A Study Of Federated Learning Based Algorithms For Skin Disease Detection

With the continuous development of artificial intelligence algorithms such as deep learning,neural networks and federal learning,aided disease diagnosis systems have been deeply integrated with the field of artificial intelligence.For example,during the new crown epidemic,computer vision technology was widely used to aid diagnosis and the issue of assisted diagnosis has become a hot issue in the field.The number of patients with dermatological diseases has been increasing in recent years,while the number of dermatologists has not increased accordingly with the number of patients.Therefore,there is a need to develop a dermatological aid to diagnosis in the field of dermatological testing.Due to the special nature of dermatological diseases and their medical datasets,there is no way for ordinary institutions and individuals to collect them directly,leading to the current problems of patient privacy leakage,small number of datasets,uneven distribution of datasets,and low diagnostic accuracy when building a dermatological auxiliary diagnosis system.In response to the shortcomings of existing research,this thesis has conducted research in the following areas:(1)A two-weight asynchronous aggregation federation learning algorithm is proposed to improve the federation learning algorithm and to increase the training efficiency and communication efficiency.The core idea of the algorithm is to introduce two weight parameters for aggregating the global model and updating the local model.In this algorithm,individual clients can upload models for aggregation after training is completed,without waiting for other clients,thus greatly improving the training efficiency and communication efficiency of the models.(2)The image segmentation algorithm is optimised based on the idea of Fully Convolutional Network(FCN).In order to solve the problem of small dermatological data sets and uneven distribution of the number of samples for each type of dermatological examination,a polynomial loss function(PolyLoss)is introduced as the loss function.In order to obtain deeper image features and avoid gradient disappearance and gradient explosion in the network,a residual module is introduced to build a fully residual convolutional optimization network,which improves the accuracy of image segmentation.(3)Introduction of traditional residual network structure and attention mechanism to improve image classification algorithm.In order to enable the network to focus more on the lesioned tissue regions of the skin lesions,the Attention Residual Learning(ARL)structure combined with the attention mechanism was introduced and the image classification network was built based on the ARL module.In addition,the dataset used for training the dermatological classification network suffers from the aforementioned problems of low data volume and uneven distribution of data samples.To solve this problem,the Cross-entropy loss function(CE)was optimised to calculate the loss for different classes of data by weighting it according to the percentage of the overall samples that their classes represent.The accuracy of the image classification network is further improved by using the skin-damaged images after image segmentation and scaling in the previous step when performing the classification task.(4)A two-step dermatological disease detection algorithm consisting of a two-weighted asynchronous federation learning algorithm with image segmentation and image classification networks is fused to propose a federation learning-based skin disease detection algorithm and validated on a real ISIC 2018 dataset to demonstrate the effectiveness of the algorithm.