UltraNBA Neural Bundle Adjustment for Pose Refinement in 3D Freehand Ultrasound

V. Bugra Yesilkaynak, Vanessa Gonzalez Duque, Magdalena Wysocki, Mohammad Farid Azampour, Nassir Navab, Diana Mateus; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 4363-4372

Imprecise tracking presents a major challenge in the reconstruction of 3D freehand ultrasound volumes, as even small errors can lead to significant misalignment. Calibration inaccuracies and the reliance on noisy sensor data further exacerbate this issue. State-of-the-art approaches typically align pixel intensities across overlapping frames. However, changes in ultrasound propagation paths, which depend on the sensor's position, often result in inconsistent intensities for the same spatial location, challenging the reliability of these methods. To address these challenges, we propose UltraNBA, a novel implicitly neural bundle-adjusting framework for ultrasound. By leveraging the spatial consistency of acoustic tissue properties instead of raw intensity alignment, UltraNBA corrects tracking errors while capturing stable anatomical and physical representations, yielding higher-quality reconstructions. Our method supports single and multiple sweeps, offering versatility in real-world clinical scenarios. Experimental results demonstrate a reduction in tracking errors, accompanied by enhanced image quality for rendering new frames. We make code and data open-source at http://github.com/anonymus.