Research On Tumor Recognition Based On Transfer Learning

Computer-aided diagnosis based on medical images is often regarded as the reference and "second opinion" of doctors,which can reduce the burden of doctors’ work and improve the accuracy of medical diagnosis.Labeling data of experts in the medical field is very expensive,and some cases are rare.It is very difficult and unrealistic to obtain large-scale medical data to analyze labeled sample data,and there are a few data challenges in the medical field.Machine learning in medical image analysis plays a vital role,has become the most promising field of study.Existing deep learning methods success depends on large-scale labeled data set,so using the method of transfer learning on tumor image diagnosis classification becomes extremely natural.Using medical analysis data samples as the target domain,and other irrelevant or related data samples with rich annotations as the source domain,you can use another sufficiently labeled sample in the source domain that is different from the target domain data distribution but semantically related to identify the target Domain task samples are becoming an increasingly important subject,which will undoubtedly solve the problem of low data in the medical field.Aiming at the problem of lack of large-scale data set in the task of tumor recognition,this paper constructs tumor image recognition models based on transfer learning,so as to improve the feature extraction ability of the model for tumor images and obtain better accuracy.The main work of this paper is as follows:Firstly,a tumor recognition model based on regular convolutional network and transfer learning is proposed.According to the characteristics of tumor recognition task,this model improves the VGG16 deep network,optimizes the network through L2 regularization,fine-tunes the last three layers of the pre-trained convolutional base,and successfully transfers the pre-trained network of source data set to the tumor image data set.Through a large number of comparative experiments,the feasibility and superiority of the proposed model in the tumor high-dimensional small sample data set are verified.Secondly,combined with feature-based transfer learning,this paper conducts the first study on the unsupervised domain adaptation classification of tumor pathological images,and propose the domain adaptation tumor recognition method based on the antinetwork.The algorithm trains a binary classifier on the source domain and the target domain,and makes it play with the self-built feature extractor,so as to learn the domain invariant features.Next,the source domain classifier is used for label prediction on this feature.Experiments show that the domain adaptation method based on the antagonistic network is effective and the accuracy of the malignant tumor image recognition is further improved.