Skip to main content
eScholarship
Open Access Publications from the University of California

Development of Segmentation Algorithms and Machine Learning Classification Methods for Characterization of Breast and Spine Tumors on MRI

  • Author(s): Wang, Xinxin
  • Advisor(s): Gulsen, Gultekin
  • et al.
Abstract

As the second most leading cause of cancer death in women, breast cancer has attracted wide attention in MRI studies. However, the segmentation of non-mass lesions in breast cancer MRI images remains challenging due to their comparatively high variations in kinetic characteristics and typical morphological parameters. Meanwhile, the segmentation of spinal tumors in axial and sagittal views has been tough task because of the complexity to determine lesion area in one single view.

In this thesis, we proposed a method for automatic non-mass tumor segmentation, which was threshold oriented and based on region growing algorithm. It turned out that the segmentation results were satisfactory where only less than 5% pixels compared to the total number of segmented tumor pixels were corrected. We also introduced a method for spinal tumor segmentation to determine the location of lesions in spinal cord cancers. Normalized cut and region growing were applied in this study so to locate spinal tumors. Seven types of spinal cancers were analyzed and the overall behavior of the method was good in most cases except for some cases due to influence of unclear boundaries of the lesion areas. Finally, we applied a machine learning based method to differentiate three subtypes of breast cancer tumors and three types of spinal cancers. ANOVA and random forest algorithm was utilized to select the most significant features and cross-validation-based logistic regression was then used in this step in order to build a classifier. The performance of the classifier can be improved if more cases are included and affecting factors like biomarkers in spinal cancers can be rejected from the classification process.

Main Content
Current View