Research On Key Technologies Of Classification For Medical Imbalanced Sample Sets

Medical image classification is one of the most important tasks in medical image analysis.In practice,it is often faced with the problem of imbalanced sample set,that is,the number of samples in different classes is greatly different,and manifested as the number of positive samples is often less than the number of negative samples.However,the problem of sample imbalance is natural in the medical field,so it is an unavoidable problem in medical image analysis,and also one of the main challenges in the task of medical image classification and recognition.This thesis mainly focuses on the class imbalance in medical image classification.The main work is as follows:1.A method of using Capsule Network for medical imbalanced sample sets classification is studied.The capsule network can overcome the shortcoming of the convolutional neural network’s dependence on a large amount of training data,and show better robustness on imbalanced datasets.According to the two characteristics of data imbalance and limited data in medical image datasets,capsule networks have the advantage of being more suitable for medical image analysis tasks.The experimental results verify the feasibility and effectiveness of capsule networks in medical image classification tasks.2.The loss function of the original capsule network is improved.In order to effectively solve the problem of class imbalance in medical image classification,and avoid the risk of overfitting or important data loss caused by data sampling,this paper introduces the class-balance loss from the algorithm level to improve the loss function of the capsule network.The improved loss function reweights each class according to the number of effective samples of each class to balance the contribution of each class to the overall loss.The experimental results show that the improved loss function can effectively guide the model to learn on the imbalanced data sets,and the classification accuracy is obviously improved.3.A model of class-balanced attention capsule network is proposed An attention module was added between the first layer of the original convolution layer and the second layer of the primary capsule layer,which solved the problem that the original capsule network ignored local feature when it emphasized the position or rotation spatial relationship between features,enriched the model’s hierarchy to promote the model to learn better features.The experimental results show that the proposed network model can effectively deal with the problem of class imbalance,improve the classification accuracy of imbalanced data,and provide new research ideas and methods for the classification of imbalanced data.4.An application system based on class-balanced attention capsule network classification algorithm is implemented.The system uses the medical assistant platform as the carrier,and mainly implements the functions of auxiliary diagnosis and case management,and has positive significance and important role in assisting doctors in improving diagnosis accuracy and reducing missed diagnosis.