結合髖關節X光影像特徵萃取及機器學習建構骨質疏鬆之預測模式

Combination Hip X-ray Image Features Extraction and Machine Learning Predictive Osteopenia and Osteoporosis

洪詩凱(Shih-Kai Hung) ； 許巍嚴(Wei-Yen Hsu) ； 施欣怡(Hsin-Yi Shih) ； 林弘智(Hung-Chih Lin) ； 胡雅涵(Ya-Han Hu)

《中華放射線技術學雜誌》 40卷2期 (2016/06) Pp. 59-67

摘要

雙能量X 光吸收儀為目前診斷骨質疏鬆症的標準工具，在醫療資源有限的情況下，臨床醫師通常會先評估病患臨床危險因子或初篩檢查數據，針對危險性高的族群再以雙能量X 光吸收儀進一步確診。本研究使用骨質疏鬆症的重要臨床危險因子，並結合髖關節股骨頸影像特徵萃取，嘗試使用機器學習技術找出具有代表性的數值來幫助分類，進而發展骨質疏鬆初步篩檢的預測模型。本研究使用X 光數位影像並採用回溯性方式收集50 歲以上男性及停經女性的髖關節X 光影像，並依照世界衛生組織所訂定的骨質疏鬆症診斷標準，依T 値分為正常與異常兩類。通過圖像形態特徵萃取的方法，直接計算X 光片上的骨小梁型態，將所得參數結合骨鬆重要危險因子，並應用類神經網路、支援向量機以及邏輯斯迴歸的分類技術，透過人工智慧的演算來建構骨質疏鬆之預測模式，並與雙能量X 光吸收儀的結果進行比較。本研究結果以類神經網路ANN 為最佳分類器，其中停經婦女正確率為86.20%、敏感性為84.71%、特異性為87.66%、ROC 曲線為0.929；男性50 歲以上正確率為85.82%、敏感性為81.64%、特異性為90.02%、ROC 曲線為0.953。本研究提供另一種有別於以往的方法作為未來臨床醫師評估骨質疏鬆之初步篩檢的參考，協助找潛在具骨鬆風險之患者，必要時再轉介雙能量X 光吸收儀以確立診斷。

關鍵字

雙能量X 光吸收儀 ； 骨質疏鬆症 ； 機器學習 ； 特徵萃取

並列摘要

Dual energy X-ray absorptiometry (DXA) is typically used to diagnose osteoporosis. Due to limited medical resources, clinical risk factors (CRFs) and initial screening results are usually assessed by clinicians for the very beginning. For patients with high risk of osteoporosis, a further diagnosis using DXA will be made. In order to develop a prediction model for initial screening of osteoporosis, we used CRFs and features extraction of femoral neck in combination with machine learning to find the most representative factors, which can help for further classification. We retrospectively collected the hip X-ray from male individuals with ages over 50 year's old and female individuals with postmenopausal women, respectively. According to osteoporosis diagnostic criteria established by World Health Organization, T-score was divided into two cohorts. Combination hip X-ray image features extraction and machine learning predictive osteopenia and osteoporosis. By using artificial neural network (ANN), support vector machine (SVM) and logistic regression (LR), these scores along with important risk factors of osteoporosis were analyzed by machine learning algorithm to construct a prediction model for osteoporosis. These prediction results were compared with the results of DXA. The results show that the best prediction model is artificial neural network (ANN). Regarding the prediction model for postmenopausal women, the accuracy, sensitivity, specificity, and ROC curve were 86.20%, 84.71%, 87.66% and 0.929, respectively. Regarding the prediction model for the men ages over 50 years old, the accuracy, sensitivity, specificity, and ROC curve were 85.82%, 81.64%, 90.02% and 0.953, respectively. Our method provides an alternative of initial prognosis for osteoporosis, especially for asymptomatic patients and those with high fracture risk.

並列關鍵字

Dual energy X-ray absorptiometry ； Osteoporosis ； machine learning ； features extraction