基于卷积神经网络的胸片肺结节检测

Lung nodules detection via convolutional neural networks in chest radiographs

针对目前胸片的肺结节检测方案的检出率较低,且存在大量的假阳性的问题,提出了一种新的基于卷积神经网络(CNN)的肺结节检测方案。增强肺结节区域的图像信号;选择正、负样本训练卷积神经网络模型,检测结节时用滑动窗口的方法对增强后的图片进行处理得到候选区域;根据候选区域的面积排除假阳性。方案中省略了传统方法中的肺区分割步骤,避免了因此可能丢失的肺结节图像。在日本放射技术学会(JSRT)数据库上测试结果显示,系统在平均每幅图5.0个假阳性水平下敏感度为86%,对不明显和非常不明显的结节检出率达到了84%,优于当前相关文献报道的方法。

Aiming at the problem that detection rate of lung nodules detection scheme based on rabat is low and has a lot of false positives,propose a new nodules detection scheme based on convolutional neural network（ CNN）.In the scheme,enhance chest radiograph,and then pick positive and negative samples to train the CNN. Process the enhanced image using sliding windows method with the pre-trained network to get the region of interest（ ROI）,and exclude the false positives by using the size of the ROI at last. The proposed scheme omits the procedure of segmentation of lung field in traditional schemes. And this can avoid losing nodules which are excluded by the segmentation procedure. The JSRT database is used to evaluate the system. The scheme achieves a sensitivity of86 % for all nodules and a detection rate of 84 % with 5. 0 FPs per radiograph for very subtle and extremely subtle nodules which outperform the current reported methods.