發展可利用膝關節三維幾何以預測自行車踩踏時該關節運動之人工智慧技術

Abstract

自行車運動在臨床上對於膝關節常被拿來做為傷後復健的手段，但不當的騎乘自行車也會對膝關節造成傷害，故了解自行車運動時的膝關節生物力學是相當重要的。以往研究膝關節力學的方式主要分為活體研究與試體研究：活體研究雖能量測人體膝關節之運動學資訊，但卻無法得知關節內部詳細之力學表現；而透過六軸機械手臂輔助之試體研究雖能得知動作中之生物力學表現，但因無從得知屬於試體之活體功能性運動學資訊，故無法模擬其動作並觀察其中生物力學。若能得知屬於試體、具活體意義的運動學資訊，未來便能利用試體實驗模擬自行車運動，並觀察過程中的生物力學變化，以對臨床復健及運動層面上的自行車運動給予建議。 本研究以57隻膝關節骨模型與24位受試者踩踏自行車的運動學資訊建立統計模型，並通過最佳化找出24組對應的係數組合作為類神經網路的訓練資料。透過決定初步類神經網路結構、評估應用範圍、改進預測表現等三個步驟，本研究成功從骨頭幾何外型預測運動學資訊，並對類神經網路的設置與應用範圍給出建議。 結果顯示以最佳化方法重建膝關節的平均方均根誤差於股骨與脛骨分別為0.68±0.07 mm與0.76±0.10 mm，而運動學資訊誤差至多只有0.51度與0.37 mm，顯示本方法能良好的重建幾何外型及運動學資訊。而本研究發現經模型對應後股骨體積介於174至220立方公分、脛骨體積介於132至161立方公分者有高的預測精確度，在運動學的平均預測誤差為1.77度與0.98毫米。本研究所開發之方法能以高準確度重建膝關節骨模型以及自行車踩踏運動學，並用類神經網路預測幾何外型所對應之運動學，以期未來能應用於基於機械手臂之試體實驗以了解膝關節於踩踏運動時的詳細生物力學。

Cycling is often taken as a rehabilitation treatment for the patients with injury of lower extremities. However, pedaling wrongly may cause injuries of knee joints. Therefore, revealing detail biomechanics of a pedaling knee is important. Biomechanics of knee had been studied mainly by in vivo and in vitro ways. With in vivo method, kinematics of a living individual can be measured, but not the detail biomechanics. On the other hand, detail biomechanics can be measured by in vitro studies, but not in living, functional movements. Therefore, providing the living kinematics for cadavers to in vitro experiments could be helpful for figuring out the biomechanics of pedaling. This study collected 57 knee models and pedaling kinematics data of 24 subjects to establish database. Through establishing statistical model of geometry and kinematics, this study simplified the geometry and kinematics with few simple coefficients and figured out the relationship between with artificial neural network (ANN), and the scope of application of this method was given. The reconstruction error of 3D model were 0.68±0.07 mm for femur and 0.76±0.10 mm for tibia. The reconstruction errors of kinematics were up to 0.51 degree for rotation and 0.37 mm for translation. Leave-one-out tests were done with the joints with volume after model correspondence in 174 to 220 cm3 for femur and 132 to 161 cm3 for tibia, the predicted kinematics had error of 1.77 degree for rotation and 0.98 mm for translation. This study developed a method to predict the knee kinematics from the geometry to provide a meaningful kinematics for cadavers which could be applied to in vitro experiments, and might help us revealing the detail biomechanics while pedaling.