Faster R-CNN内窥镜息肉检测

Faster R-CNN based polyps detection in endoscopic videos

为了提高内窥镜下肠道息肉检测率,提出一种基于Faster R-CNN的息肉检测方法。在数据预处理阶段,利用中值滤波的非线性滤波特性去除图像反光区域,通过数据增强方法扩充样本数据集。在网络结构上,使用残差网络提取多尺度特征送入区域候选网络,得到息肉候选区域;再通过更快的区域神经网络进行训练直至网络收敛,经过微调得到最终检测网络模型。实验结果表明,该方法能够准确检测息肉并标记息肉位置,准确率为96.9%,召回率为95.8%。

In order to enhance the detection rate on polyps of intestinal under endoscopic,a polyp detection method based on Faster R-CNN is proposed.In this method,during the data processing stage,the nonlinear filtering characteristics of the median filter is used to remove the image reflection area,and the data enhancement method is used to expand the data set.In terms of network structure,the ResNet-50 is used to extract multilevel features,which is then sent to the region proposal network(RPN)to generate candidate regions to get the candidate region of polyp.Fast R-CNN network is then used to train until the network converges,and the final detection model is obtained by fine tuning.Experimental results show that this method can accurately detect polyps and mark the position of polyps.The accuracy rate is 96.9%and the recall rate is 95.8%,which proves to be an effective clinical detection method.