Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RapidVol: Rapid Reconstruction of 3D Ultrasound Volumes from Sensorless 2D Scans

Mark C. Eid, Pak-Hei Yeung, Madeleine K. Wyburd, João F. Henriques, Ana I. L. Namburete

TL;DR

RapidVol introduces a hybrid implicit–explicit framework for rapid 3D ultrasound reconstruction from sensorless 2D scans by compressing the 4D cost volume with Tri-Planar or CP tensor decompositions and decoding with a small MLP. The method renders high-fidelity 3D fetal-brain views from 2D inputs with known poses, optimized using a SSIM-based loss. Empirically, RapidVol is up to 3× faster per epoch and up to 46% more accurate than prior ImplicitVol implementations, with added robustness to pose errors and the option to accelerate further using an atlas prior. This approach enables quicker, more accessible 3D ultrasound reconstructions in clinical settings using standard 2D probes.

Abstract

Two-dimensional (2D) freehand ultrasonography is one of the most commonly used medical imaging modalities, particularly in obstetrics and gynaecology. However, it only captures 2D cross-sectional views of inherently 3D anatomies, losing valuable contextual information. As an alternative to requiring costly and complex 3D ultrasound scanners, 3D volumes can be constructed from 2D scans using machine learning. However this usually requires long computational time. Here, we propose RapidVol: a neural representation framework to speed up slice-to-volume ultrasound reconstruction. We use tensor-rank decomposition, to decompose the typical 3D volume into sets of tri-planes, and store those instead, as well as a small neural network. A set of 2D ultrasound scans, with their ground truth (or estimated) 3D position and orientation (pose) is all that is required to form a complete 3D reconstruction. Reconstructions are formed from real fetal brain scans, and then evaluated by requesting novel cross-sectional views. When compared to prior approaches based on fully implicit representation (e.g. neural radiance fields), our method is over 3x quicker, 46% more accurate, and if given inaccurate poses is more robust. Further speed-up is also possible by reconstructing from a structural prior rather than from scratch.

RapidVol: Rapid Reconstruction of 3D Ultrasound Volumes from Sensorless 2D Scans

TL;DR

RapidVol introduces a hybrid implicit–explicit framework for rapid 3D ultrasound reconstruction from sensorless 2D scans by compressing the 4D cost volume with Tri-Planar or CP tensor decompositions and decoding with a small MLP. The method renders high-fidelity 3D fetal-brain views from 2D inputs with known poses, optimized using a SSIM-based loss. Empirically, RapidVol is up to 3× faster per epoch and up to 46% more accurate than prior ImplicitVol implementations, with added robustness to pose errors and the option to accelerate further using an atlas prior. This approach enables quicker, more accessible 3D ultrasound reconstructions in clinical settings using standard 2D probes.

Abstract

Two-dimensional (2D) freehand ultrasonography is one of the most commonly used medical imaging modalities, particularly in obstetrics and gynaecology. However, it only captures 2D cross-sectional views of inherently 3D anatomies, losing valuable contextual information. As an alternative to requiring costly and complex 3D ultrasound scanners, 3D volumes can be constructed from 2D scans using machine learning. However this usually requires long computational time. Here, we propose RapidVol: a neural representation framework to speed up slice-to-volume ultrasound reconstruction. We use tensor-rank decomposition, to decompose the typical 3D volume into sets of tri-planes, and store those instead, as well as a small neural network. A set of 2D ultrasound scans, with their ground truth (or estimated) 3D position and orientation (pose) is all that is required to form a complete 3D reconstruction. Reconstructions are formed from real fetal brain scans, and then evaluated by requesting novel cross-sectional views. When compared to prior approaches based on fully implicit representation (e.g. neural radiance fields), our method is over 3x quicker, 46% more accurate, and if given inaccurate poses is more robust. Further speed-up is also possible by reconstructing from a structural prior rather than from scratch.
Paper Structure (13 sections, 3 equations, 3 figures, 3 tables)

This paper contains 13 sections, 3 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Methodology
  4. Problem Setup
  5. Tensor Decompositions
  6. CP Decomposition:
  7. Pipeline of RapidVol
  8. Experimental Setup
  9. Technical Details
  10. Experiments Performed
  11. Results and Discussion
  12. Conclusion
  13. Acknowledgements.

Figures (3)

  • Figure 1: Diagram showing the two different types of tensor decomposition employed, and how the value of a voxel in $\mathcal{T}$ can be retrieved.
  • Figure 2: Pipeline of our proposed method RapidVol. During the reconstruction process, a set of $\Pi$ images and their corresponding poses $\Lambda$ are required as input. Once the brain is reconstructed, only the pose $\Lambda$ at which one wishes to see is required as input, and parameters are not updated. Nb. Ultrasound probe image adapted from Flaticon.com.
  • Figure 3: Testing accuracy profile curves. Training was done to 5,000 epochs, accuracy reported every 250 epochs. For consistent timings, these runs were done on an isolated but much slower GPU. $\Pi$ = 160 axial slices, Testing dataset = 160 coronal slices rotated $360^\circ$ about the Vertical Axis.