| 摘要: | 聽神經瘤、腦膜瘤和轉移性腦瘤是常見的腦部腫瘤。聽神經瘤和腦膜瘤常見的治療方式為加馬刀放射治療，但是術後容易產生水腫，而轉移性腦瘤容易導致腦部產生水腫，因此水腫的觀察也是很重要的事情。目前針對腫瘤以及水腫都是透過醫師肉眼判斷，但這種判定方式不僅耗時且存在醫師主觀上判定所產生的誤差，目前沒有客觀且準確的工具來進行辨識，所以不利於水腫消長過程、量化嚴重程度和病例差異性的研究。
我們的研究是利用深度學習針對聽神經瘤的腫瘤和水腫以及腦膜瘤和轉移性腦瘤的水腫進行分割和量化。聽神經瘤腫瘤的部分，我們使用的是T1-weighted post-contrast- enhancement (T1C)權重影像，且利用DeepMedic進行腫瘤的分割和量化。聽神經瘤、轉移性腦瘤和腦膜瘤的水腫部分，我們使用的是T2權重影像(T2-weighted image)，先利用Mask R-CNN (Region Convolutional Neural Network)將腦遮罩提取出來，再利用DeepMedic進行水腫的分割及量化。
聽神經瘤腫瘤的部分，資料集是44位病患，44筆資料，在經過五折交叉驗證後，我們的模型取得了平均Dice係數 91.9 %。腦遮罩提取的部分，資料集是60位病患，60筆資料，在經過五折交叉驗證後，我們的模型取得了平均Dice係數 94.3 %。水腫的部分，聽神經瘤水腫資料集是10位病患，44筆資料，在經過五折交叉驗證後，聽神經瘤水腫平均Dice係數為55.2%；轉移性腦瘤水腫資料集是33位病患，66筆資料，在經過五折交叉驗證後，轉移性腦瘤水腫平均Dice係數為83.6%；腦膜瘤水腫資料集是20位病患，130筆資料，在經過五折交叉驗證後，腦膜瘤水腫平均Dice係數為76.8%。我們也開發了圖形使用者介面，使醫師能在臨床上方便使用。
這個研究可以實現自動化分割及量化聽神經瘤腫瘤及其水腫，以及腦膜瘤和轉移性腦瘤的水腫，幫助醫師判斷水腫消長過程、量化嚴重程度和病例差異性的研究，使醫師在決定治療方向時有更準確及客觀的數據提供參考，進而增進醫師的診斷效率以及準確率。最後也有開發圖形使用者介面，使醫師方便操作。
;Acoustic neuroma, meningiomas, and metastatic brain tumors are prevalent brain tumors. Gamma knife radiation therapy is a common treatment for neuromas and meningiomas, but postoperative edema is a usual complication. Metastatic brain tumors often lead to brain edema, making the observation of edema crucial. Currently, tumor and edema assessment rely on visual inspection by physicians, which is time-consuming and subject to subjective errors.
This research focused on deep learning for the segmentation and quantification of tumors and edema associated with acoustic neuroma as well as the segmentation and quantification of edema associated with meningiomas and metastatic brain tumors. The tumor segmentation of acoustic neuroma was conducted on T1-weighted post-contrast- enhancement (T1C) images using DeepMedic. The edema segmentation of acoustic neuroma, metastatic brain tumors, and meningiomas was conducted on T2-weighted images by first extracting the brain mask using Mask R-CNN and then applying DeepMedic for edema segmentation and quantification.
For the auditory nerve tumor segment, our dataset comprised 44 patients with 44 data points. After undergoing five-fold cross-validation, our model achieved an average Dice coefficient of 91.9%. Regarding the brain mask extraction, our dataset consisted of 60 patients with 60 data points. After five-fold cross-validation, our model achieved an average Dice coefficient of 94.3%. Concerning the edema segment, the auditory nerve tumor edema dataset included 10 patients with 44 data points. After five-fold cross-validation, the average Dice coefficient for auditory nerve tumor edema was 55.2%. For metastatic brain tumor edema, with a dataset of 33 patients and 66 data points, the average Dice coefficient after five-fold cross-validation was 83.6%. Lastly, for meningioma edema, with a dataset of 20 patients and 130 data points, the average Dice coefficient after five-fold cross-validation was 76.8%. We have also developed a graphical user interface to facilitate easy clinical use by physicians.
This study enables the automated segmentation and quantification of auditory nerve tumor and its edema, as well as edema associated with meningioma and metastatic brain tumors. It aids physicians in assessing the progression, quantifying severity, and studying case variations in edema. The research provides more accurate and objective data for physicians to reference when determining treatment directions, ultimately enhancing diagnostic efficiency and accuracy. Additionally, a graphical user interface has been developed to facilitate convenient operation for physicians. |
