Towards real-time finite element simulations with machine learning for spinal surgical pre-operative planning

我们的合作伙伴Spinologics Inc.是一家专门从事脊柱相关医疗器械开发的小型公司，使用有限元建模（FEM）。多年来，他们开发了一个生物力学数值实验室来模拟人体与医疗设备（如植入物或假体）之间的相互作用。Spinologics考虑使用人工智能和机器学习来及时预测脊柱侧凸患者的手术矫正。在机器学习和人工智能方面的专业知识在他们的研究人员中是不存在的。因此，他们寻求我们在脊柱机器学习方面的专业知识。脊柱矫正取决于患者脊柱的生物力学。这种矫正与患者的脊柱运动范围有关，这是脊柱椎间柔韧性以及变形脊柱形状所固有的。将根据不同的患者解剖结构和模拟集生成大量模拟;将提取特征以学习模拟的参数。这种技术将是非常有趣的，因为它将从模拟中学习生物力学模型，并在几秒钟内为临床医生提供视觉更新。将引入深度神经网络（DNN）来预测输出形状。DNN能够对输入和输出变量之间复杂的非线性关系进行建模。在我们的项目中，我们希望了解应用哪些力来获得输出轮廓。用于训练模型的输入参数将包括：患者的术前脊柱形状、脊柱柔韧性和目标轮廓。学习模型的输出是可行的脊柱形状。然后，训练的模型将用于预测给定目标轮廓的特定患者的合理脊柱形状。然后，外科医生可以操纵目标脊柱轮廓，并对可行的脊柱形状进行近实时反馈。该项目可能为全自动患者特定建模铺平道路，它可能导致脊柱手术预测的重大进展，它可以改善加拿大患者的医疗保健，同时对脊柱外科医生有用。

Our partner, Spinologics Inc. is a small company specialized in spine related medical device development using finite element modelling (FEM). Over the years, they have developed a biomechanical numerical laboratory to simulate the interactions between the human body and medical devices, such as implants or prosthesis. Spinologics considers using artificial intelligence and machine learning to predict the surgical correction of scoliosis patients in a timely manner. Such expertise in machine learning and artificial intelligence is not available in their research staff. Thus, they seek our expertise in machine learning of the spine. Spinal correction is dependent on the biomechanics of the patient's spine. Such a correction is linked to the patient's spinal range of motion, which is inherent to spinal intervertebral flexibility as well as the shape of the deformed spine. A large number of simulations would be generated from different patient anatomy and from the set of simulations; features would be extracted to learn the parameters of the simulation. Such technique would be of high interest, since it would learn the biomechanical model from the simulations and provide clinicians with a visual update within seconds. Deep neural networks (DNN) will be introduced for the prediction of the output shape. DNNs are capable of modelling complex, nonlinear relationships between input and output variables. In our project, we wish to learn which forces to apply to obtain the output profile. The input parameters to train the model would include: the patient's pre-operative spinal shape, spinal flexibility, and target profiles. The output of the learning model is the feasible spinal shape. The trained model would then serve to predict the plausible spinal shape for a specific patient given a target profile. The surgeon could then manipulate the target spinal profile and have a near real-time feedback on the feasible spinal shape. This project may pave the road for a fully automatic patient-specific modelling and it could lead to significant advances in spine surgery prediction and it could improve health care to Canadian patients, while being useful to spine surgeons.