深度學習在醫學影像的應用使用Torch

Medical image recognition based on Deep Learning using Torch

黃晨誠(Chen-Cheng Huang)

指導教授 ： 秦群立

中山醫學大學/健康管理學院/醫學資訊學系碩士班/碩士(2017年)

摘要

隨著近年醫療技術蓬勃發展，許多學者紛紛投入發展以醫學影像進行判斷的電腦輔助診斷系統，其中利用了許多過去傳統的分類器，並經由不斷的改變及調整來結合出一個較高精確的分類器，讓醫生除了可以依靠臨床的經驗外，也可以透過電腦輔助系統的分辨結果做為參考的依據。由於近年電腦科技有所突破，讓人工智慧中的深度學習逐漸成為主流且漸漸地應用在許多領域當中，如電腦視覺、圖形識別、語音識別及自然語言處理。
本論文中，我們提出了一個利用深度學習的特性使醫學影像在判讀上能有較高的辨識準確率，且透過統一計算架構(CUDA)的技術讓學習速度能有大幅度的提升，因此我們開發了「深度學習在醫學影像的應用使用Torch」系統，此系統除了應用NVIDIA的CUDA技術及圖形處理器外，我們也使用了卷積類神經網路應用在Facebook所提供的深度學習開發工具Torch，作為我們深度學習演算法的架構。此系統的醫學影像資料集則是透過中山醫學大學附屬醫院(Chung Shan Medical University Hospital)的醫學影像放射科醫生所提供，資料集分別為鼻咽癌良惡性腫瘤資料集、肺腺癌EGFR基因突變資料集及腦中風為是否有缺血性腦中風資料集，並利用資料擴增的方式提升影像數據量，輸入至Deep CNN深度學習分類器中進行學習。
本論文的貢獻為提出了利用深度學習應用在醫學影像，並且在鼻咽癌良惡性腫瘤資料集的分類精確度上達到68.75%、肺腺癌EGFR突變資料集的分類精確度達到97.656%及缺血性腦中風資料集的分類精確度達到92.968%。

關鍵字

深度學習 ； 卷積神經網路 ； 統一計算架構 ； 圖形處理器 ； 資料擴增 ； 鼻咽癌 ； 肺腺癌EGFR基因突變 ； 缺血性腦中風

並列摘要

Many scholars actively engaged in the development of the use of medical imaging-based computer-aided diagnostic system. In the past, these computer-assisted systems use traditional classifiers to perform medical image analysis and recognition Doctors can not only rely on clinical experience, but also through the computer-aided system analysis results as a basis for diagnosis. In recent years, computer technology has been a breakthrough. So, deep learning in artificial intelligence is becoming mainstream and gradually applied in many areas such as computer vision, pattern recognition, speech recognition and so on.
In this thesis, we propose a medical image recognition system based on deep learning using Torch system. Our system has a higher recognition accuracy, and through the CUDA technology it makes the learning speed can be greatly improved. The recognition algorithm of our system is convolution neural network (CNN). We use the Torch developed by Facebook to achieve CNN. The medical datasets is provided by chung shan medical university. It is divided into three parts: the dataset of benign and malignant tumor of Nasopharyngeal Carcinoma, the dataset of lung adenocarcinoma EGFR gene mutation, and ichemic stroke datasets. Finally, we use data augmentation method to increase the amount of image data.
The contribution of thesis is to apply deep learning method to analysis medical image. And in the Benign Nasopha-ryngeal Tumors and Malignant Na-sopharyngeal Tumors of Nasopharyngeal Carcinoma data set classification accuracy reached 68.75%, the classification accuracy of EGFR mutation data of lung adenocarcinoma is 97.656%, and the classification accuracy of Ischemic Stroke data set reached 92.968%.

並列關鍵字

Deep Learning ； Convolution Neural Networks ； CUDA ； NVIDIA GPU ； Torch ； data augmentation ； Nasopharyngeal Carcinoma ； Lung adenocarcinoma EGFR gene mutation ； Ischemic Stroke

參考文獻

[23] Yoganand Balagurunathan, Yuhua Gu, Hua Wang, Virendra Kumar, Olya Grove, Sam Hawkins, Jongphil Kim, Dmitry B. Goldgof, Lawrence O. Hall, Robert A. Gatenby and Robert J. Gillies, “Reproducibility and prognosis of quantitative features extracted from CT images,” Translational Oncology, vol. 7, pp. 72-87, February 2014.

[1] Weibo Liu, Zidong Wang, Xiaohui Liu, Nianyin Zeng, Yurong Liu and Fuad E., Alsaadi, “A survey of deep neural network architectures and their applications,” Neurocomputing, vol. 234, pp. 11-26, April 2017.

[2] J. Schmidhuber, “Learning complex, extended sequences using the principle of history compression,” Neural Computation, vol. 4, pp. 234-242, 1992.

[3] Y. LeCun, L. Bottou, Y. Bengio and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, pp. 2278-2324, 1998.

[10] Haoqiang Fan and Erjin Zhou, “Approaching human level facial landmark localization by deep learning,” Image and Vision Computing, vol. 47, pp. 27-35, March 2016.