Vessel tortuosity can lead to workflow disruptions like catheter exchanges, increasing risks.
- Automated detection of these subtle anomalies is challenging with single data sources.
Purpose of the Study:
- To develop an unsupervised, multi-modal anomaly detection framework for CAG procedures.
- To improve the identification of workflow disruptions and potential complications.
Main Methods:
- A variational autoencoder (VAE) framework was developed using video and pose data from 191 CAG procedures.
- Data was segmented into 10-second clips, with features extracted using 3D CNNs and fused via modality-specific LSTM encoders.
- Reconstruction error served as the anomaly score, with thresholds calibrated on a separate dataset.
Main Results:
- Anomalies, such as extra catheter exchanges, occurred in 29.8% of procedures, increasing mean duration by 34%.
- The multi-modal VAE achieved superior performance over unimodal models, with an F1-score of 0.83, AUC-ROC of 0.88, and AUC-PR of 0.90.
- Detection accuracy was highest during periods with the most frequent anomalies.
Conclusions:
- Multi-modal fusion is effective for unsupervised anomaly detection in complex surgical workflows like CAG.
- This technology can support the development of intelligent monitoring systems for catheterization laboratories.
- Improved anomaly detection can enhance patient safety and procedural efficiency.