超声图像中前列腺分割方法的研究

发布时间：2018-07-14 19:27

【摘要】：前列腺癌是最常见的男性癌症之一。在欧美等国家,前列腺癌在男性癌症死亡率中排名第二位。在中国,前列腺癌死亡率近年来也呈现出上升的趋势。在前列腺疾病的诊断和治疗中,经直肠超声(Trans-rectal Ultrasonography,TRUS)图像由于具有实时性,且廉价无辐射等优点,而得到了广泛应用。在实际临床中,包括前列腺穿刺活检时穿刺针的精确定位,前列腺癌治疗中放射性物质的空间位置分布,以及测量前列腺的体积等,都需要从TRUS图像中精确分割出前列腺轮廓。但是,临床中常采用的手动分割方法存在工作量大、分割结果依赖操作者经验、重复性差等问题。TRUS图像中前列腺的自动分割方法研究为近年来的一个研究热点,但是超声图像中存在前列腺组织钙化的弱边界,严重的斑点噪声和低信噪比等因素,要精确地分割出前列腺的边界仍然是一项巨大的挑战。本文提出一种新方法,利用前列腺图像特征和先验统计形状从TRUS图像中自动地分割出前列腺边界。首先,利用图像的致密尺度和旋转不变特征变换(Dense SIFT),提取特征向量并进行分类,能够快速地从TRUS图像中定位前列腺,得到初始分割结果。其次,根据该初始结果从建立的多个平均模型中选择最优的形状模型作为分割的初始化轮廓。在分割过程中,前列腺伪影区域所缺失的边界信息就可以通过形状模型估计得到。最后,在该初始轮廓的指导下,利用灰度梯度特征模型和局部灰度高斯分布构建能量函数,使其最小化从而实现前列腺的精确分割。分割过程使用了一种由粗到细的多分辨率方式,提高了鲁棒性和计算的效率。通过大量实验得到的结果证实了该自动分割方法的有效性。本方法的分割结果与专家手动分割前列腺边界的平均绝对距离的均值是1.03±0.27毫米,豪斯多夫距离的均值是3.37±0.65毫米,Dice相似性系数的均值是91.9±0.8%,灵敏度的均值为94.7±2.1%,精确度的均值为99.2±0.2%。本文所使用的方法相比于改进的主动形状模型,在分割精度上有了较大程度的提高。
[Abstract]:Prostate cancer is one of the most common cancers in men. In Europe and the United States and other countries, prostate cancer in male cancer mortality rate ranked second. In China, prostate cancer mortality has also shown an upward trend in recent years. In the diagnosis and treatment of prostate disease, Trans-rectum Ultrasound (TRUS) image has been widely used because of its advantages of real time, low cost and no radiation. In clinical practice, including the accurate location of the puncture needle in prostate biopsy, the spatial distribution of radioactive material in prostate cancer treatment, and the measurement of prostate volume, it is necessary to segment the prostate contour accurately from TRUS images. However, the manual segmentation method often used in clinic has many problems such as heavy workload, the segmentation result depends on the operator's experience, and the repeatability is poor. The research on automatic prostate segmentation in TRUS image is a hot topic in recent years. However, there are some factors such as weak boundary of calcification of prostate tissue, severe speckle noise and low signal-to-noise ratio in ultrasound image, so it is still a great challenge to segment the boundary of prostate accurately. In this paper, a new method is proposed to automatically segment the prostatic boundary from TRUS images by using the prostate image features and prior statistical shapes. Firstly, feature vectors are extracted and classified by dense scale and rotation invariant feature transformation (Dense sift). The prostate can be located from TRUS image quickly and the initial segmentation results can be obtained. Secondly, according to the initial results, the optimal shape model is selected as the initial contour of the segmentation from the established multiple average models. In the process of segmentation, the missing boundary information of prostate artifacts can be estimated by shape model. Finally, under the guidance of the initial contour, the energy function is constructed by using the grayscale gradient feature model and the local gray Gao Si distribution, so that the prostate can be accurately segmented. The segmentation process uses a coarse-to-fine multi-resolution approach to improve robustness and computational efficiency. A large number of experimental results show the effectiveness of the automatic segmentation method. The average absolute distance between this method and the expert manual method is 1.03 卤0.27 mm, the average of Hausdorf distance is 3.37 卤0.65 mm, the average of Dice similarity coefficient is 91.9 卤0.8, the average of sensitivity is 94.7 卤2.1 and the average of accuracy is 99.2 卤0.2. Compared with the improved active shape model, the method used in this paper improves the segmentation accuracy to a great extent.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：R737.25;TP391.41

【共引文献】