UltraSoundNeRF: Sonographic neural reflection field for novel view synthesis

Area of Science:

• Medical Imaging
• Computational Imaging
• Acoustic Physics

Background:

• Current novel view synthesis methods excel in natural scenes but struggle with ultrasound imaging's semantic accuracy and physical plausibility.
• Previous Ultra-NeRF method achieved visual realism but lacked interpretability in acoustic parameters.
• This limits practical utility and in-depth analysis of ultrasound's acoustic properties.

Purpose of the Study:

• To enhance the physical accuracy and interpretability of neural fields for ultrasound image reconstruction.
• To develop a method that ensures the physical accuracy of the underlying neural field, moving beyond visual plausibility.
• To improve the practical utility and analytical capabilities of neural fields in ultrasound imaging.

Main Methods:

• Revisiting neural fields for ultrasound by introducing the Sonographic Neural Reflection Field (UltraSoundNeRF or USNeRF).
• Redesigning the differentiable forward synthesis model from Ultra-NeRF.
• Incorporating physics-inspired regularization derived from ultrasound imaging principles.

Main Results:

• Extended the Ultra-NeRF dataset with patient lower leg data, an ex-vivo phantom, and a calibration phantom.
• Demonstrated reconstruction of real biological tissue using the ex-vivo phantom.
• Showcased improved attenuation estimates reflecting physical values with regularization on the calibration phantom.

Conclusions:

• UltraSoundNeRF (USNeRF) significantly enhances the interpretability of acoustic properties compared to prior methods.
• Reconstruction accuracy is maintained while improving the physical fidelity of the neural field.
• The method offers greater practical utility for in-depth analysis of ultrasound imaging data.