深度學習於醫學影像處理之應用
摘要
近年來深度學習是個非常熱門的議題,許多事情都可以用深度學習來解決,本論文希望能將深度學習應用在醫學上,希望能幫助醫師節省判斷時間並降低誤判率,也可以讓就醫的人可以及早發現問題,及早治療。 本論文用CNN建立一套可以自動偵測心臟血管鈣化系統,系統主要有三個功能。一、結合影像處理的方法,擷取出心臟血管影像;二、利用CNN架構來訓練偵測鈣化的模型;三、使用影像處理的方法來定位鈣化斑塊的所在位置。 另外本論文也用Mask R-CNN建立另一套偵測脊椎壓迫性骨折的系統,系統主要有兩大功能。一、利用Mask R-CNN訓練分割脊椎骨影像的模型;二、利用Mask R-CNN訓練偵測四個點頂的模型,並用結果來評估是否為壓迫性骨折。 本系統的心臟血管擷取演算法有成功擷取出87%的血管,鈣化偵測準確率也有接近8成,誤判的部分每1張影像中僅會出現0.21個鈣化斑塊誤判,鈣化位置定位效果也有7成以上,這些數據說明本論文所使用的方法能有效的找出心臟血管鈣化斑塊及鈣化位置。 分割脊椎骨影像演算法能夠將97%以上的脊椎骨影像分割出來,找脊椎骨影像上的六個頂點正確率達到96%,偵測脊椎壓迫性骨折形狀及程度正確率也都有達到7成以上,應能夠輔助醫師有效的偵測出脊椎壓迫性骨折。 關鍵字:醫學影像、影像處理、深度學習、鈣化偵測、壓迫性骨折偵測

關鍵字: Kinect、視覺障礙者、卷積神經網路、物品偵測

 

 

Applications of Deep Learning in Medical Image Processing
Abstract
In recent years, deep learning is a very hot topic. People can solve a lot of problem with deep learning. This paper hopes to apply deep learning in medical area, hoping to help doctors save their judgment time and reduce misjudgment rate. Thus doctors can help patient to find diseases earlier and treat it earlier. This paper uses Convolutional Neural Network to build a system that can automatically detect calcified plaque in cardiac blood vessels. The system has three main features. First, extract cardiac blood vessels images by using image processing methods. Second, use Convolutional Neural Network to train a calcified plaque detection model. Third, locate calcified plaque by using image processing methods. Moreover, this paper also builds another system for detecting vertebral compression fractures by Mask R-CNN. The system has two major features. First, train a model which can separate spines images from vertebral medical images by using Mask R-CNN. Second, train another model to detect the four vertex by using Mask R-CNN. Finally using these results to do the vertebral compression fractures detection. The cardiac blood vessels extraction algorithm of this system has succeeded extract 87% of cardiac blood vessels completely, and the accuracy of calcified plaque detection almost reach 80%. The misjudgment is about 0.21 calcified plaques per image. And also location accuracy of calcified plaque is more than 70%. For these tests proves that the method we used in this paper can efficiently find the calcified plaque on cardiac blood vessels. The segmentation of vertebral images algorithm can separate 97% or more of vertebral bones images, and also find 6 vertices of each vertebral bones images correctly. 70% or more of vertebral compression fractures has been detected by the system. These results indicate that the system can assist doctors to detect vertebral compression fractures efficiently. Keywords: medical image, image processing, deep learning, calcified plaques detection, vertebral compression fractures detection