白質病變(White matter hyperintensity)是腦部疾病的一個預測因子，與中風、頑固性憂鬱症、癡呆和死亡的風險有顯著的相關性。白質病變的觀測需要對T2加權磁共振成像(T2W MRI)進行資料視覺化，讓白質病變以亮度增加的形式表現在影像上。本研究的目的是使用人工神經網路在T2加權磁共振成像快速且準確地自動分割白質病變，並且自動計算出病變的體積，輔助醫師的診斷及治療。 一般醫學影像辨識的AI輔助大多採用U net人工神經網路架構，本研究所採用的是運算速度更快且可以實例分割的Mask R-CNN，以兩階段神經網路的方式偵測。第一階段是以Mask R-CNN進行影像前處理，影像的前處理包括去除頭骨、眼球等經常是被人工神經網路誤判為白質病變的區域，只保留腦實質的部分作為後續輸入。因為頭骨與腦組織的影像強弱關係會受到機型與參數影響，本研究對留下的腦實質影像進行影像正規化，來降低不同參數與機型生成的磁共振成像差異，並提升白質病變與非白質病變組織的影像特徵差異。第二階是以Mask R-CNN對白質病變進行圖像分割，將分割的結果依據磁共振成像的像素間距計算白質病變體積的毫升數。人工神經網路的訓練資料來源是荷蘭UMC Utrecht醫院20名有白質病變的病人影像，測試資料來源是聯新國際醫院收集的10位有白質病變的病人影像，第一階段的訓練資料標記是基於FMRIB Software Library的Brain Extraction生成，並經過手動修正的二值化影像標記。 研究結果顯示使用Mask R-CNN進行腦提取的Dice係數達到96.06%。白質病變自動分割的Dice係數為55.69%，對白質病變自動分割的體積與實際體積計算相關性分析達到90.3%，運用迴歸分析計算量化白質病變的成果，判斷係數R2達到0.949。 研究結論為Mask R-CNN對於白質病變的分割及量化可以輔助醫師做為臨床參考，結合影像正規化與運用神經網路剔除干擾資訊的方法，可有效的提高自動辨識的準確率。;White matter hyperintensity (WMH) is a predictor of brain diseases and has a significant correlation with the risk of stroke, refractory depression, dementia, and death. White matter lesions are usually visible in T2-weighted magnetic resonance imaging (T2W MRI) as increased brightness. The purpose of this thesis is to use artificial neural networks to develop an automated algorithm for fast and accurate segmentation and quantification of white matter lesions with the aim of assisting physicians in diagnosis, prognosis and treatment. The artificial neural network architecture used in this research was the Mask R-CNN, which is capable of fast instance segmentation. The WMH segmentation was preceded by image preprocessing as the first stage of the entire process. The so-called brain extraction was conducted using Mask R-CNN to scrape and remove the images of the skull, eyeballs, etc., which would often be mistakenly identified as white matter lesions. Only the brain parenchyma was extracted for subsequent processing. Image normalization on the brain parenchyma was performed to compensate for the intensity difference caused by MRI acquisitions using scanners with different brands and parameters. The second stage was to segment the white matter lesions with Mask R-CNN and calculate the volume of the white matter lesions based on the pixel sizes. The training data were the MRIs of 20 patients with white matter lesions from UMC Utrecht Hospital, the Netherlands, which is open data for public use. The test data were the MRIs of 10 patients with white matter lesions collected in Landseed International Hospital, Ping-Jhen, Taoyuan. The brain masks for brain extraction were generated with the FMRIB software library’s brain extraction system and manually corrected to binarized masks. The research results show that the Dice coefficient of brain extraction using Mask R-CNN reached 96.06%. The Dice coefficient of the automated white matter hyperintensity segmentation was 55.69%. The correlation coefficient between the automatedly segmented volume and the actual volume reached 90.3%. The R2 value of the linear regression to predict the actual WMH volume from the automatedly segmented volume was 0.949. This study concludes that segmentation and quantification of white matter lesions using Mask R-CNN can attain a promising performance. Measures such as the extraction of brain parenchyma and intensity normalization have been helpful in eliminating the effect of interferences.
