Stress urinary incontinence (SUI) and pelvic organ prolapse (POP) are important health issues affecting millions of American women. Investigation of the cause of SUI and POP requires a better understand of the anatomy of female pelvic floor. In addition, pre-surgical planning and individualized treatment plans require development of patient-specific three-dimensional or virtual reality models. The biggest challenge in building those models is to segment pelvic floor structures from magnetic resonance images because of their complex shapes, which make manual segmentation labor-intensive and inaccurate. In this dissertation, a quick and reliable semi-automatic segmentation method based on a shape model is proposed. The model is built on statistical analysis of the shapes of structures in a training set. A local feature map of the target image is obtained by applying a filtering pipeline, including contrast enhancement, noise reduction, smoothing, and edge extraction. With the shape model and feature map, automatic segmentation is performed by matching the model to the border of the structure using an optimization technique called evolution strategy. Segmentation performance is evaluated by calculating a similarity coefficient between semi-automatic and manual segmentation results. Taguchi analysis is performed to investigate the significance of segmentation parameters and provide tuning trends for better performance. The proposed method was successfully tested on both two-dimensional and three-dimensional image segmentation using the levator ani and obturator muscles as examples. Although the method is designed for segmentation of female pelvic floor structures, it can also be applied to other structures or organs without large shape variation.
展开▼
