A modified faster R-CNN method to improve the performance of the pulmonary nodule detection

摘要

In order to realize the accurate and quick positioning of pulmonary nodules in hundreds of two-dimensional CT chest images and reduce the burden of radiologist, the paper proposes a modified faster R-CNN method to improve the performance of the pulmonary nodule detection. Firstly, data enhancement technology is adopted to expand the dataset. Secondly, the image is fed into VGG-16 with de-convolution to extract the shared convolution features. Then, the shared convolution feature is sent to the region proposal network (RPN) to output candidate lung nodule region. Finally, the candidate lung nodule region and the previous shared convolution features are input into ROI pooling layer at the same time, and the characteristics of the corresponding candidate area are extracted. Through the connection layer, a multi task classifier is used to position the regression of the candidate region. According to the features of complex chest image background, large detecting object range and relatively small size of pulmonary nodule compared with natural objects, we design a smaller anchor box to accommodate changes in lung nodule size. In order to get the more accurate description of the characteristics of pulmonary nodules, we add a de-convolution layer with 4, 4, 2 and 512 for nuclear size, step size, filling size and number of nuclei respectively after the last layer of VGG-16 network conv5_3, resulting in a higher de-convolution feature resolution. Finer granularity can be restored compared with the original feature map. The experimental results show that the average detection accuracy is up by 6.9 percentage points compared with the original model. This model can well detect solitary pulmonary nodules and pulmonary nodules and small nodules, showing certain clinical significance for early screening of lung cancer.