Multivariate Analysis of Tumor Shape Features for Ultrasound Breast Cancer Diagnosis

• Main Authors: Guo-Shian Hung, 洪國賢
• Other Authors: Huihua Kenny Chiang
• Format: Others
• Language: zh-TW
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/47974864261820453563

碩士 === 國立陽明大學 === 醫學工程研究所 === 88 === The implementation of ultrasound image diagnosis greatly depends on the experience of physicians. Therefore, the diagnostic process is strictly depends on the physicians, and is also lacking the experience sharing media and reference standard in the diagnostic process. In order to establish the experience sharing platform, in this research, we developed a computer-aided diagnosis system. To establish multivariate analysis model of tumor shape features for ultrasound breast cancer diagnosis, we evaluated 111 ultrasound images of solid breast nodules ( 71 infiltrative ductal carcinomas and 40 fibroadenomas ). The shapes of tumors were classified by expert radiologists, and a multivariate regression algorithm with least square method and cross-validation method was used to classify the tumor as benign or malignant. Another, the shapes of tumors were quantified by programs, and a logistic regression algorithm with the best prediction variables selection was used to classify the tumor as benign or malignant. The accuracy of multivariate regression model for classifying malignancies was 93.69%, the sensitivity was 90.14%, and the specificity was 100%. The accuracy of logistic regression model for classifying malignancies was 91%, the sensitivity was 97.2%, and the specificity was 80%. According to the experience of expert radiologists and regression analysis algorithm, we have established a computer-aided diagnosis system applied to ultrasound of solid breast nodules. This system differentiated solid breast nodules with relatively high accuracy and helped inexperienced operators to avoid misdiagnoses.