Dou, Yimeng; Li, Yin; Varghese, Tomy. (2026).听.听Biomedical Physics & Engineering Express, 12(4), 045004.听
Creating accurate聽3D ultrasound聽images from multiple聽2D freehand ultrasound聽scans is challenging because slight errors in the position or orientation of the ultrasound probe can cause the images to become misaligned, leading to distortions in the final reconstruction. Recent approaches have used聽neural radiance fields (NeRFs)鈥攁n artificial intelligence technique that learns a continuous 3D representation from 2D images鈥攂ut these methods are particularly sensitive to positioning errors. To address this problem, the researchers developed聽GAU-NeRF (Geometric Aware Ultrasound NeRF), a new approach that stabilizes the model during training and improves its ability to correct probe position errors. The method was evaluated using both simulated and real ultrasound datasets and consistently outperformed existing reconstruction techniques. Compared with previous methods, GAU-NeRF substantially improved image quality, including increases of up to 132% in聽peak signal-to-noise ratio聽and 133% in the聽structural similarity index, while reducing image reconstruction errors by up to 350% based on a perceptual image quality metric. These findings suggest that GAU-NeRF can produce more accurate and reliable 3D ultrasound reconstructions, which could improve applications that rely on freehand ultrasound imaging.
Figure 1.聽(a) Overlap of two freehand US images acquired during two different sweep sequences for a tissue-mimicking abdominal phantom, where misregistration between two sweeps occurs. (b) Reconstruction using distance weighting [], which shows incorrect 3D geometry. (c) Reconstruction result from Ultra-NeRF. (d) Reconstruction with our GAU-NeRF method, which correctly recovers the underlying 3D structure.