Unsupervised multi-modal variational autoencoder for anomaly detection in coronary angiography

Journal Article (2026)
Author(s)

Emanuele Frassini (TU Delft - Mechanical Engineering)

Rick M. Butler (TU Delft - Mechanical Engineering)

Benno H.W. Hendriks (TU Delft - Mechanical Engineering, Philips Healthcare Nederland)

Maarten van der Elst (TU Delft - Mechanical Engineering, Reinier de Graaf Gasthuis)

John J. van den Dobbelsteen (TU Delft - Mechanical Engineering)

Research Group
Medical Instruments & Bio-Inspired Technology
DOI related publication
https://doi.org/10.1080/13645706.2026.2705585 Final published version
More Info
expand_more
Publication Year
2026
Language
English
Research Group
Medical Instruments & Bio-Inspired Technology
Journal title
Minimally Invasive Therapy and Allied Technologies
Page Views
73
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract

Background
Coronary angiography (CAG) is among the most common procedures in catheterization laboratories. Patient-specific vessel tortuosity may require additional catheter exchanges, increasing procedure duration and complication risk. Detecting such workflow disruptions automatically is challenging because anomalies are subtle and difficult to capture using a single data modality.

Methods
We propose a novel unsupervised multi-modal anomaly detection framework based on a variational autoencoder (VAE). Video and pose data from 191 real-world CAG procedures were segmented into 10-second clips. Features extracted with 3D convolutional networks were fused through modality-specific LSTM encoders. Reconstruction error (mean squared error) was used as the anomaly score, and thresholds were calibrated on a held-out dataset.

Results
Procedures requiring additional catheter exchanges accounted for 29.8% of the dataset and had a 34% longer mean duration than normal procedures (47.6 vs. 35.5 min). The multi-modal VAE outperformed unimodal models, achieving an F1-score of 0.83, AUC-ROC of 0.88, and AUC-PR of 0.90. Detection performance peaked during the period when anomalies most frequently occurred.

Conclusions
This study demonstrates the effectiveness of multi-modal fusion for unsupervised anomaly detection in complex surgical workflows. These findings support the development of intelligent workflow monitoring systems for catheterization laboratories.