- 學位論文
影像處理於肩盂唇核磁共振影像判讀之應用
Application of Image Processing for Shoulder Labrum Interpretation on Magnetic Resonance Image
摘要
肩盂唇（Shoulder labrum）是否完整，為醫師診斷肩關節是否受傷的依據之一。核磁共振（Magnetic Resonance Imaging，簡稱MRI）雖可提供良好的醫療解剖影像，但不同型號所產生的MRI影像，可能會造成診斷上的差異性。故本研究之目的是利用影像處理在肩盂唇核磁共振影像判讀上的應用，藉此幫助臨床醫師的判讀。 首先篩選孟唇損傷最常見的SLAP（Superior Labrum from Anterior to Posterior）tears和Bankart tears的MRI的病患。擷取其顯示受傷的切面影像後，使用銳利化（sharpening）去除模糊強化邊界。接著透過閥值來選取有興趣的ROI區域。最後再請醫師評估處理過的影像。系統評估研究中使用70張MRI的影像，其中5張是無受傷的影像，其餘的65張皆有呈現受傷的切面。65張裡GE有42張、SIEMENS有23張。 實驗結果顯示：（1.）GE 42張與SIEMENS 23張來做系統評估其結果為：靈敏度分別為0.92、0.91；專一性都為1；準確率分別為0.93、0.92；Kappa都為0.6。（2.）診斷為Bankart 26張與SLAP 33張系統評估其結果為：靈敏度分別為0.92、0.88；專一性都為1；準確率分別為0.93、0.89；Kappa分別為0.63、0.56。（3.）依MRI切面為Axial 42張與Coronal 17張的系統評估其結果為：靈敏度分別為0.87、0.88；專一性都為1；準確率分別為0.88、0.89；Kappa分別為0.43、0.59。（4.）將全部70張MRI影像來做整體評估的系統處理分析：靈敏度為0.87、專一性為1、準確率為0.88、Kappa為0.5。以上結果顯示系統對於不同來源、損傷及切面等，都能達到預期的效果。 在診斷效率方面，選取15張原圖影像（5張無受傷，10張受傷），請資深及資淺的2位放射科醫師來進行判讀，計算所需花費的時間分別為147秒、223秒，之後將這15張影像進行影像處理，隔一段時間並打亂其順序後再請2位醫師進行判讀，其所需花費的時間分別為123秒、207秒。顯示影像經過後處理後所需的診斷時間皆有所減少。 綜合上述結果，影像處理在肩盂唇MRI判讀上的應用是有助於臨床醫師在判讀上之參考。希望未來能將其他影像處理的技術應用於此，並且在實驗資料數據充足之下，讓系統能更加精準完善。
並列摘要
Integrity of the Shoulder Labrum is one of the bases for a doctor to diagnose whether the shoulder is injured. Although Magnetic Resonance Imaging (MRI) provides a good medical anatomical image, the number of images is limited for diagnosing Shoulder Labrum injury. The MRI images generated by different brands or types of scanner may also result in different diagnosis. The purpose of this study is to utilize image processing to improve MRI Interpretation of the Shoulder Labrum. The results could provide a reference for clinician diagnosis. First off, we chose several patients’ MRI images of Superior Labrum from Anterior to Posterior (SLAP) Tears and Bankart Tears, which are two of the most common Shoulder Labrum injuries. Secondly, after we captured the section images showing lesions, we utilized sharpening method to remove blurred parts and sharpen the borders in the images. Thirdly, we selected region of interest (ROI) by setting appropriate threshold value in the image processing method. Finally, we compared the image processing result with clinician diagnosis. In this study, totaling 70 images，five of them were unaffected images, the rest of the 65 images were presented with an injured section. We systematically analyzed 65 images from GE and MRI scanners with 42 and 23 images respectively. The performance of system shows that: (1) Sensitivity is 0.92 and 0.91 and Accuracy is 0.93 and 0.92 respectively for images from GE and SIEMENS MRI scanners; Specificity is 1, Kappa is 0.6 for both scanners. (2) Sensitivity are 0.92 and 0.88, Accuracy are 0.93 and 0.89, Kappa are 0.63 and 0.56 for Bankart and SLAP tears, respectively; Specificity are 1 for both tears. (3) Sensitivity are 0.87 and 0.88, Accuracy are 0.88 and 0.89, and Kappa are 0.43 and 0.59 for Axial and Coronal section, respectively; Specificity are 1 for both sections. (4) For all 70 images analyzed in this study, Sensitivity is 0.87; Specificity is 1; Accuracy is 0.88; and Kappa is 0.5. The results showed that the studied image processing system could achieve expected performance for MRI images from different scanners, injuries and sections. To analyze the improvement in diagnostic efficiency utilizing studied image processing system, we chose 15 original images (5 uninjured images and 10 injured images) for 2 radiologists (1 senior and 1 junior) to interpret. Time costs by senior and junior radiologist are 147 and 223 seconds, respectively. Next, we utilized the studied image processing system to process the 15 aforementioned images. After a period of time, we shuffled the order of the 15 processed images for 2 aforementioned radiologists to interpret. Time costs by senior and junior radiologist are 123 and 207 seconds, respectively. It showed that the diagnostic time cost is shortened by utilizing the studied image processing system. Summarizing the results of this study, it shows that the application of image processing could provide a good reference for clinician diagnosis. We hope that other image processing technologies can be utilized into this area in the future, and the system can be improved to more accurate and complete by imputing more sufficient testing data.