Medical Image Detection Based On Augmentation Module Neural Network

With the improvement of computer hardware performance and the significant increase of available data,convolutional neural network has achieved vigorous development in various fields.People have been promoting the neural network toward faster reasoning speed and better fitting ability.However,convolutional neural network is still weak in the face of small sample data sets,and few methods can well expand the feature domain of small sample data sets.At present,the data sample set available in the field of medical imaging is far less than that of coco,and different diseases need to build data sets separately,which leads to the lack of data samples of most diseases except some hot diseases,and convolutional neural network is difficult to play an ideal role.In the face of small sample data sets,data amplification is the most commonly used and the most reliable method to improve the robustness of the model.The existing data amplification methods can be roughly divided into single sample amplification and multi sample linkage amplification.The common methods of single sample amplification are rotation,clipping,adding noise,fuzzy processing,etc.,while the common methods of multi sample linkage amplification are mixup,samplepairing,cutmix,mosaic,fmix,etc.Although the existing amplification methods can make up for the lack of data to a great extent,there are few methods to make the network learn multiple images at the same time.Even using mosaic,which uses four pictures in the amplification process,the input network only has one amplified sample graph,which means that the neural network can only extract features from a single graph,which greatly limits the learning ability of the neural network itself.In this paper,an amplification module is proposed,which can transform any traditional convolutional neural network into an amplification module network.(1)Three images can be input at the same time,and the reasoning process of the middle image is called mainstream,and the two sides are called tributaries,so that the neural network can learn the information of the three images at the same time.(2)The main stream and the two branches each output one result.In this way,it takes only one inference time to obtain three results,which can further improve the generalization ability of the network by means of set module without additional time consumption.(3)The output characteristic graphs of the mainstream and the two tributaries are fused together as the final output characteristic graph of the mainstream,which ensures that the final characteristic graph scale of the mainstream network is consistent with that of the traditional convolutional neural network,consumes less parameters,and speeds up the reasoning speed of the model.The experiment is based on the data set of hyperosteogeny.There are 9005 pieces of data in the data,which are divided into four categories.There are three kinds of contrast experiments.(1)Comparative experiment between amplification module network and traditional network.(2)Comparative test of amplification module network under different input modes.(3)Set model experiment.The results show that the precision of convolution amplification is 13.5%less than that of traditional network module.After the set model experiment,the accuracy is further improved to 94%,the highest AUC value of the four categories is 0.9964,and the lowest is 0.9831.The experimental results show that the model reasoning speed and robustness of the amplification module network are better than those of the traditional convolution network.