Research On Small Sample Biomedical Data Analysis Based On Deep Learning

With the development of information in the medical field,various types of biomedical data that can be obtained by experimental research also become rich.The body itself is as an important source of biomedical data,including blood pressure,body weight,data used drugs,doctors data,medical image data,data of blood,all kinds of hormone data,and some of the genetic data,etc.,and with the progress of medical technology can obtain the biometric data will be more.Although more and more data types are available,biomedical data are often characterized by small sample size,small feature size and highly unbalanced samples due to privacy,expensive collection cost and scarce special samples.Conventional deep learning models such as Res Net152 and VGG are based on massive training data sets.For small sample data such as biomedicine,forced use of these models for training often leads to over-fitting and failure to generalize or fall into local optimal problems.Aiming at this problem,this paper tries to use more knowledge methods of small sample learning field to solve the training problem of small sample biomedical data.In this paper,model construction,knowledge transfer and data expansion are studied and analyzed.DEM Model(An Deep Embedding Model for Few and Imbalanced Biomedical Data)is proposed in this paper.This model is based on metric learning and introduces Feature Selection and Feature Embedding.And use KNN(K-nearest neighbors),CSDNN(cost-sensitive Deep Neural Network),SMOTE(The Synthetic Minority Oversampling)Technique and FSSNN(Siamese Neural Network with Feature Selection)completed a sufficient comparative study.In terms of knowledge transfer,this paper makes an in-depth study on heterogeneous transfer learning and proposes A DDA Model(A Deep Domain Adaptive Model For Autism Detection),which introduces A Domain Adaptive layer and A Domain confusion loss.In order to verify the performance of the model proposed in this paper,previous research work was introduced into the comparative experiment in this part,mainly including RESNET-152,STN(Heterogeneous Domain Adaptation via Soft Transfer Network),DDC(Deep Domain Confusion: Maximizing for Domain Invariance)In terms of data expansion,the T-SIRGAN model is proposed in this paper,which extends the training data set through the SIR epidemic model and the GAN(Generative Adversarial Nets)Adversarial network model,and then the Transformer model is used to complete the epidemic trend prediction of the Novel Coronavirus.In this paper,through many experiments and a large number of comparative experiments,the model proposed in this paper has achieved good performance.DEM model shows obvious advantages in three different highly unbalanced data sets.DDA model is also superior to previous studies in autism detection tasks.T-SIRGAN,in the task of predicting the epidemic trend of novel Coronavirus disease,showed that the overall predicted trend was basically consistent with the real data.