基于影像组学的乳腺癌病灶分割及预后分析

发布时间：2018-04-28 23:03

本文选题：直方图均衡化 + 迭代增强　；参考：《哈尔滨理工大学》2017年硕士论文

【摘要】：影像组学研究对于定量肿瘤的异质性、辅助临床决策有很大的潜力和应用价值。在乳腺癌的临床诊疗中,淋巴结是否转移对于手术方案的制定有重要指导意义。但是目前临床上缺乏预测淋巴结转移的安全有效手段。本文应用影像组学方法对乳房CT影像进行定量分析来实现乳腺癌癌淋巴结是否转移的预测。我们从广东省人民医院收集了456例满足分析要求的乳腺癌癌病例数据,首先提出了一种应用于对比增强CT乳腺癌的半自动随机游走算法进行分割。该算法先使用手动辅助并进行最大区域提取得到局部乳腺病变。然后,使用改进的迭代直方图均衡化来对乳腺病灶进行增强,使病灶与周围乳腺组织有较大对比度。最后,使用基于四个种子点的随机游走算法来分割病变区域。对分割的病变组织,从每例中提取了592个影像特征,包括肿瘤的图像强度特征、形状和大小特征、纹理特征以及小波特征;然后利用LASSO逻辑回归(LLR)来构造乳腺癌淋巴结转移的影像组学标签(Rad-score),并结合临床信息进行多元分析,构造了诺模图个性化预测模型。基于以上的方案,我们得出以下结论:LLR基本预测模型在训练集上AUC为0.712,测试集AUC为0.737,且对比不同的模型,F1值表现最高,为0.705;在诺模图上,用所有数据进行个性化预测,得到C-index为0.721(95%CI:0.676-0.768),在一致性曲线上表现较佳,可为临床决策提供有价值的信息。
[Abstract]:Imaging studies have great potential and application value for quantitative tumor heterogeneity and clinical decision-making. In the clinical diagnosis and treatment of breast cancer, lymph node metastasis is of great significance for the formulation of surgical plan. But there is no safe and effective method to predict lymph node metastasis. In this paper, quantitative analysis of breast CT images is used to predict lymph node metastasis of breast cancer. We collected 456 cases of breast cancer from Guangdong Provincial people's Hospital. A semi-automatic random walk algorithm for contrast enhanced CT was proposed to segment breast cancer. In this algorithm, the local breast lesions are obtained by manual assistance and maximum region extraction. Then, the improved iterative histogram equalization is used to enhance the breast lesions, which makes the lesions more contrast with the surrounding breast tissues. Finally, a random walk algorithm based on four seed points is used to segment the lesion region. For segmented lesions, 592 image features were extracted from each case, including image intensity, shape and size, texture and wavelet features. Then the LASSO logical regression was used to construct the radiographic tag Rad-scoreen for lymph node metastasis of breast cancer and the multivariate analysis of clinical information was carried out to construct the personalized prediction model of Northromogram. Based on the above scheme, we draw the following conclusions: the AUC of the basic prediction model is 0.712 on the training set, the AUC of the test set is 0.7377.The comparison of different models shows that the F1 value is the highest (0.705). The C-index is 0.721 ~ 95CI: 0.676-0.768, which is better on the consistency curve and can provide valuable information for clinical decision making.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R737.9;R730.44

【相似文献】