A Multi-class Computer-Aided-Diagnosis System of Breast Ultrasound Image Based on BI-RADS Categories

• Main Authors: Chen,Yun-Ru, 陳韻如
• Other Authors: Lin,Wei-Yang
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/92hc8f

碩士 === 國立中正大學 === 資訊工程研究所 === 105 === Ultrasound image is not only a non-invasive instrument of less impact on human body for breast carcinoma detection, but also a basic tool to detect breast tumor. However, there exist some problem such as color difference and different resolution in image acquisition among different types of ultrasound imaging modalities so that clinicians always can’t identify accurately the BI-RADS categories or disease severities. In the study, three types of breast ultrasound imaging modalities including PHILIPS, SIMENS, and TOSHIBA were adopted to fetch breast ultrasound images for our experimental samples. Then, processing stages such as intensity normalization, contrast enhancement, and image segmentation were performed sequentially to detect true breast tumor. Our proposed system identifies the breast tumor severities according to the Breast Imaging-Reporting and Data System (BI-RADS) rather than traditional assessment on the severity, i.e. merely using benign or malignant. After segmentation, we focused on the BI-RADS 2-5 due to the clinical practice. And, several features related to lesion severities based on the selected BI-RADS categories were fed into three machine learning classifiers including Support Vector Machine, Random Forest, and Convolution Neural Network combined with feature selection to develop a multi-class assessment of breast tumor severity based on BI-RADS. Experimental results tested on BI-RADS samples reveal that the identification accuracies with SVM, RF, and CNN are 80.00%, 77.78%, and 85.42, respectively. In order to validate the performance and adaptability of classification using different ultrasound imaging instruments, the evaluations of F-score based on CNN obtain measures higher than 75% (i.e., prominent adaptability) when samples were tested on various BI-RADS categories. We hope that the Computer-aided diagnosis (CAD) system developed based on CNN can provide diagnostic references for surgeon in interpreting BI-RADS categories so that patients can obtain appropriate treatments or avoid unnecessary surgeries.