Classification Algorithm For Chronic Gastritis Based On Deep Learning And Federated Learning

Gastric cancer is the third leading cause of cancer deaths in the world,and gastritis including intestinal metaplasia(IM)and gastritis atrophy(GA)are typical early symptoms of gastric cancer.Traditional methods relying on doctors’ visual inspection have high rates of missed diagnoses,so the study of using deep learning combined with gastric endoscopic images to achieve intelligent detection has significant practical significance.Medical images have high resolution,and mainstream deep learning models tend to compress the input images resulting in loss of lesion information.There are also problems with training deep learning models that require large amounts of annotated data,which can lead to privacy information leakage.Therefore,this thesis proposes novel deep learning and federated learning technology to overcome the aforementioned difficulties,and summarizes the main research content as follows:(1)This thesis proposes a multi-scale input integration network and a dual-transfer learning strategy to address the high information loss rate and inefficient utilization of images from similar meaning equipment caused by downsampling high-resolution images.In order to improve the information utilisation of high resolution medical images,the method uses a multi-scale input ensemble network to split the input into globally compressed and multiple local uncompressed crops,and adds attention modules at the end of each local module to improve the focus on target features.To improve the utilisation of similar device images,the dual transfer learning strategy is used to enhance the model’s recognition of the task features.The effectiveness of the proposed classification model for the high-resolution medical image classification problem is verified by comparing it with a benchmark classification model on the gastritis dataset.(2)To address the need for privacy protection of clinical data and the non-independent and identical distribution of data across multiple hospitals,this thesis proposes a federated learning approach based on an assisted aggregation strategy and a model correction strategy.To alleviate the model forgetting problem caused by the non-independent identical distribution of participant data,the method uses a model performance-based aggregation strategy to improve the model aggregation effect by recording the outstanding model cohort and fusing the outstanding model information in the aggregation process.To alleviate the drift problem during model training,a regular term based on model performance is introduced into the objective function of the client-side model to alleviate local model overfitting,and poor models are corrected by frequency-domain correction prior to server-side aggregation.The effectiveness of the proposed approach is verified by comparing it with a benchmark federated learning approach on the gastritis dataset.(3)To address the problem of high cost of accurate annotation of medical images and low utilisation of unlabelled data,this thesis proposes a federated semi-supervised learning method based on a greedy aggregation strategy and consistent regularisation of multiple auxiliary models.In order to enhance the efficiency of model aggregation using labelled data,the model with good performance is selected as the initial model for the next round of clients by comparing the performance of aggregated models and supervised learning models at the server side.To efficiently utilise unlabelled data to improve model performance,high quality pseudo-labels are generated on the client side by using the best model as the teacher model,aided by a consistent regularisation method for the local client.The effectiveness of the proposed method was verified by comparing it with a benchmark federated semi-supervised learning method on the gastritis dataset.