Automatic deep learning-based assessment of spinopelvic coronal and sagittal alignment

Spine and/or pelvis deformity refers to morphological abnormalities in the coronal, sagittal, or axial position as a deviation from the normal position. The prevalence of such deformities can be up to 2% of adults for scoliosis and consequences on quality of life are one current public health issue [1,2]. The assessment and follow-up of spinopelvic alignment relies on radiographic measurements via either conventional uniplanar x-ray systems or orthogonal biplanar acquisitions at low dose stereoradiography [3], allowing to better understand postural adaptation and the close relationship between pelvis and spine [4]. Several radiographic parameters have been developed and validated on conventional uniplanar two-dimensional whole-spine radiographs to evaluate spinopelvic harmony and the sagittal balance of the spine [5]. Cobb angle (CA) is the most widely used parameter to evaluate frontal deviation of the spine and the reference standard to diagnose and monitor scoliosis [6]. Frontal pelvic asymmetry (FPA), sacral slope (SS), pelvic tilt (PT) and pelvic incidence (PI) are the most established parameters to evaluate spinopelvic balance.

Although manual measurements are found precise and accurate by expert readers [7,8], they remain burdensome and time-consuming in the routine radiologists’ workflow. Computer-aided semi-automated techniques have been developed for the last ten years but the recent improvement of deep learning techniques offers new perspectives, with promising results on automated deep learning-based assessment of angular deviations in the coronal or sagittal plane [9], [10], [11], [12]. However, none has assessed the performance and potential impact of a full solution for coronal and sagittal measurements of spine and pelvis parameters for routine radiologic implementation.

The main objective of this study was to assess, in a pediatric and adult population, the consistency of an AI solution for coronal and sagittal spinopelvic alignment assessment compared to the gold standard of musculoskeletal radiologist measurements. Secondary objectives were: (i) to evaluate the performance of the AI solution to classify measurements based on admitted threshold values for CA, FPA and SS; (ii) to report the potential errors of young/non expert readers; and (iii), to evaluate the potential impact of the AI solution on an overall workflow.

留言 (0)

沒有登入
gif