基于多模态图像和多分类器融合的DBT肿块自动检测方法研究

Breast cancer is the most prevalent major disease threatening human health, and the leading cause of death worldwide. Scientific evidence has shown that early detection and treatment is the most effective way of breast cancer care. It can significantly reduce patients' mortality rates. Therefore, it is critical to detect and diagnose breast lesions at their early stage. Although digital breast tomosynthesis (DBT) can effectively improve the detection ability of early breast lesions, however, due to the limited scanning angle, X-ray scattering and slightly longer exposure time, the DBT volume has problems of blurring in the out-of-focus plane, low mass conspicuity and ill-conditioned background, which cause the accuracy of lesions detection was not high enough with DBT. So, it is very important to develop the research about the key science problem of DBT masses detection and its achievements have extensive foreground for the clinical application for better diagnosis and treatment of cancer. With the purpose of improving the rate of true positive effectively, on the base of not improving the rate of false positive, this research project takes the DBT reconstructed volume and synthetic image as the research object, researches and analyzes some problems of restricting the detection performance of DBT masses using the key technique of multi-classifiers fusion. The main research contents are as follows: first, devise the method for configulating the synthetic image with conspicuity masses. Second, establish the segmentation model for suspicious masses that in the synthetic image and reconstructed volume upon a contour–based detection method. Third, finish the research of multi-classifiers fusion algorithms for DBT masses detection.

乳腺癌是当前严重影响女性健康、威胁女性生命的主要恶性肿瘤之一，目前最有效的治疗方案是在乳腺癌形成的早期及时发现并实施针对性的治疗。因此，早期发现病变并诊断成为关键。数字乳腺断层合成摄影（DBT）虽能有效提高早期乳腺病变检测能力，但因扫描角有限、X线散射及曝光时间稍偏长，导致焦外平面模糊、肿块显著度低和病态背景，使通过DBT检测病变的精度不够高。因此，开展检测DBT肿块的关键科学问题研究具有重要理论和现实意义，其研究成果对于更好地诊断和治疗癌症具有广泛的临床应用前景。本项目以DBT重建体和合成图像为研究对象，以多分类器融合为关键技术，对制约DBT肿块检测性能的若干问题进行深入研析，以期在不提高假阳性率的情况下，有效提高真阳性率。本项目主要研究内容包括：（1）设计肿块显著的合成图像构造方法；（2）构建基于轮廓检测法的合成图像和重建体疑似肿块分割模型；（3）完成DBT肿块检测的多分类器融合算法。

乳腺癌是女性最常见的恶性肿瘤，目前最有效的治疗方案是早期诊断并精准治疗。DBT能有效提高早期乳腺病变检测能力，但存在扫描角有限、X线散射及曝光时间稍偏长，导致焦外平面模糊、肿块显著度低和病态背景等问题，导致DBT检测病变的精度不够高。因此，开展检测DBT肿块的关键科学问题研究具有重要理论和现实意义。为了对各种分割算法进行验证、测试和评估，本研究通过多家附属医院影像科采集病例数据，建立了两个DBT图像数据库。考虑到DBT图像常含有泊松分布噪声，导致图像边缘模糊，本研究设计了预处理算法进行滤波，提高DBT图像的质量。在此基础上，构建了基于U-Net结构的DBT图像乳腺肿块区域分割模型，能自动分割出DBT图像乳腺肿块区域。同时，为了进一步提高DBT图像乳腺肿块区域的分割精度，本研究建立了基于膨胀DCNN结构的乳腺肿块分割模型，实现了DBT图像乳腺肿块区域的准确分割。综合上述结果，本研究对于更好地诊断、治疗和处理肿瘤疾病具有重要的科学意义和临床应用价值。资助项目已发表SCI收录期刊论文6篇，EI收录期刊论文4篇，投稿在审SCI收录期刊论文3篇；申请发明专利1项。培养研究生3人，在读研究生3人。项目投入经费20万元，支出12.4018万元，各项支出与预算相符。剩余经费7.5982万元，计划用于本项目的后续支持。