基于双边双视角的乳腺癌检测与分类方法研究

China has become one of the countries where the breast cancer morbidity is with the fastest speed. Moreover, breast cancer is getting more frequent among young people in China. Early detection is the key to increase recovery rate for breast cancer. Using the theory of image processing and pattern recognition, objective and quantitative results can be obtained, which contributes to detecting and classifying the breast cancer at an early stage. The existing detection and classification methods bared some problems. Lesion detection was more on single or two mammograms, and less on bilateral and two-view information. This was not fit with radiologist reading. Lesion feature extraction and classification was more on single view, and less on the fusion of two-view information. Moreover, the difference between the lesion center and lesion periphery was ignored in feature extraction..Aiming at these problems, the following contents will be studied. Simulating the radiologist reading, mass detection based on bilateral and two-view information will be studied. Lesion feature representing the difference between the lesion center and lesion periphery will also be studied. Fusing mass features from two views, classifying mass as benign or malignant based on two-view information can be implemented. The information discipline and medical discipline will be combined in this project research. This study has important significance in theory and will also promote the early detection and classification of breast cancer.

我国已成为乳腺癌发病率增长速度最快的国家之一，且呈年轻化趋势。对乳腺癌的早期检测与诊断是提高患者治愈率的关键。基于图像处理和模式识别理论，可获取乳腺X线图像定量客观的分析结果，有助于乳腺癌早期检测与分类。现有检测与分类方法存在的主要问题有：(1) 大多基于单幅或两幅图像进行病变检测，缺少双边双视角图像的病变检测方法，其与医生阅片机制存在一定差别。(2) 大多基于单个视角提取病变特征和实现病变良恶性分类，缺少双视角病变特征融合，同时在特征提取时忽略了病变中心与外围区域的差异性。.针对上述问题，本课题的主要研究内容为：(1) 模拟医生阅片机制，研究基于双边双视角图像的肿块检测方法。(2) 研究肿块中心与外围区域差异的特征提取以及双视角肿块特征融合，实现基于双视角图像的肿块良恶性分类方法。本课题是信息学科与医学学科的交叉学科，具有重要的理论意义，同时对乳腺癌早期检测与诊断将起到积极的推动作用。

乳腺癌是女性最常见的一种恶性肿瘤，其发病率正在不断上升，我国己成为乳腺癌发病率增长最快的国家之一。乳腺癌的早期诊断是提高患者治疗率的关键。本课题完善了乳腺X线图像数据库，并在该数据库上完成了：.(1) 研究现状分析：对乳头检测方法、胸肌分割检测方法、双边视角图像匹配方法以及同侧双视角图像匹配方法、单视角病变检测方法、双边分析病变检测方法、双视角分析病变检测方法、病变检索以及病变分类的研究现状进行分析，总结现有方法的问题，并给出了可能的解决方法。.(2) 多视角融合的肿块检测：提出了三种卷积神经网络模型，并将其用于医学图像肿块检测；建立了一种基于乳房生理特征的多视图坐标系；提出了一种双侧比对的肿块检测方法.(3) 肿块分类：提出了一种基于改进控制标记分水岭算法的肿块边缘分割方法，提出了一种对区域边界敏感的交叉熵阈值分割方法。提出了一种基于同轴模板与D-texton的肿块良恶性分类方法。提出了一种结合形状特征与纹理特征的肿块良恶性分类方法。.本课题属于信息学科与医学学科的交叉学科，利用多学科理论研究乳腺X线图像乳腺癌检测的关键问题，实现了乳腺癌检测的定量分析，有助于提高我国乳腺癌检测和诊断水平。