AUTOMATIC BRAIN INFARCTION DETECTION SYSTEM ON MAGNETIC RESONANCE IMAGING AND OPERATION METHOD THEREOF

專利類型

發明

專利國別 (專利申請國家)

美國

專利申請案號

17/092,348

專利證號

US 11,042,983

專利獲證名稱

AUTOMATIC BRAIN INFARCTION DETECTION SYSTEM ON MAGNETIC RESONANCE IMAGING AND OPERATION METHOD THEREOF

專利所屬機關 (申請機關)

臺北醫學大學

獲證日期

2021/06/22

技術說明

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。在這項研究中，我們提出了一種基於深度 學習的方法，過程中添加了Apparent Diffusion Coefficient (ADC) map 來增加判斷的依據，由於 DWI 與ADC map 並非同時間拍攝，可能會產生病人頭骨方向不一致的問題，所以前處理的 第一步我們以DWI 為標準，將ADC map 對位至DWI，將影像二值化，提取腦組織部分， 濾除雜訊與頭骨。每個人大腦的影像強度具變異性，所以根據病人的不同，每個人梗塞的 影像強度值也有所不同，所以我們依據每個人影像強度與標準差，設置一門檻值，濾除大 腦中非梗塞的區域，保留偽影與梗塞部分。將保留的區域切割成數個小區塊並加入額外資 訊，再將數個小區塊經由(Convolutional neural network, CNN)去做辨別，CNN 辨別出各個小區 塊是否包含梗塞，最後將各個相鄰的小區塊合併，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 Infarct detection on cerebral MRI have been hindered by the histographic overlapping of image artifacts and infarct lesions. This study aimed at improving the detection accuracy by utilizing convolutional neural network trained by deep learning algorithm. The preprocess included the flowing 7 steps: First, image centralization; Second, registration of DWI to T1W; Third, registration of ADC to DWI; Fourth, z-axis registration of ADC to DWI; Fifth, skull masking; Sixth, intensity normalization of the cerebellar portion of excessive intensity; Seventh, removing DWI/ADC pixels with below-/above-threshold intensity. The CNN structure contained four convolutional layers that had 16, 32, 64 and 64 filters of size 4 × 4. A convolutional layer was followed by a batch normalization layer, a ReLU layer, max pooling 2d layer. The output layer is a classification layer preceded by a fully connected layer and a softmax layer. The network model was trained with deep learning using the stochastic gradient descent algorithm. The activation function ReLU was used for faster convergence and computational efficiency. The initial network weights were initialized randomly. An experiment to evaluate the developed CNN involed 15 stroke patients and 15 healthy persons. The cerebral infarcts and artifacts of the patients had been segmented by an experienced neurologist. Patches of 16 x 16 pixels were generated from all the subject MRIs. The training set consisted of 90% of the patches and the rest were used as the test set.

備註

連絡單位 (專責單位/部門名稱)

技術移轉中心

連絡電話

02-6638-2736-2006

網址