Leveraging Image Augmentations to Accurately Predict Spinopelvic Parameters in Lumbosacral X-rays using a Whole-Spine Artificial Intelligence Model

Introduction: The measurement of spinopelvic parameters is conducted radiographically via whole-spine or regionally-focused x-rays. Choosing the particular imaging view is often informed by the anatomic range of spinal pathology, severity of spinal deformity, radiation exposure, and other institutional preferences. Current tools to automate parameter measurement are often limited to one modality vs. the other which may limit their applicability in clinical practice. We previously constructed an artificial intelligence (AI) model to automatically predict spinopelvic parameters in whole-spine x-rays. In this study, we aim to extend the performance of that AI model to lumbosacral x-rays using only random image -cropping augmentations without the need for training on lumbosacral images.

Methods: The AI model was trained and validated on a set of 761 whole-spine sagittal x-rays with random cropping of images during the training process. The AI model was tested on a set of 40 lumbosacral spine x-rays to predict lumbar lordosis (LL) and sacral slope (SS). These predicted parameters were compared to annotations on the same images by a fellowship-trained spine surgeon with > 15 years of experience. Median error was recorded for each parameter and intraclass correlation coefficient (ICC) was calculated to assess interrater reliability.

Results: The median (IQR) error at each parameter was as follows: LL [2.9° (2.6°), p =0.80] and SS [1.9° (2.2°), p = 0.78]. The ICC values were LL (0.93) and SS (0.92) which both indicate excellent reliability between the model and ground truth predictions.

Conclusion : Using only random image augmentations, we were able to extend a whole-spine trained AI model to automatically predict spinopelvic parameters in lumbosacral images. With highly accurate and reliable performance on multiple radiographic modalities, our model exhibits a significantly increased range of clinical application.