顱內動脈粥樣硬化是一種動脈慢性退化的病理過程，在疾病早期可以看到血管壁增厚，最後血管狹窄可能會造成腦中風和腦缺血等疾病。如果能夠以黑血磁共振影像及早的辨識顱內動脈粥樣硬化病人中血管變化，對病程發展和治療都有非常重要的影響。因為腦血管管壁較身體其他地方的血管薄，較難被偵測到。所以在本文中，我們提出了一個新的線性模型自動計算出血管壁。在本研究中數據集使用了28個病患資料，每人有1組155個切片(厚度1 mm,畫素0.22到0.5 mm)的黑血磁共振影像。影像預處理先以人工標記基底動脈血管中心點產生中軸線模型，接著在每0.5 mm距離產生1446張(平均每人51.6+/-19.09 張，計算後填入)在基底動脈的二維連續與中軸線垂直的橫向切片，重新採樣為畫素0.1 mm。這些橫向切片數據再由人工手動標記外壁，最後以電腦程式的線性模型自動標記內壁。然後全部1446張的圖像分成1073 (80%)張用於訓練Detectron2和SE-ResNeXt50模型和373 (20%)張用於測試。兩個模型的平均IOU分別是0.75 +/- 0.096和0.73 +/- 0.091。最後測試的結果發現Detectron2沒標到血管壁2張(0.5%, 2/373) SE-ResNeXt50模型沒標到血管壁0張(0%, 0/373)。Detectron2對血管壁的偵測正確率略高於SE-ResNeXt50模型。;Intracranial atherosclerosis is a progressive pathological process in which thickening of the blood vessel wall can be seen in the early stages of the disease, and the luminal narrowing may lead to diseases such as stroke and cerebral ischemia. Early identification of vascular changes is very important in monitoring treatment effectiveness. By using Black Blood MRI to detect subtle changes as early as possible can potentially improve treatment outcome significantly. Because the walls of blood vessels in the brain are thinner than those in other parts of the body, they are more difficult to detect. In this paper, we propose a new linear model to automatically calculate the vessel inner wall automatically to ease the arterial wall labeling. In this study, 28 patients had black blood MRI performed with 155 slices at scanning thickness of 1 mm and planar pixel of 0.22 or 0.5 mm. The MRI images were preprocessed by manual annotation of the centerline of basilar artery (BA) and were resampled with 0.1 mm resolution at the planes perpendicular to the centerline at every 0.5 mm distance along a smoothened centerline. In total, 1446 (51.6 +/-19.09, std, per MRI dataset) 2D serial slices of the BA were generated. These resampled 2D images were labeled manually to define the outer wall of the BA. The inner walls of all images were automatically identified by using the linear model proposed in this study. Then all 1446 images were divided into 1073 (80 %) images for training Detectron2 and SE-ResNeXt50 models and 373 (20 %) images were reserved for test. The average IOU scores for the two models are 0.752 (+/- 0.096, std) and 0.733 (+/- 0.091, std) respectively. The results of the final test found that Detectron2 missed 2 out of 373 test images (0.5 %) and SE-ResNeXt50 missed 0 out of 373 test images (0 %). Detectron2 had a slightly higher accuracy rate in detecting BA walls.
