利用實驗測量的皮膚表面資訊進行統計形狀建模和骨骼模型的重建

Abstract

人體下肢骨骼模型在醫學、運動科學和生物力學中具有重要作用，有助於更精確地診斷和治療下肢疾病，改善手術規劃，並在假肢設計和復健中發揮關鍵作用。相比X光或CT掃描，三維下肢骨骼模型提供了更直觀的視覺參考，使醫生能更清晰地理解骨骼結構和病變部位，從而做出更準確的診斷。 目前，大多數研究和應用中使用的模型是基於大量健康個體的平均值建立的典型模型，這些模型適用於大多數情況，成本和資源投入相對較低。然而典型模型缺乏針對特定個體的個性化考量，對於處理特定個體的特殊解剖結構或病理狀況時適用性有限。 因此，若能透過非侵入式的方法取得身體資訊並用於下肢骨頭SSM以產生個人化骨模型，可降低醫療成本及風險。本研究之目的為根據個體皮膚表面資訊預測個人化骨頭模型，相較於醫學影像這種耗費成本較高的技術，皮膚表面資訊可以透過紅外線掃描較等低成本技術取得。本研究使用開源資料庫之電腦斷層影像取得三維皮膚和骨頭模型，透過統計形狀模型技術方法建立骨頭統計形狀模型，用於重建特定於受試者的個人化骨骼，從而降低個人化骨頭模型的取得門檻。 在統計形狀模型的建立過程中，首先，我們須建立訓練模型間的對應關係，透過Scaling Itrative Closet Point (S-ICP)和Bayesian Coherent Point Drift (BCPD)方法將模型之間的關係確立。接著，再透過Generalized Procrustes analysis (GPA)將模型對齊後取得平均模型和模型變異。最後，再使用Principal component analysis (PCA)進行特徵提取，以建立出統計形狀模型。 在本研究的結果中，為驗證了透過皮膚表面資訊重建特定於受試者的骨骼之猜想，成功利用30組股骨模型訓練的統計形狀模型預測訓練集中之骨骼模型。接續使用分別由30組股骨模型和29組脛骨模型訓練出的兩統計形狀模型，並以leave one out之方式進行預測。在預測的結果中，在考量位置和型態上之誤差下，股骨的預測誤差為9.55mm，脛骨的預測誤差為7.34mm；若單考量形態上之誤差，股骨的預測誤差為3.31mm，脛骨的預測誤差為3.81mm；考量預測結果與皮膚之距離和預測結果中的第四維資訊為誤差的情況下，股骨的預測誤差為6.37mm，脛骨的預測誤差為2.70mm。 本研究成功使用股骨、脛骨模型及其與皮膚之距離資訊，構建四維統計形狀模型。通過訓練模型對應的皮膚進行驗證，確認通過皮膚表面資訊預測相應的骨骼之方法具可行性。接著，使用leave-one-out方法驗證和量化此方法的精確度。雖位置和形態上的誤差相較於型態誤差較高，但型態誤差在3~4 mm，證明皮膚資訊可重建與皮膚擁有者相似的骨骼結構。而脛骨在位置和形態上之誤差及皮膚誤差都比股骨更小，是因為脛骨基本上貼合小腿，決定位置的依據更高，故誤差較小。 本研究經由最佳化的方式，重建股骨和脛骨的幾何外型，並根據訓練模型內之骨頭所對應之皮膚，成功還原了該骨頭的樣貌，說明了此方法的可行性。接著透過leave-one-out測試，我們得到股骨在位置和形態上的誤差、型態上的誤差、皮膚到骨頭距離資訊之誤差。雖考慮位置和形態上的誤差會比僅考慮形態時來得更大。然而，若僅考慮形態上的誤差大約為3到4毫米，這表明透過皮膚資訊，可以在一定程度上重建出與該皮膚擁有者相似的股骨和脛骨結構。

The model of the human lower limb skeleton plays a crucial role in medicine, sports science, and biomechanics, aiding in more precise diagnosis and treatment of lower limb diseases, improving surgical planning, and contributing significantly to prosthetic design and rehabilitation. Compared to X-rays or CT scans, three-dimensional models of the lower limb skeleton provide a more intuitive visual reference, allowing doctors to better understand the skeletal structure and pathological sites, leading to more accurate diagnoses.

Currently, most research and applications use typical subject models based on the average values of a large number of healthy individuals. These models are applicable in most cases and require relatively low costs and resources. However, typical models lack personalized considerations for specific individuals, limiting their applicability when dealing with unique anatomical structures or pathological conditions of specific individuals.

Therefore, if non-invasive methods can be used to obtain body information and applied to the Statistical Shape Model (SSM) of lower limb to generate personalized bone models, medical costs and risks can be reduced. The purpose of this study is to predict personalized bone models based on individual skin surface information. Compared to medical imaging techniques that are relatively high in cost, skin surface information can be obtained using low-cost techniques such as infrared scanning. This study uses open-source database CT images to obtain three-dimensional skin and bone models. Using the Statistical Shape Model technique, bone SSMs are established and used to reconstruct personalized bones specific to the subjects, thereby lowering the threshold for obtaining personalized bone models.

In the process of establishing the SSM, we first need to establish correspondences between training models. The Scaling Iterative Closest Point (S-ICP) and Bayesian Coherent Point Drift (BCPD) methods are used to establish these correspondences. Next, Generalized Procrustes Analysis (GPA) is used to align the models to obtain the mean model and model variaces. Finally, Principal Component Analysis (PCA) is used for feature extraction to establish the SSM.

In the results of this study, we validated the hypothesis of reconstructing bones specific to the subjects using skin surface information. We successfully used an SSM trained with 30 sets of femur models to predict the bone model in the training set. We then used two SSMs trained with 30 sets of femur models and 29 sets of tibia models respectively, and conducted predictions using the leave-one-out method. In the prediction results, considering the errors in position and shape, the prediction error for the femur was 9.55mm, and 7.34mm for the tibia; considering only shape errors, the femur prediction error was 3.31mm, and 3.81mm for the tibia; considering the prediction result’s distance to the skin and the error in the fourth dimension information of the prediction result, the femur prediction error was 6.37mm, and 2.70mm for the tibia.

This study successfully constructed four-dimensional SSMs using femur and tibia models and their distance information to the skin. Through validation with the corresponding skin of the training models, we confirmed the feasibility of predicting the corresponding bones using skin surface information. Then, using the leave-one-out method, we validated and quantified the accuracy of this method. Although the errors in position and shape were higher compared to only shape errors, the shape errors were around 3-4 mm, demonstrating that skin information can reconstruct bone structures similar to those of the skin owners. The tibia had smaller errors in position, shape, and skin compared to the femur because the tibia essentially conforms to the calf, providing higher positional determination, thus resulting in smaller errors.

This study used optimization methods to reconstruct the geometry of the femur and tibia and successfully restored the appearance of the bones based on the skin corresponding to the bones in the training models, demonstrating the feasibility of this method. Using leave-one-out testing, we obtained the errors in position and shape, shape errors, and the errors in skin-to-bone distance information for the femur. Although considering both position and shape errors resulted in higher errors than considering only shape errors, the shape errors were about 3 to 4 millimeters, indicating that through skin information, it is possible to reconstruct femur and tibia structures similar to those of the skin owners to a certain extent.