- 學位論文
利用影像深度學習對二維MRI T1和T2腰部脊椎骨折加權影像辨識之研究
The Classification of Vertebral Fracture by Using Deep Learning on MRI T1 and T2 Images
摘要
臨床常使用X光(X-ray)、電腦斷層(Computer Tomography, CT)和磁振造影(Magnetic Resonance Imaging, MRI)診斷脊椎骨折。MRI辨別脊椎骨折引發椎間盤突出的鑑別能力較CT優異，因此經常利用MRI影像診斷脊椎方面的疾病；而矢狀(Sagittal)、冠狀(Coronal)和橫狀(Transverse)切面中以矢狀切面常用於診斷，但切片數不少且細節很多，為了縮短看片時間以及提高電腦輔助診斷，因此採用深度學習建立MRI 影像脊椎骨折之判識模型。 本研究採用回顧性分組實驗設計，收集義大醫院2019年4月至2020年1月接受MRI T1及T2造影者共213例；其中實驗組為具有脊椎骨折者共108例、二維影像為1199張；對照組為無脊椎骨折者共105例、二維影像為1144張。主要收集MRI T1和T2之加權矢狀切面影像；影像為DICOM格式其Matrix Size大小為448×224、每例MRI T1及T2矢狀切面數約介於8至11張左右。深度學習採用AlexNet、VGG19及Resnt50三種深度學習方法；使用轉移學習方式(Transfer Learning)建立CNN影像分類架構；分析影像亦有三種模式，包括二維T1、二維T2以及T1、T2與T1及T2融合成RGB三通道影像。因此，共計有9種分類模型，所有模型中訓練集占比例為70%，即實驗組與對照組MRI影像各839及801張；測試集占30%，實驗組與對照組MRI影像360及343張。 測試集分類結果顯示，AlexNet對T1、T2與RGB影像之準確度、靈敏性及特異性分別為(87.8%、88.1%、89.1%)、(87.5%、87.8%、93.3%)與(88.1%、88.3%、84.8%)；VGG19對T1、T2與RGB影像之準確度、靈敏性及特異性分別為(83.6%、87.6%、87.4%)、(84.7%、86.9%、86.4%)與(82.5%、88.3%、88.3%)；ResNet50對T1、T2與RGB影像之準確度、靈敏性及特異性分別為(88.9%、90.2%、90.5%)、(88.1%、90.6%、90.3%)與(89.8%、89.8%、90.7%)。 透過影像深度學習對二維MRI T1和T2腰部脊椎骨折加權影像做辨識是可行，其中又以使用ResNet50對RGB影像的準確度為佳。因此使用T1、T2和T1與T2融合而成的RGB影像，具有提高分類鑑辨準確度之效果。
並列摘要
MRI is better than CT in distinguishing whether a spine fracture has caused disc herniation. Therefore, MRI is often used in clinical practice to differentiate spinal diseases. In sagittal, coronal, and transverse Among the (Transverse) sections, the sagittal section has the most diagnostic value. However, there are many slices and many details. In order to shorten the viewing time and improve the computer-aided diagnosis, deep learning is used to establish a discriminant model of MRI image spine fractures. This study used a retrospective study to collect a total of 213 cases from April 2019 to January 2020, including 108 spinal fractures (experimental group) included 1199 two-dimensional images and 105 in vertebral fractures (control group) included 1199 two-dimensional images. The MRIT1 and T2 weighted sagittal section images of the experimental group and the control group were obtained.DICOM format image whose size is 448 × 224, each MRI T1 and T2 sagittal section number approximately from about 8-11. and the performance evaluation of the resolution of CNN models such as AlexNet,VGG19 and Resnt50 was performed in this work. Establish CNN image classification framework using Transfer Learning. There are also three modes for analyzing images, including 2D T1 images, 2D T2 images, and T1, T2 and T1 and T2 merged into RGB three-channel images. Therefore, there are a total of 9 classification models. In all models, the training set accounts for 70%, that is, 839 and 801 MRI images of the experimental group and the control group, respectively; the test set accounts for 30%, which is the MRI image of the experimental group and the control group. 360 and 343 images. The test set classification results show that AlexNet's accuracy, sensitivity and specificity for T1, T2 and RGB images are (87.8%, 88.1%, 89.1%), (87.5%, 87.8%, 93.3%) and (88.1% , 88.3%, 84.8%). The accuracy, sensitivity and specificity of VGG19 for T1, T2 and RGB images are (83.6%, 87.6%, 87.4%), (84.7%, 86.9%, 86.4%) and (82.5%, 88.3%, 88.3%) ). ResNet50's accuracy, sensitivity and specificity for T1, T2 and RGB images are (88.9%, 90.2%, 90.5%), (88.1%, 90.6%, 90.3%) and (89.8%, 89.8%, 90.7%). It is feasible to identify two-dimensional MRI T1 and T2 lumbar vertebral fracture weighted images through image deep learning, and the accuracy of RGB images using ResNet50 is better. Therefore, the use of RGB images fused by T1, T2, and T1 and T2 has the effect of improving classification accuracy.