Transducer Adaptive Ultrasound Volume Reconstruction

Hengtao Guo, Sheng Xu, Bradford J. Wood, Pingkun Yan

TL;DR

This paper introduces a novel domain adaptation strategy for "sensorless" 3D ultrasound reconstruction, allowing deep learning models to generalize across different ultrasound transducers. By using a paired-sampling strategy that aligns feature distributions based on motion trajectories, the authors successfully reconstruct 3D volumes from 2D scans without the need for external tracking hardware, even when switching between transrectal and transabdominal probes.

What Problem Does This Paper Address?

3D ultrasound (US) provides significantly more clinical context than 2D slices, which is vital for procedures like prostate biopsies. Traditionally, creating a 3D volume from "freehand" 2D scans requires bulky external tracking devices (optical or electromagnetic) to record the probe's position in space.

"Sensorless" reconstruction uses deep learning to estimate the motion between 2D frames directly from the image data. However, these models suffer from a major "domain shift" problem: a network trained on one specific transducer (e.g., a transrectal probe) performs poorly when applied to another (e.g., a transabdominal probe) due to differences in image properties and scanning trajectories. This paper aims to make these models "transducer-adaptive."

Methods

The authors propose a three-step method called TAUVR (Transducer Adaptive Ultrasound Volume Reconstruction):

  1. Step A (Source Training): A convolutional feature extractor ($G_s$) and a 6-degree-of-freedom (DOF) regressor ($R_s$) are trained on a large labeled source dataset (transrectal scans).
  2. Step B (Domain Adaptation): This is the core innovation. To adapt the model to a new target domain (transabdominal scans), the authors use a paired-sampling strategy. They create a "pool" of source subsequences. For every target sample, they find a source sample with a matching motion trajectory (DOF). They then train a new target feature extractor ($G_t$) by minimizing the $L_2$ distance (discrepancy loss) between the feature vectors of these pairs.
  3. Step C (Inference): The adapted feature extractor ($G_t$) is combined with the original regressor ($R_s$) to reconstruct 3D volumes from the new transducer's data.

Crucially, they also implement Transformation Space Alignment as a pre-processing step to roughly align the 3D trajectories of different scans, preventing "label bias" from impairing the network.

Datasets

The study utilized clinical data from the National Institutes of Health (NIH):

  • Source Domain: 640 transrectal US video sequences (captured with a C95 transducer). These included ground-truth position data from EM-tracking.
  • Target Domain: 12 transabdominal US video sequences (captured with a C51 transducer). 9 cases were used for adaptation training, and 3 were reserved for testing.

Key Insights and Ideas

  • Adversarial vs. Discrepancy-based Adaptation: The authors argue that standard adversarial adaptation (like ADDA), which tries to "fool" a discriminator into thinking source and target features are the same, is too blunt for regression tasks. It often destroys the task-specific information (the motion signal) while trying to merge distributions.
  • Preserving the "Latent Gradient": By pairing samples with similar motion, the model learns to ignore the appearance of the transducer while remaining sensitive to the motion encoded in the speckle patterns.
  • Weakly-Supervised Learning: Even though the target domain has very few labeled samples, the model uses those labels effectively to "anchor" the target features to the well-learned source feature space.
  • Visualization via t-SNE: The authors used t-SNE to show that while other methods either failed to merge the domains or lost the motion pattern, TAUVR successfully merged the domains while maintaining a smooth "color gradient" representing the dominant motion direction.

Key Results

The TAUVR method outperformed both the "Source-only" model and standard adversarial adaptation (ADDA) on the target domain:

  • Mean Distance Error: TAUVR achieved 12.67 mm, compared to 15.73 mm for the source model and 14.88 mm for ADDA.
  • Final Drift: TAUVR reduced the final drift to 20.34 mm, down from 26.73 mm in the source-only model.
  • Qualitative Accuracy: Visual inspections of the reconstructed sagittal views showed that TAUVR produced volumes significantly closer to the ground truth than the source-only model, which tended to overfit to the source's specific scanning trajectory.

Limitations and Future Work

The primary limitation is the small size of the target dataset (only 12 sequences). While the results are promising, the authors acknowledge that a more comprehensive evaluation with larger and more diverse datasets is required to prove universal applicability. Future work will likely focus on testing the method across a wider variety of clinical applications and transducer types.

Summary updated

Abstract

Reconstructed 3D ultrasound volume provides more context information compared to a sequence of 2D scanning frames, which is desirable for various clinical applications such as ultrasound-guided prostate biopsy. Nevertheless, 3D volume reconstruction from freehand 2D scans is a very challenging problem, especially without the use of external tracking devices. Recent deep learning based methods demonstrate the potential of directly estimating inter-frame motion between consecutive ultrasound frames. However, such algorithms are specific to particular transducers and scanning trajectories associated with the training data, which may not be generalized to other image acquisition settings. In this paper, we tackle the data acquisition difference as a domain shift problem and propose a novel domain adaptation strategy to adapt deep learning algorithms to data acquired with different transducers. Specifically, feature extractors that generate transducer-invariant features from different datasets are trained by minimizing the discrepancy between deep features of paired samples in a latent space. Our results show that the proposed domain adaptation method can successfully align different feature distributions while preserving the transducer-specific information for universal freehand ultrasound volume reconstruction.
BibTeX
@inproceedings{guo2021transducer,
  author = {Hengtao Guo and Sheng Xu and Bradford J. Wood and Pingkun Yan},
  title = {{Transducer Adaptive Ultrasound Volume Reconstruction}},
  booktitle = {ISBI},
  year = {2021},
  doi = {10.1109/isbi48211.2021.9433756},
}
Figure