Search papers, labs, and topics across Lattice.
This paper introduces an unsupervised anomaly detection framework, Panda, specifically designed for real-time pelvic MRI, addressing the challenges posed by physiological motion and tissue deformation. By utilizing a frozen DINOv3 Vision Transformer encoder and a noisy MLP bottleneck, the method learns normative representations from healthy cases to flag deviations without the need for labeled data. The framework achieves a pixel-level AUROC of 88.06% and high specificity of 95.45%, enabling immediate feedback for radiologists during imaging procedures.
Unsupervised anomaly detection in pelvic MRI can now achieve real-time performance, offering radiologists immediate feedback during procedures.
Female pelvic diseases remain an under researched area characterized by often delayed diagnosis. While pelvic MRI offers superior soft-tissue contrast for diagnosis and image-guided procedures, real-time anomaly detection remains challenging due to physiological motion, tissue deformation, and instrument artifacts. Existing supervised approaches are impractical, as adverse events are rare, heterogeneous, and difficult to annotate. We present a Dinomaly-based unsupervised anomaly detection framework adapted for pelvic MRI that learns normative representations from healthy cases and flags deviations without requiring labels. Our approach leverages a frozen DINOv3 Vision Transformer encoder combined with a noisy MLP bottleneck and Linear Attention decoder to prevent identity mapping while maintaining computational efficiency. Anomalies are localized via per-token cosine distance between encoder and decoder representations, yielding spatial anomaly maps that provide immediate feedback at the scanner to support radiologist decision-making and adaptive protocol adjustment. Evaluated on a curated subset of the Uterine Myoma Dataset, the framework achieves a pixel-level AUROC of 88.06% and high specificity (95.45%) at frame level at 40.5 slices/s, meeting real-time clinical deployment requirements. The spatial anomaly maps and frame-level scores provide immediate, localized feedback at the scanner to support radiologist decision-making and adaptive protocol adjustment during active procedures.
