- 學位論文
深度學習演算法應用於T1與T2大腦磁振融合影像進行腫瘤分類及分割之研究
Deep learning applied to MRI T1 and T2 fusion images for brain tumor classification and segmentation
摘要
臨床上常使用磁振造影(Magnetic Resonance Image)對腦腫瘤進行診斷，其常規執行的造影Sequence包含T1、T2，兩種影像所產生的特性不同，再者每一類腫瘤的細胞類型以及含量都不一樣，因此在對腦腫瘤進行診斷以及圈選腫瘤的邊界時，皆需要根據這兩種影像進行評估。本研究採用深度學習方法，將T1、T2 MRI影像進行融合對腦腫瘤進行分類與分割。 本研究採用回顧性實驗設計。共收集200個案例，包含100例正常及100例腦腫瘤，共有118位男性(55.2±13歲)及82位女性(55.4±14.6歲)。使用八種組合將T1、T2灰階影像融合為RGB三通道影像。便採用六種卷積神經網路模型加上轉移學習進行腦腫瘤分類，包含：VGG19、VGG16、Darknet19、Darknet53、Resnet50、Resnet101。最後經由混淆矩陣評估模型；使用準確度最高之影像組別採用全卷積網路進行腫瘤分割，其採用CNN(Convolutional Neural Network)模型做為骨架含：Resnet18、Resnet50、Xception以及Mobilenetv2。 模型分類結果顯示準確度最高為Resnet101，準確度、靈敏度、特異性、陽性預測值、陰性預測值、Kappa值為0.98、0.99、0.97、0.98、0.98、0.97；FCN模型以骨架Resnet50為最佳，得到之整體準確率、平均準確率、平均交疊率、加權交疊率、平均邊界F-1分數及Dice score為0.99、0.91、0.84、0.98、0.79、0.74。 由T1、T2融合而成的RGB影像不僅提供視覺上明顯標記，同時提供更多有用且顯著的影像特徵建立模型。此外，Resent101適合用於分類而Resnet50則適合分割MRI腦腫瘤影像。
並列摘要
Magnetic resonance imaging is often used clinically to diagnose brain tumors. The conventionally performed imaging sequence includes T1 and T2. The characteristics of the two images are different, and the cell type and content of each type of tumor are different. It is not the same, so when diagnosing brain tumors and ROI-selecting the boundaries of tumors, it is necessary to evaluate them based on these two images. In this study, deep learning methods were used to fuse T1 and T2 MRI images to classify and segment brain tumors. This study uses a retrospective experimental design. A total of 200 cases were collected, including 100 normal and 100 brain tumors. There were 118 males (55.2±13 years old) and 82 females (55.4±14.6 years old). Put the T1 and T2 grayscale images into the RGB three channels for image fusion, a total of eight combinations. Six convolutional neural network models plus transfer learning are used to classify brain tumors, including VGG19, VGG16, Darknet19, Darknet53, Resnet50, and Resnet101. Finally, the model is evaluated through the confusion matrix; the image group with the highest accuracy is used for tumor segmentation using a full convolutional network (FCN), the backbone convolutional networks used include Resnet18, Resnet50, Mobilenetv2, and Xception models. Model classification results show that the highest accuracy is Resnet101, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, Kappa value is 0.98, 0.99, 0.97, 0.98, 0.98, 0.97; The FCN with backbone Resnet50 was the best-segmented performance among all investigated methods. The global accuracy, mean accuracy, mean IoU, weighted IoU, mean boundary F-1 scores and Dice score were 0.99, 0.91, 0.84, 0.98, 0.79, and 0.74. The RGB image fused by T1 and T2 not only provides visually obvious marks but also provides more useful and significant image features to build models. In addition, Resent101 is suitable for classification and Resnet50 is suitable for the segmentation of MRI brain tumor images.