| 摘要: | 腦血管疾病，包含腦中風，為台灣十大死因的第4位。急性缺血性腦中風，又佔腦中風人數的8成以上。目前主要治療手段，是在中風後3小時內注射靜脈血栓溶解劑。若超過3小時注射，可能造成腦內出血，導致生命危險，因此治療急性缺血性腦中風是刻不容緩的。臨床醫學上，腦部磁振造影(Magnetic Resonance Imaging, MRI)是診斷急性缺血性腦中風與判定腦部梗塞區域的主要工具，其中擴散權重影像(Diffusion-Weighted Imaging, DWI)對腦梗塞區域有高敏感度。
若要量化腦部梗塞體積，目前仍以醫師手動標定為主，但此法既費時又繁瑣。本文提出一種在DWI上快速且自動偵測腦梗塞區域的方法。該方法分成兩個階段，第一階段使用Mask R-CNN對DWI進行腦實質提取，濾除頭骨與背景雜訊；第二階段使用3D Multi-scale CNN對DWI腦實質範圍進行腦梗塞分割，多尺度的網路架構能同時學習梗塞的大致位置和細微的特徵。在訓練兩個網路模型前，皆會進行影像標準化、影像重採樣與資料增量等影像前處理。
本論文使用來自台北榮民總醫院218筆DWI，其中200筆執行5-fold交叉驗證，得到平均DSC、Precision、Recall分別為74.2%、77.5%、76.2%，並且剩餘18筆作為內部測試集。最終我們將5-fold交叉驗證得到的五個模型，應用於18筆測試資料上，得到平均DSC、Precision、Recall分別為74.6%、76.7%、74.8%。同時使用來自聯新國際醫院的66筆DWI作為外部測試集，並且在五個模型上得到平均DSC、Precision、Recall分別為68.9%、65.1%、77.8%。
;Cerebrovascular diseases including brain strokes are the fourth leading cause of death in Taiwan. Acute ischemic strokes account for more than 80% of brain strokes. To date, the major treatment method is to inject recombinant tissue plasminogen activator (rt-PA). This treatment must be done within 3 hours after brain stroke, because injection of rt-PA later than 3 hours post-stroke may cause life-threatening cerebral hemorrhages. Hence, an accurate diagnosis of acute ischemic stroke and prompt decision for urgent treatment are very important. In clinical medicine, magnetic resonance imaging (MRI) is the most powerful tool for visualizing stroke lesions to diagnose acute ischemic stroke.
To quantify stroke infarction from MRIs, the time-consuming and cumbersome manual labeling is still the main method clinically available. This paper proposes a method for rapidly and automatically detecting infarction on DWI, which is the most sensitive to brain infarct among different MR images. This method can roughly be divided into two stages. The first stage is to use Mask R-CNN to extract brain parenchyma from DWI to eliminate the skull and extracranial noise. The second stage is to use 3D multi-scale CNN to segment brain infarction on DWI brain parenchyma. Multi-scale network architecture can learn both rough infarction positions and detail features. Before training the two neural network models, image preprocessing such as image normalization, image resampling, and data augmentation will be performed.
This study used 218 DWI scans collected from Taipei Veterans General Hospital. Among these scans, 200 were used for a 5-fold cross-validation, which resulted in a 74.2% average Dice similarity coefficient (DSC), a 77.5% average precision, and a 76.2% average recall. The remaining 18 scans were used as the internal test set to test the five generated models. The internal test resulted in a 74.6% average DSC, a 76.7% average precision, and a 76.8% average recall. Furthermore, these models were tested with 66 DWI scans from Landseed International Hospital as an external test set. The average results of DSC, precision, and recall rate were 68.9%, 65.1%, and 77.8%, respectively. |
