Diffusion-Inspired Anchor Conditioned Noisy Vectors Segmentation For Needle Segmentation in MR Images

Accurate segmentation of needle-like structures is essential for precise localization of the needle tip in real-time Magnetic Resonance Imaging (MRI) guidance procedures. Various post-processing techniques has been employed to calculate needle axis orientation and tip positions. However, needle segmentation in real-time MR images presents significant challenges, such as low image signal-to-noise ratio (SNR), variable needle-induced signal void features, and the limited availability of clinical datasets. Recently, diffusion models have attracted large attention for their superior model convergence and sample quality as well as their inherent tolerance and affinity for noise. Inspired by these advancements, we propose an unsupervised anomaly detection (UAD) approach, where only healthy samples are required to model the normal distribution, and signal void features are treated as anomalies. Specifically, we present a self-supervised anomaly segmentation method that incorporates edge-gradient-based noisy anomaly synthesis and a norm-guided anchor condition module to reduce input variance. Comparative experiments with representative UAD methods demonstrate that our approach is fully automated, noise-robust, and shape-insensitive, simultaneously achieving Dice scores of 0.89 (needle) and 0.47 (tumor) under simulation, and highlighting its potential to streamline clinical workflows in brain biopsy procedures.