利用踝關節統計形狀與運動模型配合機器學習預測試體關節上樓梯運動以利機械手臂量測外側韌帶之應變

Abstract

在人體下肢關節中，踝關節為承擔負重的主要關節之一，並在日常生活與運動中都扮演了重要的腳色。以往研究踝關節生物力學的方式分為活體研究和試體研究：活體研究雖然能量測踝關節的運動學資訊，但無法得知踝關節內部的詳細力學表現；而試體研究雖能透過機械手臂來得知踝關節內部之運動學資訊，但卻無法模擬試體於活體時之功能性動作，並以此得知其運動學資訊。為瞭解踝關節於爬樓梯時之力學表現，必須得到屬於此試體具活體意義之爬樓梯動作的運動學資訊，才能用試體實驗來模擬爬樓梯動作，並同時用三維全域變形及應變量測系統拍攝，以此來得知踝關節於爬樓梯時外側韌帶之應變。 本研究以74隻踝關節骨模型與14位受試者爬樓梯的運動學資訊來建立統計模型，再利用最佳化找到14組對應之係數組合來當作類神經網路的訓練資料，透過決定初步的類神經網路架構與改善預測表現等方式，來建立統計形狀模型與運動學資訊之間的連接，以此來預測試體的爬樓梯運動學資訊，並應用於機械手臂之試體實驗，同時用三維全域變形及應變量測系統拍攝，推算出踝關節外側韌帶之應變。 使用最佳化方法所重建的踝關節骨模型平均方均根誤差在脛骨、距骨和跟骨為0.40 mm、0.46 mm和0.58 mm，而在類神經網路訓練的平均誤差在平移與旋轉分別是1.42 mm和2.85度，接著透過訓練好的類神經網路預測出屬於試體的爬樓梯運動學資訊，並用於機械手臂之試體實驗，根據踝關節外側韌帶應變變化可以得知，前距腓韌帶於上樓梯時的應變變化最多，其次則是跟腓韌帶，後距腓韌帶則幾乎沒有變化。

The ankle joint is one of the main weight-bearing joints in the human lower limb joints and plays an important role in daily life and physical activity. Biomechanics of the ankle had been studied mainly in vivo and in vitro ways. In in vivo studies, the kinematics of a living individual can be measured, but not the detailed biomechanics. In vitro studies, detailed biomechanics can be measured by using the robotic testing system, but not in living, functional movements. To understand the biomechanics of the ankle joint upstairs, it is necessary to obtain the living kinematics of upstairs for cadavers for in vitro experiments and can get the strain of the ankle lateral ligaments. This study collected 74 ankle models and upstairs kinematics data of 14 subjects to establish the database. Through establishing the statistical model of geometry and kinematics, this study simplified the geometry and kinematics with a few simple coefficients and figured out the relationship between them with the artificial neural network, applied it to in vitro experiments, and then calculate the strain of the ankle lateral ligament. The reconstructed ankle joint bone models using the optimization method have the root mean square error of 0.40 mm, 0.46 mm, and 0.58 mm for the tibia, talus, and calcaneus, respectively. The average errors in translation and rotation during artificial neural network training are 1.42 mm and 2.85 degrees, respectively. Subsequently, the trained artificial neural network is used to predict the kinematics of stair climbing of the cadaver and is applied in experiments involving the robotic joint testing system (RJTS). According to the results of the ankle lateral ligament strain, the strain variation of the ATFL is the most significant, followed by the CFL, and the PTFL shows minimal variation.