基于多信息深度融合的肺部GGO结节自动检测方法研究

The early detection of GGO is important for improving the survival rate and cure rate of lung cancer patients. However, both the manual detection and the existing automatic detection methods have a high probability of missed detection and false positive rate. To solve that problem, we will propose an automatic detection method based on multi-information deep fusion in this study. First, according to the character of GGO, edge blur and small size, the multi-scale information will be fused for minimizing the information loss caused by the convolution and pooling in the deep neural network, to reduce the rate of missed detection. Secondly, multi-source information will be fused. The expert knowledge obtained from the medical field and the feature information obtained from the image field merge in a model level fusion. The deep learning with the complementary and redundant information is used to generate the optimal feature description, to reduce the false positive rate. Finally, in order to meet the requirements on the efficiency of clinical application, multi-perspective information is fused for selecting the most discriminant information by sampling from sagittal plane, coronal plane and axial plane, to improve the detection efficiency. The automatic detection method for GGO in this study enriches the theory of GGO detection and can be applied to the computer-aided diagnosis and treatment of lung diseases. Therefore, the automatic detection of pulmonary GGO nodule based on the multi-information deep fusion has scientific significance and clinical application values.

GGO的早期检测有助于提高肺癌患者的存活率与治愈率，但是目前人工检测和现有的自动检测算法存在着漏检和假阳性偏高的现象，为了解决这些问题，本项目拟提出基于多信息深度融合的自动检测方法。首先，针对GGO边缘模糊和尺寸较小的特性，利用多尺度信息融合的方式，来弥补深度神经网络框架中卷积和池化造成的信息损失，从而降低漏检率；其次，利用多源信息融合的方式，将从医学领域中获得的专家知识与从图像领域获得的特征信息进行模型级的融合，深度学习其中的互补与冗余信息，产生最优的特性描述，从而降低假阳性率；最后，为了达到临床应用的时效性，利用多视角信息融合的方式，通过对矢状面、冠状面和轴向面等多个视角采样，来选择最具有判别性的信息，从而提高检测效率。本项目提出的GGO自动检测方法，丰富和完善了GGO检测的理论，并可应用于肺部疾病的诊断和治疗中，具有一定的科学意义和临床实用价值。

本项目致力于GGO的自动检测，为提高患者的存活率，甚至提高肺癌患者的治愈率做贡献。.本项目基于CT影像与肺结节的特点，检测系统从候选结节检测与假阳性去除两方面来着手，提出了多个模块来提高检测准确率与效率。候选结节检测算法设计阶段，主要利用三维卷积操作来获取结节的局部信息，设计切片注意力模块来获取结节的可判别信息，设计多尺度模块来获取结节轮廓的精细描述，利用Transformer模块来获取全局信息，利用高斯分布来提取结节中心点的热力图来提高定位效率；在假阳性去除算法设计阶段，首先利用数据增强来对真结节进行样本数量上的填充，其次，采用基于直方图的数据筛选，自信度学习机制，Balanced-MixUp的采样方法来挑选具有代表性的假结节，从面平衡正负样本，最后，得到平衡数据后，设计了基于中心区域的神经网络模型，使其重点提取候选框的中心信息，提高结节的识别准确率。.本系统提出的检测算法具有97%的准确率，及0.6秒/张的检测速度。在公开数据库上，无论是准确率还是效率均优于现有的检测方法。在临床效果评估上，也达到了与高级放射科医生相当的水平。.因此，本系统在肺癌早筛的临床诊断中，不仅能够降低医生的工作量，还能作为辅助诊断，大大提高诊断的效率和准确率，从而进行较早的治疗，降低肺癌的致死率。在未来还可以扩展到其它部位，应用于微创手术或放疗环节中。