[Paper Review] Enabling Augmented Segmentation and Registration in Ultrasound-Guided Spinal Surgery via Realistic Ultrasound Synthesis from Diagnostic CT Volume

Ang Li, Jiayi Han+3 more|arXiv (Cornell University)|Jan 5, 2023

Medical Imaging and Analysis4 citations

TL;DR

This paper proposes a realistic ultrasound (US) simulation framework that generates synthetic bone US images from diagnostic CT volumes to address the scarcity of clinical US data and annotations for spinal surgery. Using a lightweight vision transformer with a Long-range Contrast Learning Module, the method achieves high-accuracy bone segmentation (0.599 mm Chamfer distance) and CT-US registration (0.13–3.37 mm error), enabling safe, real-time US-guided pedicle screw placement without real clinical data fine-tuning.

ABSTRACT

This paper aims to tackle the issues on unavailable or insufficient clinical US data and meaningful annotation to enable bone segmentation and registration for US-guided spinal surgery. While the US is not a standard paradigm for spinal surgery, the scarcity of intra-operative clinical US data is an insurmountable bottleneck in training a neural network. Moreover, due to the characteristics of US imaging, it is difficult to clearly annotate bone surfaces which causes the trained neural network missing its attention to the details. Hence, we propose an In silico bone US simulation framework that synthesizes realistic US images from diagnostic CT volume. Afterward, using these simulated bone US we train a lightweight vision transformer model that can achieve accurate and on-the-fly bone segmentation for spinal sonography. In the validation experiments, the realistic US simulation was conducted by deriving from diagnostic spinal CT volume to facilitate a radiation-free US-guided pedicle screw placement procedure. When it is employed for training bone segmentation task, the Chamfer distance achieves 0.599mm; when it is applied for CT-US registration, the associated bone segmentation accuracy achieves 0.93 in Dice, and the registration accuracy based on the segmented point cloud is 0.13~3.37mm in a complication-free manner. While bone US images exhibit strong echoes at the medium interface, it may enable the model indistinguishable between thin interfaces and bone surfaces by simply relying on small neighborhood information. To overcome these shortcomings, we propose to utilize a Long-range Contrast Learning Module to fully explore the Long-range Contrast between the candidates and their surrounding pixels.

Motivation & Objective

• Address the critical lack of clinical intra-operative ultrasound (US) data and accurate annotations for training deep learning models in spinal surgery.
• Overcome challenges in bone surface annotation due to US image artifacts and low contrast.
• Develop a realistic, radiation-free in silico US simulation framework to generate high-fidelity synthetic US images from diagnostic CT volumes.
• Train a lightweight vision transformer with a Long-range Contrast Learning Module for accurate, real-time bone segmentation in US-guided spinal navigation.
• Validate the simulation-based training pipeline for CT-US registration and pedicle screw placement with sub-millimeter accuracy and no complications.

Proposed method

• Generate realistic US images from diagnostic CT volumes using a physics-informed simulation pipeline incorporating tissue acoustic properties, probe geometry, and tissue deformation models.
• Simulate US probe movement and tissue compression via extrusion and moving space simulation to enhance realism and diversity of synthetic data.
• Introduce a Long-range Contrast Learning Module (LCLM) that captures long-range pixel relationships to improve localization of bone surfaces in noisy US images.
• Train a lightweight vision transformer model on the synthetic US data for end-to-end bone surface segmentation with high efficiency and accuracy.
• Perform CT-US registration using coarse alignment followed by de-noised Iterative Closest Point (ICP) on segmented point clouds to align pre-operative plans with intra-operative US.
• Validate the entire pipeline using physical phantoms (spine in water, 3D-printed spine, bovine spine in agar) with ground truth from calibrated US and CT scans.

Experimental results

Research questions

• RQ1 Can a physics-based in silico US simulation framework generate realistic, contextually accurate US images from diagnostic CT volumes without requiring real US data?
• RQ2 Can a vision transformer with a Long-range Contrast Learning Module achieve superior bone surface segmentation accuracy in low-contrast, echo-rich US images?
• RQ3 Does training on synthetic US data alone enable accurate and robust CT-US registration for spinal surgery guidance?
• RQ4 Can the simulated US data support safe and accurate pedicle screw placement with minimal targeting error?
• RQ5 How does the proposed method compare to existing data augmentation techniques in terms of realism, generalization, and clinical applicability?

Key findings

• The proposed US simulation framework successfully generates realistic bone US images from CT volumes, enabling data augmentation without real clinical data.
• The vision transformer with Long-range Contrast Learning Module achieves a Chamfer distance of 0.599 mm in bone surface segmentation, indicating high precision.
• CT-US registration using simulated data achieves an error of 0.13 to 3.37 mm, with the highest error (3.37 mm) observed on the L5 vertebra due to its complex shape.
• Pedicle screw placement based on registration shows a translation error of 0.027–4.031 mm and rotation error of 3.05–4.47 degrees, confirming clinical safety and accuracy.
• The model trained solely on simulated data outperforms traditional approaches and generalizes well across different anatomical phantoms, demonstrating robustness.
• Ablation studies confirm the effectiveness of the Long-range Contrast Learning Module in improving segmentation accuracy and localization of bone surfaces.