Research On Siamese Cross Contrast Neural Network Adapting To Small Medical Dataset

Due to the particularity of medical image data,deep learning algorithms which perform well in traditional computer vision tasks are difficult to be directly and effectively applied in the field of medical image analysis.The generation difference between natural images and medical images in model performance may be attributed to the insufficient scale of medical image data and the mismatch between the underlying design physical ideas of the existing deep learning algorithm and the dynamic physiological system corresponding to the medical image data.With the increasing demand for medical image analysis today,it is of great significance to explore new methods for adapting small medical datasets in the context of deep learning.At present,relevant research mainly focuses on the three directions of expanding the scale of data,using transfer learning and designing the effective representations which fit the dynamic physiological system of medical image data better.We are inspired by the cross contrast neural network and propose a new structure named siamese cross contrast neural network which integrates three strategies into an end-to-end network enhancing the adaptability of small medical datasets.This is a new attempt to study the problem.From the structural perspective,the network is mainly composed of siamese feature extractors and auxiliary classifiers.Under the framework of metric learning,the newly designed distance metric based on the information-based similarity theory guides twin feature extractors to mine features which are discriminative and better fit the characteristics of the potential dynamic physiological system of the medical image data.The auxiliary classifiers as the downstream network of the siamese structure use the extracted features and directly output the prediction results.In this study,in order to evaluate independently the impact of the new designed distance metric on extracting features,the gradient of the auxiliary classifiers was truncated.In practical applications,you can select to liberate the gradient according to the specific problem.Siamese cross contrast neural network has the characteristics below:From the perspective of expanding data,with the help of the framework of metric learning,examples are randomly matched into pairs,which indirectly enlarges the data available for training.In addition,the comparisons between examples also contain rich information.Siamese structure simplifies the inputting and pairing process and realizes random input to a certain extent.The auxiliary classifiers simplify the prediction process of metric learning and directly output the classification results.From the perspective of designing representation,which fit the dynamic physiological system of medical image data better,the distance based on the information-based similarity theory pays more attention to the order of feature probability,reduces the need for the significance of dominant features.It is more suitable for the situation in the actual physiological system where the resolution between lesion and background tissue is constantly changing,and the lesion area is too small,or even insignificantly visible.This newly designed distance is in line with the underlying physical thinking of the medical system,and it is easier to maintain stability between the physiological systems.From the perspective of transfer learning,the basic model of the siamese feature extractors can use any effective deep convolutional network,and the initialization parameters are obtained through pre-training.In practical applications,you can choose to train all or freeze some of the parameters.This method is validated in a classification task of hepatocellular carcinoma and intrahepatic cholangiocarcinoma with abdominal CT images.This study provides detailed results together with visual analysis.The optimal accuracy results in the binary and three-categories tasks are 94.2% and 90.2% respectively.