| 摘要: | 顱內動脈粥樣硬化為一種慢性進行性的血管退化疾病，其主要病理變化包括動
脈壁的結構性退化與膽固醇斑塊的沉積。疾病早期常見血管壁增厚，隨病程進
展可能造成血管腔狹窄，進而引發缺血性腦中風或腦部灌流不足等臨床後果。
若能透過 black-blood MRI 於早期即偵測出顱內動脈的病理變化，將有助於疾
病的早期診斷、風險評估與後續治療規劃。然而，由於顱內血管壁相較於周邊
血管更為細薄，影像辨識難度較高，因此提高偵測準確性仍是一項挑戰。本研
究旨在探討運用深度學習模型，對高解析度黑血 MRI 中血管與血管壁進行自動
化標記與預測的可行性。研究資料來自 21 位患者的黑血磁振影像，每位患者皆
包含 155 張切片（切片厚度為 1 mm，畫素解析度介於 0.22 至 0.5 mm）。影像
前處理階段，首先由人工標記基底動脈的中心線，進一步建立中軸線模型，並
於每 0.5 mm 處擷取垂直於中軸線的橫向切片（平均每位患者約 51.6 ± 19.09
張），再重新採樣為 0.1 mm 畫素解析度。接著由人工標記血管外壁，並透過線
性模型自動推估內壁位置。最終共計 1259 張橫切圖像，其中 839 張（66%）用
於訓練四種深度學習模型：UNet、ResUNet、RegUNet 與 TransUNet；234 張
（19%）用於驗證，186 張（15%）用於測試。模型預測結果以 IoU
（Intersection over Union）作為主要效能評估指標，進行模型表現分析與比
較。整體 IoU 表現為 TransUNet 的 0.8657 ± 0.0294 最好。;Intracranial atherosclerosis is a chronic, progressive vascular degenerative disease,
characterized primarily by structural degradation of the arterial wall and cholesterol
plaque deposition. In its early stages, vessel wall thickening is commonly observed.
As the disease progresses, it can lead to arterial lumen narrowing, potentially resulting
in ischemic stroke or cerebral hypoperfusion. Early detection of pathological changes
in intracranial arteries using black-blood magnetic resonance imaging (MRI) can
facilitate timely diagnosis, risk assessment, and treatment planning. However, due to
the thinner walls of intracranial vessels compared to peripheral arteries, accurate
image identification remains a major challenge. This study aims to investigate the
feasibility of applying deep learning models to automatically segment and predict
vessel walls and lumens in high-resolution black-blood MRI. The dataset consists of
black-blood MR images from 21 patients, each comprising 155 slices (slice thickness
of 1 mm, pixel resolution ranging from 0.22 to 0.5 mm).In the preprocessing stage,
the basilar artery centerline was manually annotated and used to construct a central
axis model. Cross-sectional slices perpendicular to the axis were extracted at 0.5 mm
intervals (on average, 51.6 ± 19.09 std slices per MRI dataset) and resampled to a
pixel resolution of 0.1 mm. The outer vessel wall was manually labeled, and the inner
wall boundary was estimated using a linear model. A total of 1,259 cross-sectional 2D
images were obtained, with 839 images (66%) used to train four deep learning
models: UNet, ResUNet, RegUNet, and TransUNet; 234 images (19%) were used for
validation, and 186 images (15%) for testing. Model performance was evaluated using
the Intersection over Union (IoU) as the primary metric. Among the models,
TransUNet achieved the highest mean IoU score (0.8657 ± 0.0294). |
