ImplicitCell: Resolution Cell Modeling of Joint Implicit Volume Reconstruction and Pose Refinement in Freehand 3D Ultrasound

TL;DR

ImplicitCell tackles the challenge of artifacts in freehand 3D ultrasound caused by noisy pose signals by marrying a physics-based ultrasound resolution cell model with an implicit neural representation. The framework jointly optimizes volume reconstruction and pose refinement, leveraging Monte Carlo integration over subcells and a carefully regularized training strategy to mitigate drift. Across phantom, CCA, and CA datasets, ImplicitCell significantly reduces reconstruction artifacts and improves geometric and plaque morphology fidelity compared to baselines, demonstrating strong potential for clinically reliable 3DUS on cost-efficient systems. The approach advances freehand 3DUS by embedding ultrasound physics into INR-based reconstruction, enabling more robust, high-quality imaging in scenarios with noisy EM tracking.

Abstract

Freehand 3D ultrasound enables volumetric imaging by tracking a conventional ultrasound probe during freehand scanning, offering enriched spatial information that improves clinical diagnosis. However, the quality of reconstructed volumes is often compromised by tracking system noise and irregular probe movements, leading to artifacts in the final reconstruction. To address these challenges, we propose ImplicitCell, a novel framework that integrates Implicit Neural Representation (INR) with an ultrasound resolution cell model for joint optimization of volume reconstruction and pose refinement. Three distinct datasets are used for comprehensive validation, including phantom, common carotid artery, and carotid atherosclerosis. Experimental results demonstrate that ImplicitCell significantly reduces reconstruction artifacts and improves volume quality compared to existing methods, particularly in challenging scenarios with noisy tracking data. These improvements enhance the clinical utility of freehand 3D ultrasound by providing more reliable and precise diagnostic information.

Figures (9)

• Figure 1: A visual overview of the ImplicitCell framework and its key components. A). Two pixels on adjacent frames and their corresponding US resolution cells. Each resolution cell is devided into subcells in our resolution cell model. B). The overview of our proposed method. The trainable parameters are highlighted in red.
• Figure 2: Front view and top view of Lego phantom.
• Figure 3: Visual comparison of reconstruction quality across different methods under "Light Noise" (top row) and "Heavy Noise" (bottom row) conditions on phantom dataset. For each method, the top image shows the 3D rendering of the reconstructed volume, and the bottom image displays a longitudinal slice whose position is illustrated in the 3D view (green wireframe indicating the longitudinal slice position). The red dashed box highlights the differences. Numerical values below the slices indicate the LFE metric (mm, lower is better).
• Figure 4: Quantitative comparison of results on phantom datasets.
• Figure 5: Visual comparison of artery reconstruction quality across different methods and MRI reference on CCA dataset. For each method, the top row shows a 3D rendering of the artery segmentation. The middle and bottom rows display sagittal and coronal slices of the reconstructed volumes, respectively, with the artery segmentation overlaid in green. The red dashed boxes in the sagittal slices magnify the detail of artery wall, and the magnifying MRI image is enhanced for better visualization. Numerical values below each method represent "(CLD/GC)" metrics. The lower value indicates the better result.