| 摘要: | 腦海綿狀血管瘤（cerebral cavernous malformations, CCM），是腦部中的一種血管病變，由良性不正常的血管組成，在腦部某一位置的血管膨脹成團。在T2權重影像上會有低信號hypointensity（黑色）的邊緣，本體可能因為反覆出血而呈現多囊狀有如爆米花(popcorn)般的型態。目前腦海綿狀血管瘤的診斷，主要依賴醫師的目視判讀及手動標記，但人眼判讀的方法容易受外界環境、視覺疲勞影響，手動標記耗時費力，因此擁有一客觀工具來改善診斷的準確性和效率為當前需求。本文提出一種深度學習方式，在T2權重影像上自動分割及量化腦海綿狀血管瘤。首先，使用Mask Region based Convolution Neural Networks (Mask RCNN)對T2權重影像進行實質腦提取，去除頭骨、頭皮以及背景雜訊，目的為提高分割效率於實質腦範圍內的腦海綿狀血管瘤，接著對影像進行強度標準化、體素尺寸重採樣以及資料增量等影像前處理，最後使用Deepmedic多尺度3D卷積神經網路在實質腦範圍進行腦海綿狀血管瘤的分割與量化。本研究使用的資料來源為臺北榮民總醫院192筆T2權重影像，資料被隨機劃分五分之三為訓練集、五分之一為驗證集及五分之一為測試集。目前訓練模型用於腦海綿狀血管瘤自動分割在測試集上取得的模型評估指標，平均Dice、精確率(Precision)、召回率(Recall)分別為0.736、0.807和0.729。此結果顯示了所提出的深度學習方法在自動腦海綿狀血管瘤分割方面的有效性。此系統的開發提供了一種客觀工具，以提高腦海綿狀血管瘤診斷的準確性和效率。;Cerebral cavernous malformations (CCM) are vascular abnormalities in the brain characterized by benign clusters of abnormal blood vessels. Magnetic resonance imaging (MRI) is a diagnostic tool used by physicians to detect and assess the size of CCM. In T2-Weighted (T2W) images, there may be hypointensity (dark) edges, and the lesion itself may appear as a multicystic structure resembling popcorn-like morphology, possibly due to recurrent seizures. Currently, the diagnosis of CCM heavily relies on visual interpretation and manual delineation by physicians. However, these methods are subjective, prone to environmental factors and visual fatigue, and time-consuming. Therefore, there is a current need for an objective tool to improve the accuracy and efficiency of diagnosis. To address these challenges, we proposed a deep learning-based approach for automated segmentation and quantification of CCM on T2W. Firstly, a Mask Region based Convolution. Neural Networks (Mask RCNN) model is employed to extract the brain region from the T2W, removing skull, scalp, and background noise to improve segmentation efficiency within the brain region. The images are then subjected to preprocessing steps including intensity normalization, voxel size resampling, and data augmentation. Finally, a Deepmedic multi-scale 3D convolutional neural network (CNN) is used to perform CCM segmentation and quantification within the extracted brain region. The dataset used in this study consists of 192 T2W from Taipei Veterans General Hospital, which are randomly divided into training (3/5), validation (1/5), and testing (1/5) sets. The trained model for CCM segmentation achieved the following evaluation metrics on the testing set: average Dice coefficient of 0.736, precision of 0.807, and recall of 0.729. The results demonstrate the effectiveness of the proposed deep learning approach in automated CCM segmentation. The developed system provides an objective tool to improve the accuracy and efficiency of CCM diagnosis. |
