| 摘要: | 磁振造影（Magnetic Resonance Imaging,MRI）越來越多地用於診斷腦組織病變，特別是擴散加權(Diffusion Weighted Image,DWI)在檢測缺血性腦中風中具有高度敏感性。然而，MRI會因設備，非自主運動，磁化率及金屬等的產生多種偽影，而這些偽影在DWI影像中的直方圖與梗塞有很大的重疊區域，容易使醫生判斷病人時造成誤判。對於日常臨床使用而言，醫生仍然需要手動或半手動地對腦區域中的病變進行評估,手動評估病理變化太麻煩且耗費時間，也難免受到個人主觀性的影響，然而目前為止提出的數種全自動分割方法中去除偽影與正確率仍保有很大的進步空間。在此項研究中，我們提出了一種基於深度學習的方法，首先取出影像的腦組織，經由旋轉T1加權影像(T1-Weighted Image,T1WI)找出最對稱的旋轉角度，然而將T1 map移位至影像正中心，再將DWI與表觀擴散係(Apparent Diffusion Coefficient,ADC) map對位至T1 map。依每張切片的影像強度，設立自適性門檻值，濾除大部分頭骨與雜訊。判斷小腦的影像平均強度是否與大腦平均強度是否相似，若不相同則會調整小腦平均強度。由於每個人大腦影像強度存在著變異性，所以我們將利用DWI map做線性回歸的方法，提取出梗塞機率較高的區域，並將非線性回歸所得的影像，作為第三種參考依據。將之前對位的DWI與ADC map和非線性回歸影像重疊，再將重疊的影像切割成數個小影像並加入切割的位置資訊、整個大腦的影像平均強度、第幾張切片，最後利用深度學習排除偽影與非梗塞並偵測出梗塞的位置，供醫生參考。該研究對於偽影的辨識，與醫生手動評估的速度上，達到了很好的準確率與速度。;Magnetic resonance imaging (MRI) is highly sensitive to stroke lesions. However,
MRI can produce a variety of artifacts due to equipment, involuntary movement, magnetic susceptibility, and metal. The histograms of these artifact may overlap with those of infact in the DWI image, which causes infarct segmentation errors. In the traditional clinical infarct segmentation, the doctors need to use manual or semi-automated methods to detect infarct lesions in the brain area. These nonautomated methods to assess pathological changes is cumbersome, time-consuming, and easily influenced by the assessor’s personal subjectivity. There is still a lot of room for improvement in removing artifacts and accuracy in the automatic segmentation method. In this study, we propose a cerebral infarct segmentation method based on deep learning. First, the brain tissue of the image is extracted, and the T1-Weighted (T1-W) image is rotated at various angles to find the angle with respect to which the T1-W image is most symmetrical. Then the T1-W image is centralized by a shift, and the DWI and ADC map are registered to T1-W image. Based on the image intensity of each slice, an adaptive threshold is set to filter out most of the skull and noise. The average image intensity of the cerebellum is compared with that of the cerebrum. If the cerebellar intensity is higher than the cerebral intensity, the average intensity of the cerebellum is adjusted to the same level as the cerebral intensity. To accommodate the inter-person variability in the brain image intensity, a new image is generated as the third reference image in addition to the DWI and ADC map. This new image contains regions with a high probability of infarction extracted from the DWI by nonlinear regression. The previously aligned DWI, ADC image, and the non-linear regression image are all divided into patches of 16 × 16 pixels. Each patch is accompanied with the position information of the patch, the average intensity of the entire brain image, and the slice number. A convolutional neural network (CNN) is constructed and trained with the patches of over 30 patients. The trained CNN achieves a good performance in identifying infarct, noninfarct, artifact patches at a high speed. |
