AI Framework for Spine Shape Representation From 3D Ultrasound

Summary

Patients with adolescent idiopathic scoliosis (AIS) typically require spine exams every six months, resulting in frequent X-ray exposures and radiation accumulation that poses health risks. To avoid radiation accumulation, spinal ultrasound imaging has become increasingly popular in research studies for non-invasive evaluation of spinal curvature in scoliosis. However, existing scoliosis assessment methods require manual work. They are also biased and limited to 2D analysis. For automatic and accurate 3D spine anatomy analysis, we propose a deep learning framework for automated extraction of smooth 3D landmark curves from freehand 3D ultrasound, which can represent the spine shape, both quantifiably and visually. Keywords: deep learning, spine shape, vertebrae, landmarks, scoliosis, ultrasound.

Freehand 3D Ultrasound Sequences

A freehand 3D ultrasound (FH 3D US) system can conveniently provide imaging for large body structures like the spine. It is portable, compact, free of harmful radiation, and the least expensive. FH 3D US captures 3D ultrasound sequences with each frame including a distinct 2D image, and its corresponding transformation matrix obtained from a 3D position sensor. Figure 1 demonstrates the acquisition of spinal ultrasound sequences in 3D. As the sonographer operates the ultrasound probe along the spine shape, data frames are collected at a high frame rate. These are stored in two sequences: one with ultrasound images, and the other with positional information mapping images from the transducer to the world coordinate system. This enables the rendering of a 3D spinal ultrasound sequence in world coordinates.

Standardized Landmark Labeling Protocol

In current literature, extracting single-point landmarks representing the laminae is an ill-defined task—there is no formal guideline as to the most relevant part of the visible lamina bone surface in ultrasound images. We propose a protocol to solve this issue.

Figure 2, row II, presents a subjective preference diagram tailored to assist manual lamina landmark localization in ultrasound images, where only a portion of the lamina bone surface is visible. The preference diagram assigns the highest priority to the pair of points located slightly higher than the flat bottom of the left/right lamina regions, close to the spinous process and symmetrical to the center shadow. Subsequently, it monotonically decreases priorities on both sides. These preferred landmarks—typically located in the middle of high luminance horizontal bone surface curves—correspond to the paired green dots in row I.

We defined left/right lamina curve as the curve that optimally fits all left/right lamina landmarks across a 3D ultrasound sequence. Figure 2, row I, illustrates the anticipated paired vertebral lamina curves for individual lumbar vertebra and thoracic vertebra, which include paired sequences of lamina landmarks from the same vertebra. The green dotted curves represent the optimal pair, while the purple and red dotted curves correspond to the inner and outer boundary pairs, respectively. Any pair of curves within the inner and outer boundaries may be considered acceptable, provided that their central curve lies parallel to the spinal cord.

3D Visualization ꟷ Lamina Curve Comparison

Figure 3 demonstrates the comparison of 3D lamina curves, intercepting and zooming in on the subsequence of two lumbar vertebrae, extracted using our proposed deep learning framework (yellow curve), and the manual labeling method (green curve), respectively. The interweaving of two distinct colored curves affirms the effectiveness of artificial intelligence in automated extraction of 3D lamina curves.

Conclusion

We developed a deep learning model specifically tailored to 3D ultrasound sequence analysis, to automatically extract a pair of smooth 3D lamina curves. This novel deep learning framework underwent evaluation experiments on 3D ultrasound sequences, encompassing both lumbar and thoracic regions. The extracted lamina curves can be projected onto the coronal plane—offering the potential for scoliosis assessment and enable vertebral rotation analysis. Their sagittal plane projection can aid in vertebral level identification and kyphotic angle measurement using ultrasound imaging. The symmetrically distributed vertebral point clouds of lamina landmarks centered around the spinous process, potentially support 3D reconstruction of the spinal column with Statistic Shape Models. They also support 3D image registration by combining FH 3D US with an articulated shape model. Furthermore, the extracted vertebral landmarks from ultrasound images have applications in epidural anesthesia, image-to-image registration, and spinal intra-operative guidance.

Additional Information

For more information on this research, please read the following paper: S. Li, F. Cheriet, L. Gauthier and C. Laporte, "Automatic 3D Lamina Curve Extraction from Freehand 3D Ultrasound Data using Sequential Localization Recurrent Convolutional Networks," in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, doi: 10.1109/TUFFC.2024.3385698