轉移性腦瘤是由其他部位的原發性腫瘤細胞轉移至腦部而形成的。轉移性腦瘤的存在也往往會導致腦水腫的形成。腦水腫是在細胞層級之下，細胞內或是細胞間液體累積，在巨觀之下造成腦組織腫脹並且功能異常。準確評估轉移性腦瘤及其引起腦水腫的位置、大小對於制定有效的治療策略至關重要。目前轉移性腦瘤及其引起腦水腫的評估通常依賴醫師的肉眼觀察，分別使用對比劑增強T1權重影像和T2權重影像。然而，這種主觀判定的方法不僅耗時還導致結果的不一致性。 本研究提出了一種基於深度學習的自動化分割模型，以實現轉移性腦瘤及其引起的腦水腫的準確辨識和量化。我們使用了Mask-region convolutional neural network (mask R-CNN)進行腦遮罩的提取，接著經由前處理和三維卷積神經網路DeepMedic進行轉移性腦瘤及其引起腦水腫的自動分割。 經過46位病患共90筆資料的五摺交叉驗證，我們的腦遮罩提取模型獲得平均Dice係數為96.4%，而轉移性腦瘤及其引起腦水腫分割模型平均Dice係數分別為71.6%與85.1%。我們亦開發一個友善的圖形使用者介面，使臨床醫師方便使用這些模型進行影像分析。 研究結果顯示，我們提出的基於深度學習自動化分割模型為轉移性腦瘤及其引起的腦水腫的準確辨識和量化提供了一個有前景的解決方案。模型的準確性和開發的使用者介面為臨床醫師提供了一個可靠且直觀的工具，有助於更準確地評估轉移性腦瘤和腦水腫的位置、大小提供治療策略的規劃。 ;Metastatic brain tumors are formed when cancer cells from primary tumors in other parts of the body spread to the brain. The presence of brain metastases often leads to the development of brain edema. Brain edema refers to the excessive accumulation of fluid in the brain tissue. Accurate assessment of the location and size of metastatic tumors and brain edema is crucial for effective treatment strategies. Currently, the evaluation of brain metastases and brain edema heavily relies on visual observation by physicians using T1C-weighted and T2-weighted images. However, this subjective approach is not only time-consuming but also prone to inconsistencies in results. In this study, we propose a deep learning-based automated segmentation model for the accurate identification and quantification of metastatic tumors and the associated brain edema. We employ Mask R-CNN for brain mask extraction, followed by preprocessing and the utilization of a 3D convolutional neural network called DeepMedic for automatic segmentation of metastatic tumors and brain edema. Through five-fold cross-validation on a dataset comprising 90 records from 46 patients, our brain mask extraction model achieves excellent results with an average Dice coefficient of 96.4%. The segmentation models for metastatic tumors and brain edema attain Dice coefficients of 71.6% and 85.1%, respectively. Additionally, we have developed an intuitive graphical user interface that enables clinical physicians to conveniently utilize these models for image analysis. The research findings demonstrate that our proposed deep learning-based automated segmentation model provides a promising solution for the accurate identification and quantification of metastatic tumors and associated brain edema. The accuracy of the models and the developed user interface offer a reliable and intuitive tool for clinical physicians, facilitating a more precise evaluation of the location and size of brain metastases and brain edema, ultimately aiding in the selection of appropriate treatment strategies.
