基于深度学习与声影校正的全乳房超声解剖层分割新方法研究

Whole breast ultrasound image can provide complete breast anatomic information and then potentially increase the detection rate of breast lesions. However, the reading and understanding of the whole breast image data is far more difficult than traditional B-mode ultrasound and may also lower down the efficiency of diagnostic procedure. To improve the efficiency of the clinical image reading and diagnosis, this study aims to develop an automatic segmentation method to partition the whole breast ultrasound into four layers of "fat", "mammary gland tissue", "muscle", and "rib/chest wall". The clinical applications of the segmentation of four layers can be at least twofold. First, the segmentation can provide quantitative volume measurements of the specific layers for the calculation of breast density. Second, the segmentation result can be served as the prior knowledge to exclude clinically implausible false-positive findings in the computerized breast lesion detection scheme for the 3D whole breast ultrasound. In such cases, the segmentation of the four breast anatomic layers is quite valuable in the clinical practice. The segmentation scheme is proposed to be realized in three steps: (1) construction of breast anatomic recognition models with deep learning techniques to identify each specific layer in the image as the pre-segmentation; (2) detection of the locations of acoustic shadows in the image; (3) accurate segmentation and shadow correction with the region competition framework that can be initialized with the results of steps (1) and (2). The first step employs the deep learning techniques to address the high intra-class variation problem in the pattern of each anatomic layer. The result of the first step can be served as a reliable pre-segmentation for the third step. The second step will provide the prior location knowledge of acoustic shadows to prevent over-correction in the normal regions in the process of shadow correction. The third step then adopts the iterative segmentation and shadow correction fashion to seek the satisfactory segmentation result with high accuracy. The segmentation of four anatomic layers in the whole breast ultrasound image is a brand new problem with the advance of this cutting-edge imaging technique. The corresponding research methods in this proposal are original and innovative. The research outcomes can potentially be shaped into a computerized tool for the analysis of the whole breast ultrasound images. Meanwhile, the research experience exploited in this proposal can be accumulated and shared as the referential knowledge for the future studies on image segmentation and pattern recognition in the 3D ultrasound image.

全乳房超声图像由于可提供完整的乳房解剖信息，而有潜力能大大提高病灶检出率。但是全乳房数据的图像判读诊断更加的困难，也使得诊断效率低下。为了提升全乳房超声图像的分析，本项目拟研发全自动乳房解剖层分割算法，将全乳超声分割为脂肪、乳腺组织、肌肉、肋骨与胸腔壁等四层。分割结果可用于计算乳房密度和减少计算机辅助乳腺病灶检测的假阳性，具有极大的临床价值。本项目拟采用以下步骤进行分割：1）使用深度学习技术进行四个解剖层的预分割，来对付解剖层样式高类内差异的问题；2）声影区域检测，提供声影位置先验知识，避免正常区域过度校正；3）基於步骤1-2的结果以区域竞争算法为框架，借由分割与声影校正迭代的手段，求得精确的分割结果。全乳超声解剖层分割为全新的课题，本项目提出的一系列研究方案具有原始初创性，可作为临床全乳超声分析工具，也可为三维超声图像分割与模式识别等科学问题提供知识积累。

全乳房超声图像由于可提供完整的乳房解剖信息，而有潜力能大大提高病灶检出率。但是全乳房数据的图像判读诊断更加的困难，也使得诊断效率低下。为了提升全乳房超声图像的分析，本项目拟研发全自动乳房解剖层分割算法，将全乳超声分割为脂肪、乳腺组织、肌肉、肋骨与胸腔壁等四层。分割结果可用于计算乳房密度和减少计算机辅助乳腺病灶检测的假阳性，具有极大的临床价值。本项目拟采用以下步骤进行分割：1）使用深度学习技术进行四个解剖层的预分割，来对付解剖层样式高类内差异的问题；2）声影区域检测，提供声影位置先验知识，避免正常区域过度校正；3）基於步骤1-2的结果以区域竞争算法为框架，借由分割与声影校正迭代的手段，求得精确的分割结果。全乳超声解剖层分割为全新的课题，本项目提出的一系列研究方案具有原始初创性，可作为临床全乳超声分析工具，也可为三维超声图像分割与模式识别等科学问题提供知识积累。