3D vertebrae labeling in spine CT: an accurate, memory-efficient (Ortho2D) framework

Abstract

Purpose. Accurate localization and labeling of vertebrae in computed tomography (CT) is an important step toward more quantitative, automated diagnostic analysis and surgical planning. In this paper, we present a framework (called Ortho2D) for vertebral labeling in CT in a manner that is accurate and memory-efficient. Methods. Ortho2D uses two independent faster R-convolutional neural network networks to detect and classify vertebrae in orthogonal (sagittal and coronal) CT slices. The 2D detections are clustered in 3D to localize vertebrae centroids in the volumetric CT and classify the region (cervical, thoracic, lumbar, or sacral) and vertebral level. A post-process sorting method incorporates the confidence in network output to refine classifications and reduce outliers. Ortho2D was evaluated on a publicly available dataset containing 302 normal and pathological spine CT images with and without surgical instrumentation. Labeling accuracy and memory requirements were assessed in comparison to other recently reported methods. The memory efficiency of Ortho2D permitted extension to high-resolution CT to investigate the potential for further boosts to labeling performance. Results. Ortho2D achieved overall vertebrae detection accuracy of 97.1%, region identification accuracy of 94.3%, and individual vertebral level identification accuracy of 91.0%. The framework achieved 95.8% and 83.6% level identification accuracy in images without and with surgical instrumentation, respectively. Ortho2D met or exceeded the performance of previously reported 2D and 3D labeling methods and reduced memory consumption by a factor of ∼50 (at 1 mm voxel size) compared to a 3D U-Net, allowing extension to higher resolution datasets than normally afforded. The accuracy of level identification increased from 80.1% (for standard/low resolution CT) to 95.1% (for high-resolution CT). Conclusions. The Ortho2D method achieved vertebrae labeling performance that is comparable to other recently reported methods with significant reduction in memory consumption, permitting further performance boosts via application to high-resolution CT.

Multi-modality registration of preoperative MR and intraoperative long-length tomosynthesis using GAN synthesis and 3D-2D registration

“…Recognizing the potential for deformations, we will employ a locally rigid registration approach (for each vertebra) to account for changes in spinal curvature. 8 This approach will be combined with automatic initialization using CNN-based vertebrae labeling presented in our early work 9,10 to extend the capture range, parallelize registrations, and reduce algorithm runtime.…”

Deformable 3D–2D image registration and analysis of global spinal alignment in long‐length intraoperative spine imaging

“…The proposed framework involves a few important inputs, including labeling of the vertebrae in preoperative CT, which forms the basis for subvolume masking in the multi‐scale approach. Such labeling can be accomplished via automatic methods, including those based on appearance models, 43 probabilistic models, 44 and convolutional neural networks 37,45 . The vertebrae labels can also be used for the analysis of GSA.…”

“…Such labeling can be accomplished via automatic methods, including those based on appearance models, 43 probabilistic models, 44 and convolutional neural networks. 37,45 The vertebrae labels can also be used for the analysis of GSA. A method for labeling vertebrae in long-length 2D images is currently under development, providing a means for fast initialization of the multi-scale registration framework via matching of labels between preoperative 3D and intraoperative 2D images.…”

“…Volumetric masks are computed from a 3D spline curve fitted to the vertebral centroids defined in preoperative CT or CBCT, and no additional user input is required during the progression of the framework. Vertebral centroids were defined manually on the preoperative images and may be defined by automated methods, 36,37 which is the subject of other work. Based on the size of the masked volume and the projective geometry, real and synthetic images were cropped to the region of support to reduce memory consumption.…”