Shadow Reduction in Ultrasound Imaging using Radiance Field Decomposition

Abstract

Acoustic shadows can hide clinically relevant structures in ultrasound images, especially in fetal brain scans where the skull attenuates the signal before it reaches deeper anatomy. RFlash is a physics-informed post-processing method that reduces these shadows without requiring scanner hardware access. It decomposes an ultrasound image into attenuation and scatter maps, then uses a differentiable ultrasound simulator to render a shadow-reduced image.

Across fetal brain and abdominal ultrasound data, RFlash improves visual contrast consistency and reduces shadow-related intensity differences compared with the original images and a classical Hughes baseline. The same decomposition also produces shadow confidence masks that can support downstream shadow segmentation.

Overview

Ultrasound is portable, low risk, and widely used, but acoustic shadows can obscure anatomy behind highly attenuating structures such as bone. In fetal brain imaging this can make the proximal hemisphere harder to assess, limiting the visibility of bilateral structures that are clinically important.

RFlash treats shadow reduction as an image-formation problem. Instead of learning a direct image-to-image correction, it estimates physically meaningful maps that explain how ultrasound energy is attenuated and scattered through tissue. A shadow-reduced rendering can then be interpreted as virtually moving the probe through the tissue so that each depth depends less on structures closer to the transducer.

Methods

RFlash optimises explicit attenuation and scatter-intensity maps so that a differentiable ultrasound simulator can reproduce the input image. After this decomposition, the simulator renders a second image with the shadow-forming attenuation effect reduced. This keeps the method tied to ultrasound physics while remaining usable as post-processing on already beamformed images.

Results

Fetal brain: Original vs RFlash

RFlash

Original

Fetal brain: Hughes vs RFlash

RFlash

Hughes

Abdominal ultrasound: Original vs RFlash

RFlash

Original

Abdominal ultrasound: Hughes vs RFlash

RFlash

Hughes

Drag each slider to compare the original or Hughes-corrected image with RFlash.

Age Prediction

To test if the recovered structures from the proximal hemisphere are clinically useful, we trained a deep learning model to predict gestational age from the proximal hemisphere on normal ultrasound data and shadow reduced data. We tested on both hemisphere and found that shadow-reduction improves predictability, with the largest improvements of the predictability of the proximal hemisphere.

Shadow Masks

The sliders below compare shadow mask from Meng et al. overlaid with the confidence masks derived from RFlash.

Shadow mask example 04

Ours

Quinn

Shadow mask example 07

Ours

Quinn

Shadow confidence overlays from Quinn compared with RFlash.

Cite

@misc{bacher2026shadow,
      title={Shadow Reduction in Ultrasound Imaging Using Differentiable Simulation and Radiance Field Decomposition}, 
      author={Valentin Bacher and Pak Hei Yeung and Bernhard Kainz and Madeleine K. Wyburd and Nicola K. Dinsdale and Michael Gray and Ana I. L. Namburete},
      year={2026},
      month={9},
      eprint={2609.29373},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.29373}, 
}