Characterization of spine and torso stiffness via differentiable biomechanics.

We present a methodology to personalize the stiffness response of a biomechanical model of the torso and the spine. In high contrast to previous work, the proposed methodology uses controlled force-deformation data that mimic the conditions of spinal bracing for scoliosis, which leads to personalized biomechanical models that are suitable for computational brace design. The novel methodology relies on several technical contributions. First, a prototype system that includes controlled force measurement and low-dose radiographs, with low-encumbrance for its implementation in the clinical protocol. Second, a model of differentiable biomechanics of the torso and the spine, which becomes the key building block for robust parameter estimation. And third, an optimization procedure for parameter estimation from force-deformation data, which relies on differentiability of the biomechanics and the image generation process. We demonstrate the application of the methodology to a cohort of 7 subjects who underwent scoliosis check-ups, and we show quantitative validation of the estimated personalized parameters and the improvement over default parameters from the bibliography.