Research On Pneumonia Detection Based On Federated Learning

With the increase of air pollution and the spread of unknown bacteria in recent years,pneumonia,a common disease of the respiratory system,has become the second leading cause of death in children.In medical diagnosis,it has become exceptionally difficult to aid in the diagnosis of pneumonia using radiographs.A radiologist with expertise in the detailed observation of chest X-rays is required for this task.However,this may delay the diagnosis of pneumonia and thus affect the patient’s subsequent treatment.Excellent advantages in target detection have been shown by deep learning with the development of artificial intelligence.However,a large amount of high-quality data is required for training with deep learning.For medical images such as pneumonia,most of the data contains patient privacy and is highly specialized in labeling,so there are "data silos" in real life for such data.At the same time,most of these medical image data are stored in different medical institutions in different locations,and the data distribution has great variability due to different equipment and different parameters,which brings great challenges to the construction of deep learning models in medical treatment.Therefore,a method for detecting pneumonia based on federal studies is proposed in this thesis in response to the above questions.An efficient federal pneumonia screening model can be trained without localizing the data to train multiple medical institutions using this approach.The details of the study are as follows:Federated learning is introduced into pneumonia detection in this thesis to address the problems of "data islands" in chest medical images of pneumonia and data leakage in deep learning.The privacy of federated learning is used to prevent data leakage,and a federated pneumonia detection model FL-PD(Pneumonia Detection Model Based on Federated Learning)is built by using federated learning to solve the problem of "data islands" in pneumonia images with joint training of multiple clients.The attention of the model to deep features from both channel and space aspects is improved by introducing the CBAM module into the client-side local model.At the same time,a sparse gradient compression method optimized by Adam is introduced in the client-server transmission process to reduce the communication overhead of the model while ensuring the model effect.The accuracy of model detection can reach 94.2%,and the communication overhead is excellent under the condition that the data is not out of locality,according to experimental results.In addition,the original Fed Avg algorithm is improved based on the FL-PD model,and a federated learning algorithm called Fed Pso(Federated Learning Algorithm Based on Particle Swarm Optimization)that integrates particle swarm optimization is proposed in this thesis.It is used to solve problems such as reduced model accuracy and increased communication when the client’s data presents a non-independent and identically distributed state.The locally optimal particles from the client are uploaded to the server by using the search ability of particles,reducing the impact of data heterogeneity on edge clients.The experimental results show that the proposed method has good robustness in processing non-independent,identically distributed data,and the accuracy of model detection reaches 94.6%.It also has more efficient performance in independent,identically distributed scenarios.