Research On Medical Image Classification Technology Based On Deep Learning

Medical imaging is an important research category in computer vision,and the classification task is a classic machine learning task.In recent years,in the classification task of computer vision,research methods have emerged one after another,and these methods have not only created huge economic value,but also produced positive social benefits.Based on the specific analysis of the advantages and disadvantages of the existing medical imaging models,this thesis proposes the Softmix model and the Mix T model from the perspectives of data augmentation and neural network models,and validates them through a variety of experiments.In terms of data augmentation,the innovation of this thesis is based on the fusion of multiple images,which optimizes the drawbacks caused by boundary distortion,and proposes a data augmentation model Softmix.First,two boxes of different sizes are set to divide the image area.Next,the calculation formulas of the mask values of different divided regions are designed.Finally,the different images are fused using the mask value.At the same time,combined with related work,this thesis comprehensively analyzes the advantages and disadvantages of our model design,and conducts two experiments of existing model comparison and hyperparameter setting.The results show that the data augmentation model in this thesis has a certain accuracy improvement.This thesis also explores ways to improve the accuracy of medical image classification in terms of neural network models.In this regard,this thesis compares and discusses the advantages and disadvantages of the convolutional neural network and Transformer models,and proposes the Mix T model.The main innovation of the model is to use the convolutional position embedding layer to improve the model’s ability to acquire location information;use the pooling-based local information aggregation branch to improve the model’s ability to acquire local detail information;use the multi-head self-attention branch to Improve the model’s ability to obtain global information.In the overall design,this part sets up a hierarchical structure in the form of a feature pyramid.At the same time,comparative experiments and ablation experiments are also carried out,and the results show that the Mix T model generally improves the classification accuracy of medical images.In summary,this thesis proposes two model structures at the data and model levels,thereby improving the classification accuracy of medical images.How data augmentation can enrich the number and content of samples while minimizing the redundant and negative information introduced by humans provides a reference for how to give full play to the role of receptive field in improving accuracy in model design.