深度學習於骨折醫療影像辨識系統之應用

人工智慧出現於1950年代，來自當時電腦科學領域的先驅，被廣泛應用於各領域協助解決許多問題，皆得到不錯的結果。機器學習是人工智慧領域中相當重要的部分，已經成功的被運用在許多方面，例如：影像分類、語音識別、數位助理以及自動駕駛等。機器學習的原理是透過大量的資料學習，建立能正確反應或判斷的資訊模型。在醫療保健的應用範圍，包括從一維生物信號分析和對諸如癲癇發作，及心臟驟停等醫學事件的預測，到支持臨床決策分析的電腦輔助檢測診斷系統，藥物發現、輔助治療選擇、藥物基因學，提高手術效率，及電子健康分析記錄等。在醫療保健領域中，深度學習的廣泛使用仍處於剛起步階段，但是整個學術界都有強力的措施，許多大型公司正在基於機器學習的醫療保健項目中進行研究。不僅是醫療科技公司，還包括 Google Brain、DeepMind、Microsoft和IBM等大型公司。其中骨折醫療影像辨識系統也是機器學習中重要的應用領域，骨折是急診室中最常見的損傷之一，根據衛生福利部全民健康保險醫療統計年報統計，骨折人數：108 年的前臂骨折人數為 101965 人，股骨骨折人數為73088人，小腿骨折包含踝共95125人，針對骨折診斷首用的技術主要就是 X 光片，這是一種已經使用了一百多年並且仍然經常使用的方法，然而醫師在放射影像上準確地識別骨折是困難的，因為骨折具有其獨特性，它們可以發生在任何骨骼之中，其外觀取決於區域解剖結構和 X 光片的投影。由於醫師可能超負荷工作及長時間閱讀片子產生疲勞，或是在緊急情況下必需採取行動，年輕醫師經驗不足且在 X 光片解釋時經常都是在沒有第二專家意見中發生等，這些因素都可能導致骨折判讀的遺漏。骨折誤診是急診室最常見的診斷錯誤及醫療疏失索賠的最常見原因，可導致治療延誤和骨頭癒合不良、關節炎甚至是長期的殘疾。深度學習可以輔助醫師準確的識別骨折，這種方法可以避免大多數判讀時所發生的錯誤。論文中回顧了深度學習應用於醫學影像及骨折醫學影像相關的研究，討論的骨折辨識研究中AUC 數據平均為 0.96，敏感性平均為 93%，特異性平均為 91.7%，表示經由卷積神經網路的深度學習模型非常適用於醫學骨折圖像識別，再深入探討研究論文發現，深度學習對於即使是像舟狀骨這種不易辨識的隱匿性骨折也具有良好的效果。如果有一套好的醫學影像電腦輔助診斷系統，將可有效幫助醫師減少誤判，進而提供病人更正確及時的治療。本論文對每個研究進行匯整和評估，以統整歸納一套可靠的基準來針對骨折進行分類及檢測方法，建立一套標準提供醫院建置深度學習骨折醫療影像辨識系統。

關鍵字

人工智慧；深度學習；骨折醫療影像辨識；卷積神經網路；反向傳播算法；隱匿性骨折

並列摘要

Artificial intelligence emerged in the 1950s. It came from a pioneer in the field of computer science at that time. It was widely used in various fields to help solve many problems, and got good results. Machine learning is a very important part in the field of artificial intelligence, and has been successfully used in many aspects, such as;Image classification, speech recognition, digital assistant and self-driving cars, etc. The principle of machine learning is to learn from a large amount of data to build an information model that can correctly react or judge. Applications in healthcare include from one-dimensional biosignal analysis and prediction of medical events such as epileptic seizures, cardiac arrest, and other medical events, to computer-aided detection and diagnosis systems that support clinical decision-making and analysis, drug discovery, auxiliary treatment options, pharmacogenetics, improve the efficiency of surgery, and electronic health analysis records, etc. In the field of healthcare, the widespread use of deep learning is still in its infancy, but the entire academic community has strong measures, and many large companies are conducting research in healthcare projects based on machine learning. Not only medical technology companies, but also large companies such as Google Brain, DeepMind, Microsoft and IBM. Among them, the fracture medical image detection system is also an important application field in machine learning. Fractures are one of the most common injuries in the emergency room. According to the National Health Insurance Report, the number of fractures on 2019, forearm fractures was 10,1965, femur fractures was 73088, and low leg fractures included ankle was 95,125 of the people. The first technology used for fracture diagnosis is mainly X-ray film, which is a method that has been used for more than one hundred years and is still used frequently. However, it is difficult for doctors to accurately identify fractures on radiographic images, because fractures have their unique characteristics, they can occur in any bone, and their appearance depends on the regional anatomy and the projection of X-rays. Because doctors may be overworked and tired from reading films for a long time, or must take action in an emergency, young doctors are inexperienced and the interpretation of X-rays often occurs without a second expert opinion, etc., Fractures may be missed due to these factors. Fracture misdiagnosis is the most common diagnosis error in the emergency room and the most common cause of medical negligence claims. It can lead to treatment delays and poor bone healing, arthritis and even long-term disability. However, deep learning can assist physicians in accurately identifying fractures. This method can avoid most errors in interpretation. The paper reviews the application of deep learning to medical imaging and fracture medical imaging. The average AUC data in the fracture identification study discussed is 0.96, the average sensitivity is 93%, and the average specificity is 91.7%, Indicates that the deep learning model via convolutional neural network is very suitable for medical fracture image recognition, and further research papers found that deep learning has a good effect even for occult fractures such as scaphoid bones that are not easy to identify. If there is a good computer-aided diagnosis in medical imaging system, it will effectively help doctors reduce misjudgments and provide patients with more accurate and timely treatment. This thesis summarizes and evaluates each study, uses a set of reliable benchmarks to classify and detect fractures, and establishes a set of standards to provide hospitals with a deep learning fracture medical imaging detection system.