NeRF-US👤👤: Neural Radiance Fields for Accurate Ultrasound Imaging in the Wild

Rishit Dagli1,2, Atsuhiro Hibi2,3,4, Rahul G. Krishnan1,5, Pascal Tyrrell2,4,6
Departments of 1 Computer Science; 2 Medical Imaging, University of Toronto, Canada
3 Division of Neurosurgery, St Michael's Hospital, Unity Health Toronto, Canada
4 Institute of Medical Science; Departments of 5 Laboratory Medicine and Pathobiology; 6 Statistical Sciences, University of Toronto, Canada
Paper arXiv Code Data

NeRF-US trains a NeRF on sound fields.



NeRF-US Teaser Image
NeRF-US turns ultrasounds captured in the wild into artifact-free 3D reconstructions.
Novel view ultrasound renders from NeRF-US.


Abstract

Current methods for performing 3D reconstruction and novel view synthesis (NVS) in ultrasound imaging data often face severe artifacts when training NeRF-based approaches. The artifacts produced by current approaches differ from NeRF floaters in general scenes because of the unique nature of ultrasound capture. Furthermore, existing models fail to produce reasonable 3D reconstructions when ultrasound data is captured or obtained casually in uncontrolled environments, which is common in clinical settings. Consequently, existing reconstruction and NVS methods struggle to handle ultrasound motion, fail to capture intricate details, and cannot model transparent and reflective surfaces. In this work, we introduced NeRF-US, which incorporates 3D-geometry guidance for border probability and scattering density into NeRF training, while also utilizing ultrasound-specific rendering over traditional volume rendering. These 3D priors are learned through a diffusion model. Through experiments conducted on our new ``Ultrasound in the Wild'' dataset, we observed accurate, clinically plausible, artifact-free reconstructions.

Motivation

There are a few challenges common when using previous medical NeRF methods and standard methods: the need for high-quality, diverse datasets, capturing intricate details like tissue interface locations critical for medical diagnosis, accurately modeling transparent and reflective surfaces. There are quite a few NeRF artifacts that appear when using these methods in the wild. In contrast to this, our approach (as shown) produces artifact-free reconstructions with minor details accurately reconstructed.
Showcasing boundaries.

How does NeRF-US work?

Our goal is to produce a 3D representation given a set of ultrasound images taken in the wild and their camera positions. The first step of our approach relies on the training of a 3D diffusion model, which can serve as geometric priors for our NeRF model. This diffusion model produces an 32 x 32 x 32 occupancy grid. To create this diffusion model, we finetune the 3D diffusion model on a small dataset of voxels around the human knee generated synthetically.

How to create diffusion model?

Figure: An overview of how our diffusion model is fine-tuned, we use 323-sized patches to LoRA-finetune a 3D diffusion model trained on ShapeNet.

We now train our NeRF model that takes in a 3D vector (denoting positions in 3D) and learns a 5D vector (attenuation, reflectance, border probability, scattering density, and scattering intensity). While training this NeRF, we run the outputs through the diffusion model and obtain guidance vectors for border probability and scattering density. These are added to the photometric loss. We finally train the NeRF with this final loss we calculated.
How to train our model?

Figure: An overview of how our method works. We train a NeRF model that uses ultrasound rendering to convert the representations into a 2D image after which we infer through a 3D diffusion model which has geometry priors through which we calculate a modified loss definition to train the NeRF.

Visual Results

Ours
Nerfacto [Tancik 2023]
Ours
Nerfacto [Tancik 2023]
Ours
Nerfacto [Tancik 2023]
Ours
Ultra-NeRF [Wysocki 2024]
Ours
Ultra-NeRF [Wysocki 2024]
Ours
Ultra-NeRF [Wysocki 2024]

Figure: Qualitative Results. We demonstrate the results of our method and compare it qualitatively with Nerfacto [1], Gaussian Splatting [3], and Ultra-NeRF [2]. Our approach, NeRF-US, produces accurate and high-quality reconstructions as compared to the baseline models on novel views (best viewed with zoom).

Depth comparisons

Figure: Qualitative Results. We demonstrate the results of depth maps produced from our method and compare them qualitatively with Nerfacto [1], Gaussian Splatting [3], and Ultra-NeRF [2] (best viewed in color and with zoom).

Ultrasound in the wild Dataset

Here we show some instances of our new ultrasound in the wild dataset, we limit the visualizations of the dataset to the first 10 seconds of some of the scenes in our dataset. For visualization, we pre-process these videos with a script.

Examples from our ultrasound in the wild dataset.
The following works were mentioned on this page:

[1] Tancik, Matthew, et al. "Nerfstudio: A modular framework for neural radiance field development." ACM SIGGRAPH 2023 Conference Proceedings. 2023.

[2] Wysocki, Magdalena, et al. "Ultra-nerf: neural radiance fields for ultrasound imaging." Medical Imaging with Deep Learning. PMLR, 2024.

[3] Kerbl, Bernhard, et al. "3d gaussian splatting for real-time radiance field rendering." ACM Transactions on Graphics 42.4 (2023): 1-14.

Citation

@misc{dagli2024nerfusremovingultrasoundimaging,
      title={NeRF-US: Removing Ultrasound Imaging Artifacts from Neural Radiance Fields in the Wild}, 
      author={Rishit Dagli and Atsuhiro Hibi and Rahul G. Krishnan and Pascal N. Tyrrell},
      year={2024},
      eprint={2408.10258},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.10258}, 
}

Acknowledgements

This research was enabled in part by support provided by the Digital Research Alliance of Canada. This research was supported in part with Cloud TPUs from Google's TPU Research Cloud (TRC). The resources used to prepare this research were provided, in part, by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute.