RESUMO
Objective: The measurement of in vivo intervertebral disc (IVD) mechanics may be used to understand the etiology of IVD degeneration and low back pain (LBP). To this end, our lab has developed methods to measure IVD morphology and uniaxial compressive deformation (% change in IVD height) resulting from dynamic activity, in vivo, using magnetic resonance images (MRI). However, due to the time-intensive nature of manual image segmentation, we sought to validate an image segmentation algorithm that could accurately and reliably reproduce models of in vivo tissue mechanics. Design: Therefore, we developed and evaluated two commonly employed deep learning architectures (2D and 3D U-Net) for the segmentation of IVDs from MRI. The performance of these models was evaluated for morphological accuracy by comparing predicted IVD segmentations (Dice similarity coefficient, mDSC; average surface distance, ASD) to manual (ground truth) measures. Likewise, functional reliability and precision were assessed by evaluating the intraclass correlation coefficient (ICC) and standard error of measurement (SEm) of predicted and manually derived deformation measures. Results: Peak model performance was obtained using the 3D U-net architecture, yielding a maximum mDSC â= â0.9824 and component-wise ASDx â= â0.0683 âmm; ASDy â= â0.0335 âmm; ASDz â= â0.0329 âmm. Functional model performance demonstrated excellent reliability ICC â= â0.926 and precision SEm â= â0.42%. Conclusions: This study demonstrated that a deep learning framework can precisely and reliably automate measures of IVD function, drastically improving the throughput of these time-intensive methods.
