Lung Nodule Detection Based on Deep Learning Technology

• Main Authors: Chen Han, 陳涵
• Other Authors: Lien Chen-Chang
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/5dhd73

碩士 === 中華大學 === 資訊工程學系 === 106 === With the advancement of artificial intelligence, computer aided detection (CAD) has been continuously developed to improve the diagnosis rate of diseases and reduce the burden on doctors. The mortality rate of lung cancer is one of the highest in cancer. If it can detect early malignant nodules in the lungs, it can effectively reduce the mortality of lung cancer. Because the lungs are a large organ in the human body, and the diameter of the tumor is very small, the detection is difficult because of the low detection rate and excessive nodule misjudgment. With the recent rise of deep learning and the development of hardware technology, deep learning can be applied to CAD. It has been proposed to use a Convolution neuron network (CNN) to detect the area where the nodule is located, but there are too many misjudgments. Some studies use two thresholds to reduce the number of misidentified nodule. Although there are good results, the setting of the two thresholds is too heuristic. It has also been proposed that Faster R-CNN plus 3-D CNN to detect nodules and reduce the number of nodule misjudgments. This paper proposes a technique for detecting lung nodules based on a deep learning architecture, using a two-stage deep learning network. The first phase of the deep learning network proposed 3-D CNN to achieve nodule detection, and the second phase of deep learning network proposed 3-D GoogLeNet to achieve nodule misjudgment reduction. In the experimental part, we use the dataset of LUNA16 to analyze the nodule detection rate and misjudgment. After the analysis of the experimental results, we can achieve 90.9% average detection rate with average of 10 false positives or 93.3% average detection rate with average of 16.1 false positives.