使用機器學習於大腸內視鏡影像結合人工特徵與機器特徵判別息肉惡性程度

Taipei Medical University Institutional Repository > 醫學科技學院 > 醫學資訊研究所 > 博碩士論文 > Item 987654321/58633

請使用永久網址來引用或連結此文件: http://libir.tmu.edu.tw/handle/987654321/58633

題名:	使用機器學習於大腸內視鏡影像結合人工特徵與機器特徵判別息肉惡性程度 Using colorectal endoscopic images in machine learning combined with artificial features and non-artificial features to determine the degree of polyposis
作者:	葉宇軒 Yeh, Yu-Hsuan
貢獻者:	醫學資訊研究所陳俊璋
關鍵詞:	大腸息肉;機器學習;深度學習 Colorectal polyps;Machine learning;deep learning
日期:	2019-06-27
上傳時間:	2020-02-11 11:41:08 (UTC+8)
摘要:	根據衛福部國健署公布的台灣癌症登記報告，2014至2016年癌症死亡人數以大腸癌1萬9335人最多，大腸癌長年以來是台灣癌症死亡榜首。大腸息肉是直腸內常見的增生性細胞，如未經治療切除有機會發展成直腸癌，大腸息肉分為四種，增生性息肉、贅生性息肉、發炎性息肉以及誤生息肉。在臨床上，要確定息肉是哪種類型通常都必須做切片檢查，此研究希望透過影像特徵分析結合機器學習，快速的從大腸鏡的影像中直接判別息肉種類，以減少病患因為開刀傷口所引起的併發症。在機器學習方面，我們有使用人工特徵提取來進行分類模型的建造，也有使用機器特徵當中的深度學習卷積神經網路，來比較這兩者所建造出分類模型差異。我們的影像來自台北醫學大學附設醫院消化內科大腸鏡患者的臨床資料，我們總共蒐集了1881位大腸鏡臨床患者的影像，共有1991張影像，其中增生性息肉有1053張、腺瘤狀息肉及腺癌共938張。我們在人工特徵值的Gabor配合21種機器學習中，以Ensemble Subspace KNN達到了最高93.2%準確率。而在機器特徵深度學習Resnet-101方面我們取得98.4%的更高準確率。 According to the Taiwan Cancer Registration Report published by Taiwan health promotion administration ministry of health and welfare, the number of new cancers in 2014 was the highest in colorectal cancer, with a total of 19,576 people. In the cancer ranking, colorectal cancer ranked first in cancer for the ninth time. A colorectal polyp is a polyp occurring on the lining of the colon or rectum. Untreated colorectal polyps can develop into colorectal cancer. And to determine which type of polyp is usually necessary to do a biopsy. Colorectal polyps are divided into four types, hyperplastic polyps, neoplastic polyps, inflammatory polyps, and misdiagnosed polyps. It is more difficult to determine the type of polyp in the polyp of patients with colorectal black lesions in the clinic. This study hopes to quickly identify polyp types from colonoscopy images through image feature analysis combined with machine learning. This is used to reduce the complications of the patient because of surgery. In machine learning, we use artificial feature extraction to build the classification model, and also use deep learning among non-artificial features to compare the difference between the two models. Our images were obtained from the clinical data of patients with gastroenterology in the Department of Gastroenterology, Taipei Medical University. There were 1991 images, including 1053 hyperplastic polyps, 938 adenomatous polyps and adenocarcinomas. We use the artificial eigenvalue Gabor with 21 machine learning algorithms, of which Ensemble Subspace KNN has 93.2% with the highest accuracy. In the deep learning of machine features, we achieved a higher accuracy rate, and achieved a higher accuracy rate of 98.4% in Resnet-101.
描述:	碩士指導教授：陳俊璋委員：邱泓文委員：羅崇銘
資料類型:	thesis
顯示於類別:	[醫學資訊研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	80	檢視/開啟

檢視Licence

在TMUIR中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....