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.
Introduction
3D ultrasound (US) imaging offers significantly more clinical context than traditional 2D slices, which is critical for precision procedures such as prostate biopsies. However, generating a 3D volume from "freehand" 2D scans typically requires external tracking hardware—such as electromagnetic or optical sensors—to record the probe's spatial coordinates. These devices are often bulky, expensive, and complicate the clinical workflow.
"Sensorless" reconstruction has emerged as a promising alternative, utilizing deep learning to estimate inter-frame motion directly from image data. Despite its potential, these models face a significant "domain shift" challenge. A network trained on a specific transducer (e.g., a transrectal probe) often fails when applied to a different one (e.g., a transabdominal probe) due to variations in image properties, speckle patterns, and scanning trajectories. This research introduces a transducer-adaptive approach to overcome these generalization barriers.
Methodology
The proposed framework, TAUVR (Transducer Adaptive Ultrasound Volume Reconstruction), employs a three-stage process to achieve cross-transducer generalization:
- Source Training: A convolutional feature extractor ($G_s$) and a 6-degree-of-freedom (DOF) regressor ($R_s$) are first trained on a large, labeled source dataset consisting of transrectal scans with known ground-truth motion.
- Domain Adaptation via Paired-Sampling: To adapt to a target domain (e.g., transabdominal scans), the authors implement a novel paired-sampling strategy. Instead of standard adversarial methods that can strip away task-relevant motion information, TAUVR identifies pairs of source and target sequences with matching motion trajectories. A target feature extractor ($G_t$) is then trained by minimizing the $L_2$ distance (discrepancy loss) between the latent features of these matched pairs.
- Inference: The adapted target feature extractor ($G_t$) is paired with the pre-trained source regressor ($R_s$) to perform 3D reconstruction on the new transducer data.
To further refine the process, the authors utilize Transformation Space Alignment as a pre-processing step to align the 3D trajectories across domains, preventing label bias from degrading network performance.
Experimental Results
The study utilized clinical data from the National Institutes of Health (NIH), comprising 640 transrectal sequences (C95 transducer) for the source domain and 12 transabdominal sequences (C51 transducer) for the target domain.
Quantitative Performance
TAUVR demonstrated superior performance compared to both the baseline "Source-only" model and standard adversarial adaptation (ADDA):
- Mean Distance Error: TAUVR achieved 12.67 mm, significantly lower than the source-only model (15.73 mm) and ADDA (14.88 mm).
- Final Drift: The final drift was reduced to 20.34 mm, a marked improvement over the 26.73 mm observed in the source-only baseline.
Feature Space Analysis
Visualization using t-SNE confirmed that TAUVR successfully aligned the source and target feature distributions. Unlike adversarial methods that often collapse the feature space, TAUVR preserved the "latent gradient"—the smooth transition of features corresponding to motion direction—ensuring that the regressor could still accurately estimate movement despite the change in image domain.
Conclusion
This work presents a robust domain adaptation strategy for sensorless 3D ultrasound reconstruction. By leveraging a paired-sampling approach that preserves motion-specific information while aligning transducer-invariant features, TAUVR enables deep learning models to generalize across different clinical scanning modalities. While the current study is limited by a small target dataset, the results provide a strong foundation for developing universal, hardware-free 3D ultrasound systems.
Summary updated