Ultrasharp: Beltrami Transformers for Ultrasound Super-Resolution

Ultrasound (US) super-resolution (SR) is challenged by anisotropic point spread functions (PSFs), multiplicative speckle noise, and the clinical need for edge fidelity at tissue interfaces. We propose UltraSharp, a geometry- and physics-aware transformer that injects Beltrami image geometry into attention while enforcing formation-consistent reconstruction. The proposed framework employs edge-aware Beltrami positional encodings to capture geodesic structure, anisotropic geodesic attention to limit cross-boundary feature mixing, and physics- and speckle-aware losses to preserve realistic texture statistics. Across cardiac, breast, and fetal US, UltraSharp improves edge quality and texture while maintaining fidelity. We report standard SR metrics and USspecific measures, including contrast-to-noise ratio (CNR), edge spread, speckle SNR, and downstream segmentation accuracy.