Logo image
Automatic segmentation of breast ultrasound images for finding tumor regions by using a distance regularized level set evolution combined with texture feature-based initialization and post-processing
Thesis

Automatic segmentation of breast ultrasound images for finding tumor regions by using a distance regularized level set evolution combined with texture feature-based initialization and post-processing

Hsu, Yung Hsuan
Masters, 國立清華大學, 電機工程學系
2015

Abstract

超音波 乳癌 病理特徵 影像切割 距離規則化水平集演化 紋理分析 碎形維度 灰階共生矩陣 ultrasound breast cancer histologic characteristics image segmentation distance regularized level set evolution texture analysis fractal dimension GLCM
Breast cancer is the most common type of cancer in women worldwide, which inspires researchers worldwide to develop computer-aided-diagnosis with ultrasound images. In this thesis, we propose the research of combining texture features with tumor segmentation based on level set evolution in breast ultrasound image. Furthermore we try to find if the texture features coming from ultrasound images have correlation with histological features or some specific types of breast cancer. First a segmentation method, which combines the texture features and one rubout weak-edge segmentation algorithm, called combined distance regularized level set evolution (cDRLSE), is developed for capturing the contour of tumors. The proposed cDRLASE consists of the following steps. First, apply the texture features for support vector machine (SVM) to decide the initialization area. Second, re-initialize by improved DRLSE, where the edge indicator of external energy was improved so that the contour can easily and precisely reach to the weak boundary. Third, Gaussian filtering is used to smooth noise, and apply DRLSE for roughly capturing the tumor area. Last, with the post-processing to find a proper tumor area. Meanwhile, manual segmentation on BUS images under the supervision of Dr. Chou, an experienced doctor, are performed to define the contour separating the tumor and non-tumor region and used as ground truth contour for evaluation. Comparing the cDRLSE method with two other segmentation methods in Appendix, cDRLSE outperforms the other two in the average. Then, BUS tumor image texture features are extracted by gray scale intensity analysis and texture analysis. The image features includes mean, variance, skewness, kurtosis, entropy, fractal dimension and gray-level co-occurrence matrix. During the procedure of computing gray level co-occurrence matrix, various distance parameters are included, which lead us to build a large number of texture features. Finally, we investigate if there are any correlations between texture features extracted from breast ultrasound images, pathological features and the histological type by SPSS. According to the analysis, most of the features have no relation among histological type, except only histological grade and category show relations. Also, few histological features were found correlate with texture features; size, lymph nodes and tumor necrosis are the ones associate with some texture features.

Metrics

1 Record Views

Details

Logo image