基于超像素分析和无监督聚类模型的乳腺肿瘤超声图像分割研究

The computer-aided diagnosis based on breast ultrasonic images can help doctors avoid missed diagnosis and misdiagnosis, make it easier to find breast tumors, and enable the medical process optimization and automation. Focusing on the analyzing the breast lesions in ultrasound images, this project proposes an automatic tumor segmentation for joint superpixel clustering and unsupervised finite mixture model. To overcome the artifacts in ultrasonic images, i.e. noise, intensity inhomogeneity, low contrast, the shadowing effect, blurred boundaries, we explore the kernel function method which is superior to traditional Euclidean distance, and propose a compact superpixels generation model that adhere well to object boundaries. The results of superpixel clustering are then utilized to provide multi-scale superpixel feature representation for the collaborative optimization model. To break the bottleneck that the existing models depends on the shape of the breast lesion, the clustering algorithm based on unsupervised finite mixture model is proposed, and the sub-region information with strong correlation in the superpixel image obtained above is studied. According to the clustering results, the loss function of the collaborative optimization model is constructed. Finally, the above two processes are integrated into a single model to realize the collaborative updating of data clustering and depth features, and to construct a more accurate data clustering optimization model, aiming at solving the problem of lesion segmentation in breast ultrasound images. This project will vastly promote the theoretical development and extensive applications of superpixel clustering and unsupervised finite mixture model, which is very meaningful and valuable in both theory and practice.

基于乳腺超声数据的计算机辅助诊疗有助于帮助医生避免漏诊、误诊，更加便捷地探测乳腺肿瘤,使得医疗流程智能化。本项目针对乳腺超声图像中病变分析问题，研究基于超像素分析和无监督聚类的协同优化模型的乳腺肿瘤超声图像分割方法。针对乳腺超声图像中噪声、灰度不均匀、低对比度、阴影效应和模糊边界等缺陷，探究优于欧式距离的核函数方法，提出一种能够粘附目标边界的紧致的超像素生成模型，提高协同优化模型对病灶区域边界的鉴别力。突破算法对病变形态的依赖，深入开展基于无监督混合模型的聚类算法研究，挖掘上述得到的超像素图像中关联性强的子区域信息，根据聚类结果构造协同优化模型的损失函数。最后，将上述两个过程集成到单个模型中，实现数据聚类和深度特征的协同更新，构造具有强大特征表示能力的数据聚类优化模型，旨在解决乳腺超声图像中病变区域的分割问题，为临床诊断提供有效的分析依据。

本项目以临床需求为先导, 提炼科学问题, 着重解决超声影像数据的乳腺肿瘤病灶探测问题. 由于超声影像数据自身存在缺陷, 如噪声、低对比度、模糊边界等，以及病理征象的复杂性, 导致现有算法不能够准确探测到病灶边界. 本项目围绕关键科学问题，设计包括如何鉴别乳腺超声图像病灶区域、背景区域以及模糊区域? 这里的模糊区域所包含的像素点可能属于病灶区域，可能属于背景区域. 如何从模糊区域中挑选属于病灶的像素? 项目研究内容紧密联系, 相互支持. 项目研究成果为高质量学术论文, 研究成果已超出预期研究成果的论文数目, 现在已发表高质量学术论文3篇, 正在审稿论文4篇...项目研究的解决需要新的数学思想和新的数学方法和技术。无论是在理论上还是临床 应用方面均具有十分重要的理论意义和应用价值。随着大数据，云计算和人工智能等技术 的飞速发展，其在医学影像分析领域也得到了不同程度的应用。如何在影像人工智能领域 中提升价值，这将为临床医生和相关学科领域的科研人员，如，计算机科学，信息科学， 数据科学和数学等领域的科研人员，提供新方向，同时需要理工医结合沉下心来针对每个 场景做踏踏实实的创新和积累。深入开展基于乳腺肿瘤超声数据的分割方法研究具有广泛 的发展可能和应用前景，并将对相关学科及高、新技术的发展带来积极影响。