SEMINAR 
School of Medical Instrument and Food Engineering, USST

題目：Automatic Brain Metastasis Detection by Convolutional Neural Network

基于卷積神經網絡的腦轉移自動檢測

報告人：Prof. Jyh-Cheng Chen（陳志成教授）

National Yang-Ming University, Taipei, Taiwa（臺灣陽明大學）

時間：2019年5月15日（周三）, 上午10:00

地點：上海理工大學醫療器械與食品學院，綜合樓C區3樓302會議室

報告人簡介：陳志成，博士，臺灣陽明大學教授，生醫影像與放射系主任。1983年獲臺灣中央大學 (又稱中國南京大學) 物理學學士學位, 1988年和1995年獲得美國亞利桑那州大學物理學碩士學位和光學科學博士學位。1995年在臺灣工業技術研究所光電和系統實驗室工作。1996年在臺灣陽明大學醫學技術系放射科學與技術系工作，任放射科學研究所副教授。2005年任臺灣大學生物醫學成像和放射科學 (BIRS) 教授, 從事分子成像物理和儀器領域的研究。目前的研究領域是圖像處理、分析和重建。包括使用micro PET、micro SPECT、micro CT進行圖像重建、處理和分析, 用于動物分子成像研究。自主設計制作了micro CT、FMTT和 PET/CT小動物成像系統。共發表學術論文100余篇, 多次撰寫專業著作章節，并擁有多項專利。獲得了臺灣陽明大學教師學術卓越獎等研究獎項。他是 IEEE、SNM 和 FASMI 的會士，擔任《國際生物醫學成像雜志》的編委委員、《計算機化醫學成像和圖形》特刊的特邀編輯、超過15份國際科學期刊的評審員，以及ANMMI和JRS的聯合編輯。

報告內容：Objective：Brain metastasis is the most common cerebral neoplasm in adults. Post-contrast MR has the best detectability to detect brain metastasis among all imaging modalities, yet recognition of the pathology is cognitive demanding even for experts. The purpose of our study was to use convolutional neural network (CNN) algorithm to facilitate automatic detection of metastasis in MR images. Materials and methods: Fifty metastatic patients with post-contrast axial T1 weighted MR of 3-mm slices to cover the whole brain were used as the whole dataset. Metastasis lesions were manually annotated by an experienced radiologist and served as the gold standard. Two-dimensional slices were cropped and normalized, and then were transferred to CNN structure for training. Various internal structures were attempted with different filter size. We tried some parameters such as 28 layers and 24 batch size. We chosed the exponential linear unit as activation function. The initial learning rate was 0.05, and decayed by 0.6 every 200 steps. The cost function consists of mean square error and Dice coefficient. The performance was evaluated with Dice coefficient and area under ROC curve (AUC). All the experiments were performed using Tensorflow in Python 3.5. Results: The CNN model with chosen parameters achieved a Dice score of 0.68 and AUC score of 0.84 on our testing data. Conclusion: CNN effectively facilitates the recognition of metastatic lesions in post-contrast MR images. Further evaluation with more datasets obtained by different protocols and institutions may improve its detection accuracy.

歡迎廣大老師、科研人員和研究生參加！