Modeling and visualizing vertebral anatomy from freehand 3D ultrasound

Freehand 3D Ultrasound (FH 3D US) is emerging as a viable, non-invasive alternative to X-ray imaging for assessing spinal deformities like Adolescent Idiopathic Scoliosis (AIS), which is the most common form of scoliosis in adolescents and is characterized by curvatures exceeding 10 degrees. Unlike X-rays, which rely on ionizing radiation and pose risks with frequent use, FH 3D US offers a radiation-free, cost-effective, and real-time imaging solution that enhances safety for young patients requiring regular follow-ups. This technology also allows for dynamic imaging and portability, and it can assess patients in their natural standing posture, thereby reflecting the spine curvature. However, the inherent nature of ultrasound imaging presents challenges in spinal analysis. Factors such as image artifacts, variations in tissue contrast, and the difficulty of maintaining consistent probe alignment during freehand scanning can complicate the interpretation of ultrasound images. This thesis focuses on developing novel methods for the automatic extraction of vertebral features from FH 3D US data to facilitate non-invasive analysis of spine anatomy. A key contribution of this work is the identification of paired 3D lamina curves as a novel and robust representation of spinal shape. These curves are comprised of pairwise anatomical landmarks of the laminae, creating a robust and information-rich representation. Each lumbar lamina landmark is further employed to guide a deep learning model in accurately targeting and extracting the vertebral bone surface. To establish a foundation for automated analysis, two standardized manual labeling protocols were designed and validated by spinal ultrasound experts. These protocols ensure accuracy and consistency in identifying pairwise point landmarks on the laminae for 3D spine shape representation and for delineating the vertebral bone surface. Building upon this foundation, a Convolutional Neural Network (CNN)-based framework was developed for the automatic extraction of pairwise lamina landmarks from ultrasound images. We addressed challenges such as identifying optimal criteria for laminae structure identification, optimizing hyper-parameters for robust training, and selecting suitable CNN models. The performance of the CNN-based framework was assessed using K-Fold cross-validation on data from three participants. The results showed a mean distance error of 2.1 ± 1.3 mm and 1.8 ± 1.2 mm (3mm is acceptable for scoliosis assessment) for left and right lamina landmarks, respectively. This work demonstrated the feasibility of lamina landmark extraction from individual 2D ultrasound images, paving the way for real-time applications. To further enhance the robustness and smoothness of lamina curve extraction, we extend the CNN-based framework to Sequential Localization Recurrent Convolutional Network (SL-RCN), accommodating the 3D sequential nature of FH 3D US data. This advancement allows for the integration of 3D spine shape constraints into the extraction process, leading to more accurate representations. A 7-fold cross-validation is conducted on data from 7 participants, employing the leave-one-participant-out strategy. In contrast to the CNN-based framework, SL-RCN generates reduced left/right mean distance errors from 1.62/1.63mm to 1.41/1.40mm, and normalized discrete Frechet distances from 0.591/0.639 to 0.428/0.457. The experiment results demonstrated the effectiveness of SL-RCN in extracting accurate and smooth paired lamina landmark curves and ablation studies verified the utility of the architectural components comprising SL-RCN. Finally, this thesis explores the application of the Segment Anything Model (SAM) Zero-Shot for segmenting vertebral surfaces in ultrasound images, aiming to capture a complete vertebral visualization. The performance of SAM’s automated and prompt-based segmentation methods was evaluated, and a novel method leveraging landmark-prompted SAM and image intensity distribution was proposed for accurate extraction of vertebral bone surfaces. This method is specifically tailored for suboptimal transverse ultrasound images, addressing the limitations of invisible vertical edges in transverse spinal ultrasound images. The acoustic shadow masks beneath the extracted bone surface were evaluated against manually labeled masks, achieving a mean Intersection over Union above 0.92. This approach demonstrated promising results in reconstructing 3D meshes of lumbar vertebrae. In conclusion, this thesis presents novel methods for automatic extraction of vertebral features from FH 3D US data, paving the way for non-invasive analysis of spine deformities and facilitating the clinical application of spinal ultrasound imaging. The proposed methods for extracting paired 3D lamina curves and vertebral bone surfaces, coupled with the standardized manual labeling protocols, contribute significantly to the advancement of FH 3D US as a viable tool for spine imaging and intervention.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.