Research On Esophageal Cancer Detection Based On Convolution Neural Network In Barium Esophagram

Esophageal cancer is the eighth most common cancer worldwide.Barium esophagram is an inexpensive,noninvasive,and widely available method for esophageal cancer detection.Deep learning system based esophageal cancer detection can provide doctors with effective information on esophageal cancer diagnosis,which is of great significance for medical clinical treatment.However,the variation of tumor size,the interference of surrounding tissues and esophageal deformation makes tumor detection very challenging.In this paper,an esophageal cancer detection system based on convolutional neural network is proposed for identifying positive patient cases and detecting esophageal cancer.The system mainly consists of a detection network,a classification network,and a majority voting module.The RoI(Region of Interest)detected by detection network from the images of different positions of patients are used as the input of classification network to judge whether the RoI contain tumor.On this basis,patients are judged whether they are a patient with esophageal cancer through a majority voting method.If so,the lesion is output through detection network.In order to improve the detection performance of esophageal cancer,this paper has made two improvements to Faster R-CNN.To solve the problem that the esophagus area of the feature map obtained by the backbone network of Faster R-CNN is not obvious and the background area is relatively large,this paper proposes CBAM_Faster R-CNN(CAFR).It adds CBAM(Convolutional Block Attention Module,CBAM)in the last convolutional module of backbone network to improve the feature saliency of the esophagus region in the feature map.In order to solve the problem that a single prediction feature layer has poor prediction effect on lesions of different sizes,and conventional convolutional layer cannot sample deformed esophagus area adaptively,the second improved Deformable_FPN Faster R-CNN(DFFN)is proposed in this paper.DFFN uses ResNet50 backbone network with stronger feature extraction capability.In the last Bottleneck Block of the ResNet50 backbone network,the deformed esophageal wall is sampled adaptively using Deformable Convolutional Network instead of convolutional layers.At the same time,the improved ResNet50 backbone network is used in combination with the FPN network to predict lesions of different sizes.In the feature of region proposals mapping,DDFN uses RoIAlign rather than RoIPooling to avoid the loss of some feature information in the feature mapping process.RoIAlign does not perform any rounding.Considering the small resolution of RoI,we design a classification network more suitable for RoI.This network consists mainly of three convolutional blocks,one attention module and two linear layers.These blocks are used to extract the feature of RoI,the attention module is used to enhance the feature significance of the esophagus in the feature map,and finally output the RoI classification results through two linear layers.The deep learning system outputs whether a patient has esophageal cancer through a majority voting module.The module is recurved from the RoI classification results output by the classification network to the barium esophagram classification results,the position classification results,and the classification results of this case.If so,the RoI obtained by the detection network are lesions.The experimental dataset are provided by Tongji Hospital,Tongji Medical College,Huazhong University of Science and Technology.The deep learning system was evaluated in 40 positive and 53 negative cases with 100%(40/40)and 84.90%(45/53)accuracy in identifying positive and negative cases respectively.In the 40 positive cases,the AP metrics of CAFR and FRDF is 58.94 and 63.30 respectively,which is 2.2 and 6.59 higher than that of Faster R-CNN of 56.71.