Research On Medical Image Classification Algorithm Based On Federated Learning

With the development of artificial intelligence and Internet of things technology,a large amount of data that can be used for deep learning model training has been generated in image processing,speech recognition and other fields.Using these data,we can train a better model.However,in the medical field,how to train effective models without revealing patient privacy(or data)has become a big challenge in artificial intelligence technology.As a distributed machine learning framework,the Federated Learning algorithm effectively solves the above problems.Federated learning allows multiple institutions to train a model together in a distributed manner without sharing data,thus realizing data privacy protection.It allows the model to be trained in different locations and distributed environments,so Federated Learning method can use more data and computing resources to improve the accuracy of the model.However,in the medical field,because different hospitals have different sampling methods for patient data,different areas and cycles of patients,there is a Data Heterogeneity problem,which is one of the issues urgently addressed in the Federated Learning field.That is,how to reduce the impact of heterogeneous data on model performance and train a model with high precision and strong generalization ability under the condition of highly heterogeneous data.The specific research content is as follows:(1)A new medical image classification model based on Federated Contrastive Learning is proposed.In each round of Federated Contrastive Learning,the training objective of the model is to reduce the distance between the feature representation of the local model learning of the current round and the feature representation of the global model learning of the last round,and increase the distance between the feature representation of the local model learning of the current round and the feature representation of the global model learning of the last round,so that the local model gradually approaches the global model,learns better features and increases the generalization ability of the model.The experiment shows that,compared with other Federated Learning methods,this method has higher accuracy in the melanoma skin disease detection dataset.At the same time,this method has achieved 93.43% accuracy in the detection dataset of COVID-19,which has good application significance.(2)A new medical image classification model based on Enhanced Transfer Learning is proposed.In the process of local training,if the local model is updated by the global model directly,the local knowledge learned by the local model will be eliminated.Enhanced Transfer Learning can enable the global model and the local model to learn from each other and integrate their global knowledge and local knowledge,thus reducing the performance regression caused by Data Heterogeneity.In addition,the model also uses a small amount of sample data to train the global model on the server side,which further corrects the impact of class imbalance data on the global model.Extensive experiment shows that by integrating these two strategies,the classification accuracy of the model on two datasets has been effectively improved.In addition,compared with the Federated Contrastive Learning method,this method achieves better results in the metrics of accuracy and specificity,and its performance on the COVID-19 detection dataset is 94.73% and 95.44%,respectively.