建構基於人工智慧姿態辨識評估肌少症之分類模型

Taipei Medical University Institutional Repository > 醫學院 > 人工智慧醫療碩士在職專班 > 博碩士論文 > Item 987654321/63348

請使用永久網址來引用或連結此文件: http://libir.tmu.edu.tw/handle/987654321/63348

題名:	建構基於人工智慧姿態辨識評估肌少症之分類模型 Constructing a Human Pose estimation Model Based on Artificial Intelligence For Evaluating Sarcopenia
作者:	黃李琪 HUANG, LI-CHI
貢獻者:	醫學院人工智慧醫療碩士在職專班康峻宏
關鍵詞:	人工智慧;姿態辨識;肌少症 Artificial Intelligence;Human Pose estimation;Sarcopenia
日期:	2023-07-11
上傳時間:	2023-12-15 14:37:24 (UTC+8)
摘要:	背景：肌少症是一種以骨骼肌肉質量和功能逐漸喪失為特徵的疾病。為了預防和治療肌少症，持續的飲食管理和規律運動以及患者的自我管理固然重要。然而，傳統的肌少症診斷和監測方法難以由患者自己定期監測，因為他們必須到醫療機構就診，並且需要特殊設備，例如雙能X射線吸收測定法、握力測量和步行速度測量。為了改善傳統檢測方法的缺點，本研究預建構一個分類模型，評估老年人的三公尺計時起身行走測試（Timed-up and GO test, TUG）表現，使用OpenPose的關鍵點模型分析，藉由深度學習技術建立一個能夠評估肌少症的動作分析模型。研究方法：本計畫共招募肌少症患者30名及非肌少症受試者各39名，被招募的肌少症患者及非肌少症受試者的年齡應在65歲以上。依亞洲肌少症工作小組（Asia Working Group for Sarcopenia, AWGS）的診斷標準，依序完成生物電阻抗分析 (Bioelectrical impedance analysis, BIA)、握力測試、簡短身體功能量表(short physical performance, SPPB)及坐站起走測試。以手機攝影機攝影側面角度的影像，作為研究影片資料。影片資料透過OpenPose 1.7版，得到人體25個關鍵點的JavaScript Object Notation (JSON)檔資料。利用開發的程式擷取25個關鍵點座標；藉由人體姿態及各參數的變化做相關特徵選取，再透過深度學習架構讓一維卷積神經網路分類肌少症及非肌少症的動作表現。並驗證分類器分類能力的準確性。研究結果：研究結果顯示，使用三公尺計時起身行走測試，透過卷積神經網路建構肌少症動作偵測模型是可行的，且所建置的模型準確率達84%。結論：本研究結果可用來開發肌少症自我監測技術的基礎研究。 Background：Sarcopenia is a disease characterized by a progressive loss of skeletal muscle mass and function. To prevent and treat sarcopenia, continuous dietary management and regular exercise, and self-management of the patients are crucial. However, the conventional sarcopenia diagnosis and monitoring methods are difficult for regular monitoring by the patients themselves because they have to visit a medical institution and it requires many procedures and special devices such as DXA, handgrip strength measurement, and walking speed measurement. To improve this shortcoming of the conventional method, this study implemented a classification model for predicting the sarcopenia with deep learning techniques using OpenPose data which are measured when assessing the TUG performance ability. Materials and Methods：A total of 30 patients with sarcopenia and 39 non-sarcopenia participants will be recruited for this study. The recruited patients with sarcopenia and non-sarcopenia participants should be aged 65 years or older. According to the diagnostic criteria of the Asia Working Group for Sarcopenia (AWGS), the following assessments were conducted in sequence: bioelectrical impedance analysis (BIA), grip strength test, short physical performance battery (SPPB), and sit-to-stand test. The video data will be processed using OpenPose 1.7 to obtain JSON files containing the coordinates of 25 key points of the human body. A developed program will extract the coordinates of the 25 key points, and feature selection will be performed based on the changes in body posture and various parameters. A one-dimensional convolutional neural network will be trained using a deep learning framework to classify the movement performance as sarcopenia or non-sarcopenia. The accuracy of the classifier's classification ability will be validated. Results：The research results indicate that it is feasible to construct a sarcopenia movement detection model using a convolutional neural network through the Timed Up and Go test, and the accuracy of the constructed model reaches 84%. Conclusion：This study can be used as a basic research for the development of self-monitoring technology for sarcopenia.
描述:	碩士指導教授:康峻宏委員:郭柏齡委員:陳弘洲委員:康峻宏
資料類型:	thesis
顯示於類別:	[人工智慧醫療碩士在職專班] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	64	檢視/開啟

在TMUIR中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....