Unsupervised anomaly detection based on diffusion models for needle detection in MR-guided interventions

Motivation: Rapid and robust needle detection are crucial for real-time interventional MRI, but it is easily disturbed by low image signal-to-noise ratio (SNR) and diverse needle void features Goal(s): To develop a rapid needle detection approach in MR-guided interventions, which is noise-robust and insensitive to the diverse needle void features Approach: We present an unsupervised anomaly detection scheme based on diffusion model. Moreover, we dynamically select informative features from a pre-trained VAE to enhance representation. Results: Our method was evaluated on a simulated human brain intervention dataset, and the results showed effective detection and recognition of different noise levels and various needle-like features. Impact: The proposed method is noise-robust and insensitive to shape characteristics, has the potential to be applied in real-time interventional magnetic resonance imaging (i-MRI)