基于医学影像结构和功能混合特征的周围型肺癌计算机辅助诊断方法

Early diagnosis of peripheral lung cancer can effectively reduce the mortality rate of lung cancer, which has the highest morbidity and mortality rate among all kinds of cancer. While the application of mixed features of functional image (single photon emission computed tomography) and structural image (lesion structure contrast imaging) makes it possible to use medical image features to make early definite diagnosis of peripheral lung cancer, this method cannot reduce the misdiagnosis rate effectively, probably due to such issues as its stability and the availability of doctors who have experience with this method. To improve the effectiveness of this diagnosis method, the proposed project aims to build self-feedback computer-aided diagnostic model by combining image visual features from original image, diagnostic imaging features from our image recognition algorithms, and medical diagnostic semantic from current medical practice. Our research will first extract and quantify the hybrid imaging features of functional and structural images of peripheral lung cancer. Then, a 3D mixed-feature model of lesions and image feature library will be constructed by using lesions slice images segmented by functional image. Furthermore, with machine learning techniques, we will integrate image visual features, diagnostic imaging features and medical diagnostics semantic to establish our self-feedback computer-aided diagnostic model. We expect that our model can effectively improve the efficiency and accuracy of peripheral lung cancer diagnosis and provide technical support for automatic diagnosis of early lung cancer. The main contribution of this project is the exploration of the theory and application of using computational techniques from image recognition and artificial intelligence in accurate clinical diagnosis.

目前，癌症中发病率和死亡率最高的是肺癌，周围型肺癌如果能实现早期确诊，可以有效降低肺癌的死亡率。尽管肺部功能影像（单光子发射断层成像）和结构影像（病灶结构造影成像）的混合特征在周围型肺癌诊断中的应用，为基于医学影像特征的早期确诊提供了可能，但是医师的人数、经验和诊断稳定性等瓶颈的存在，依然不能降低周围型肺癌的误诊和漏诊率。本课题通过提取和量化周围型肺癌功能影像和结构影像的混合成像视觉特征，利用功能像分割出的病灶断层结构图像，重建病灶三维混合特征模型和图像特征库；利用机器学习方法将图像视觉特征、影像学诊断特征和医学诊断语义进行有效融合，建立周围型肺癌的诊断关联模型；并进一步建立自反馈的计算机辅助诊断模型，有效提高周围型肺癌的诊断效率和准确率，为肺癌的早期自动诊断提供技术支持；为图像识别和人工智能等计算机技术在临床精确诊断中的应用进行理论和实践探索，具有一定的理论意义和学术价值.

肺结节的早期诊断有助于提高肺癌的治愈率。本项目通过对PET-CT混合成像的肺结节的研究，具体包括肺实质分割、肺结节的分割、二维和三维特征提取及三维重建、传统方法和深度学习对肺结节自动诊断模型的研究，哈希技术在肺结节图像检索中的研究，最终搭建出肺结节自动诊断模型。在序列肺实质分割方法研究中，项目组提出基于超像素和自生成神经森林的序列肺实质图像分割方法，利用六边形聚簇和形态学优化的超像素序列图像分割算法对图像进行过分割，再通过在超像素块间的距离计算中构造自适应权重系数，实现肺结节的精准定位。在对肺结节的分割方法的研究中，为解决以往分割算法对血管粘连型结节分割不准确以及分割效率较低等问题，提出基于超像素和密度聚类的序列肺结节图像分割方法；与此同时提出了基于超像素和稀疏子空间聚类的序列肺结节图像分割方法；在序列CT图像上，提出了基于多模态数据和超体素的肺结节图像3D分割方法，通过种子点的自动选取和拟定的增长规则进行自动超体素三维区域增长，最终分割得到三维肺结节。项目组还对肺结节的三维特征的提取以及深度特征的提取进行研究，以达到提高肺结节诊断性能的目的。通过图像哈希技术，引入概率超图方法和卷积神经网络模型，提出了将哈希码的学习问题转变为容易解决的概率超图的划分问题，同时用卷积神经网络技术对图像特征进行充分的提取。最后，搭建了五个肺结节分类和诊断模型：深度信念网络的肺结节良恶性分类模型、栈式极限学习机肺结节分类及诊断模型、双模态深度极限学习机自编码的肺结节诊断模型、基于卷积和反卷积神经网络肺部微小结节检测模型、基于局部感受野和半监督深度自编码的肺结节检测模型。本项目对肺癌早期诊断具有一定的理论意义和学术价值。