基于解剖语义的医学超声图像分割与理解

Segmentation and classification of medical images are playing increasingly important roles in clinical computer-aided diagnosis. However, due to a large amount of speckle noises in ultrasound images, currently used techniques which highly rely on the visual features of the image itself cannot make precise segmentation and accurate classification for medical ultrasound images. This project innovatively proposes to introduce the anatomical semantic information of human physiological tissue structures to reconstruct 3D ultrasonic anatomical semantic models. Mimicking the understanding process of clinician when reviewing a medical ultrasound image, the proposed idea iteratively refines the segmentation results by a novel iterative scheme called “segmentation - classification - feedback learning”, and eventually achieves the classification and understanding of ultrasound images using the anatomical semantics. This idea takes into account conventional techniques for image segmentation and classification, incorporates a number of information processing technologies (e.g. the simulation of ultrasound images, image classification based on semantics, multi-objective optimization, and parallel computation) into the methods, and self-corrects the results of segmentation and classification by feedback learning. Intrinsically, therefore, the proposed algorithmic framework regards the problems of segmentation and understanding on medical images as a whole, in which the two problems affect each other and promote each other, effectively improving the accuracies of segmentation and classification for ultrasound images. The proposed project will be a beneficial, reasonable, and novel attempt for medical ultrasound image analysis.

医学图像的分割与识别在临床辅助诊断过程中发挥着日益重要的作用。然而，由于超声图像中含有大量斑点噪声，现有技术手段主要依赖图像自身的视觉特征，难以做到精确分割和准确识别。本项目创造性提出引入人体生理组织结构的解剖学语义信息，构造三维超声解剖语义模型，模拟医生对医学超声图像的理解过程，通过分割-识别-反馈学习的迭代过程，不断优化分割结果，实现最终的基于解剖语义的图像识别和理解。该思路融合传统的图像分割与识别技术，结合超声图像仿真、语义图像分类、多目标问题优化、并行计算等技术，通过学习反馈机制自我修正分割与识别结果。因此本质上，该项目提出的算法架构将医学图像的分割与理解有机的合并成同一个问题，借助解剖语义的先验知识，两者相互影响、相互促进，有效提高超声图像的分割与识别准确性，是一次有益的、合理且具有创新性的尝试。

医学图像的分割在临床辅助诊断过程中发挥着日益重要的作用。然而，由于超声图像中含有大量斑点噪声，现有技术手段主要依赖图像自身的视觉特征，难以做到精确分割。目前，精确的超声图像分割依然是学术界的难题。本项目创造性提出了一种新的基于语义分类和超像素图像块的乳腺肿瘤分割方法。采用对图像局部区域即超像素提取高层次特征（视觉词汇），达到对区域的理解，实现局部分类，从而实现整体的分割。首先，我们研究超像素的语义表示方法，提取图像特征后建立词袋模型用于语义表示；然后，研究对超像素进行分类的方法，利用BP神经网络（BPNN）进行初始分类，采用k近邻（KNN）方法进行重新分类。最后，我们把超像素分类后的边缘相连接，得到最终的分割结果。此外，我们与医院合作，获取了大量病例，并对其中的典型病例进行了标注。在获取的病例上，我们进行了大量的对比实验。与传统方法相比，我们的算法取得了更优秀的结果，得到了肿瘤轮廓的更优近似。本研究首次提出的分割方法，对提高超声图像分割的准确率有很大的帮助。同时，也是一种全新的分割思路，对后来的研究者而言也有很大的学术意义。