胸部CT图像中孤立性肺结节良恶性快速分类

目的：为突破医学影像诊断学依据医学征象进行定性诊断准确度不高的瓶颈，针对胸部CT图像中孤立性肺结节（SPN）定性诊断问题，寻求能够有效表示SPN病理特性的图像特征，实现快速准确地SPN良恶性计算机辅助诊断系统。方法：首先，采取交互式分割方法从胸部CT图像中提取出SPN；其次，直接计算SPN图像的多分辨率直方图得到768维空间信息特征样本集；然后，充分利用具有处理高维数据集优势的支持向量机（SVM）构造SPN良恶性分类器；最后，通过测试样本集对经训练后的SVM分类器进行测试以评价分类性能。结果：经214例病例实验结果表明：240个SPN图像的768维特征计算所用时间为4.83秒，SVM分类器训练测试所用时间为2.24秒，敏感性71.33%,特异性70%，准确度71.67%，接受者操作特性曲线(ROC)下面积(AUC) 0.7864。结论：该系统提取的高维图像空间信息特征能够有效表示SPN特性；没有考虑医学征象进行SPN定性诊断的准确度就达到了71.67%，同时分类速度比传统纹理算法提高了近50倍，为医学影像学解决SPN定性诊断问题提供了便捷、客观的辅助手段。

Fast classification of benign and malignant solitary pulmonary nodules in CT image

Objective: In order to break through the bottleneck of low accuracy diagnosis with medical signs in medical imaging diagnostics, effective image features of solitary pulmonary nodules (SPN) need to be found for the computer-aided diagnosis system quickly and accurately differentiating benign and malignant SPNs in chest CT images. Method: First, SPNs are extracted from chest CT images using interactive segmentation. Second, the multi-resolution histograms of SPNs are directly calculated to receive a high-dimensional features sample set with spatial information of SPNs.Then the classifier for differentiating benign and malignant SPN is constructed with making full use the advantage of SVM which is good at dealing with high dimensional data sets. Finally, the performance of classification is evaluated by testing the trained SVM with the test sample set. Result: The test results by 214 cases show that it takes 4.83s for computing 768 dimensional features of 240 SPNs, 2.24s for training and testing the SVM classifier. Receiver Operating Characteristic (ROC) analysis of classification performance of the proposed approach shows that the sensitivity is 73.33%,specificity is 70%,accuracy is 71.67%,and the Area Under Curve (AUC) is nearly 0.7864. Conclusion: Image spatial information can effectively express the characteristics of SPN, the system classification accuracy of benign and malignant SPNs is up to 71.67% without medical signs, and the classification velocity is about 50 times than traditional texture methods. It provides a feasible, simple, objective method for solving the problem in medical imaging diagnosis of the SPN.