基于多视角深度信念网络的肺结节识别方法

针对传统肺癌计算机辅助诊断系统中肺结节检出过程烦琐,且存在假阳性高的问题,提出一种基于多视角深度信念网络的肺结节识别方法。该方法首先将肺结节进行三维重建并将重建后不同大小的肺结节归一到不同尺度的立方体中,然后将不同视角的2.5D切片作为深度信念网络的输入数据,最后通过不同的融合策略完成对肺结节的识别。在肺部图像数据库联盟(LIDC)数据集上大量实验表明:相比于传统肺癌识别系统本文方法敏感性为(92.8±0.25)%,平均每组病例假阳性个数为2.4±0.3,该方法能有效降低肺结节自动检测过程中的假阳性率。

Pulmonary Nodule Detection Method Based on Multi-view Deep Belief Network

Aiming at the complex process of pulmonary nodule detection in traditional computer aided diagnosis system of lung cancer and the existence of false positives, in this paper proposed a method for pulmonary nodules detection based on multi-view deep belief network. Firstly, the three-dimensional reconstructed pulmonary nodules are resized to different sizes of cube, and 2.5D slices of different angles are used as the input data of the deep belief network. Finally, the fusion strategy is used to identify the pulmonary nodules. Compared with traditional CAD system, the sensitivity of this method is 92.8±0.25% and the 2.4±0.3 false positives per scan. The large number of experiments on LIDC dataset shows that this method can effectively reduce the false positive rate of pulmonary nodule detection.