Research On Dermoscope Image Classification Based On Personalized Federated Learning

Skin cancer is often overlooked by patients due to its less prominent early symptoms,resulting in a high misdiagnosis rate.It ranks among the top types of cancer in both China and Western countries.Early diagnosis and surgical treatment can significantly increase the patient’s survival rate,making skin cancer one of the cancers that require special attention.To better diagnose skin cancer,medical professionals use dermoscopy to collect image data of suspicious areas and visually discriminate whether there is any cancerous lesion.In recent years,with the increasingly advanced development of artificial intelligence technology,collecting a large number of dermoscopy images and automatically classifying skin lesions in these images can greatly improve diagnosis speed and reduce patient mortality.However,in actual medical institution application scenarios,the category labels of dermoscopy images are based on the gold standard of pathological biopsy,which makes large-scale collection difficult.Moreover,due to patient privacy issues,data sharing is prohibited by relevant laws and regulations.In addition,differences in ethnic distribution,environmental factors,and equipment among medical institutions in different countries and regions lead to large differences in data distribution,which further amplifies the difficulty of joint training between institutions.In order to address the series of difficulties encountered in the process of implementing automatic classification of dermoscopy images,this paper conducts a series of studies as follows:(1)To address the problem of poor convergence accuracy of the model caused by the highly Non-I.I.D data from multiple medical institutions and the long-tail distribution of global labels,this article proposes a personalized federated dermoscopy image classification method,called p-MFC,which consists of two key modules,MD and PPA.Firstly,the PPA module is responsible for personalized aggregation of models uploaded by various medical institutions,and controls the degree of personalization to balance between global generalization and local personalization in optimizing training targets.Then,the MD module extracts global knowledge from models of various medical institutions with the assistance of Gaussian noise,calibrates it into the personalized model prepared by the PPA module,and distributes the personalized model to the corresponding medical institution.Through the coarse aggregation of the PPA module and the fine calibration of the MD module,p-MFC provides a personalized solution for federated dermoscopy image classification without additional auxiliary data.(2)Since the p-MFC method relies on noise-based distillation calibration,the performance of its MD module cannot be further improved due to data quality limitations.To address this issue,this article proposes p-AMFC,personalized Adversarial Mutual Distillation for Federated Dermoscope Image Classification,based on adversarial mutual distillation on top of p-MFC.In p-AMFC,a conditional generator is introduced on the server side to enhance the performance of the MD module.Furthermore,a method is presented to train the generator solely based on dermoscopy image classification networks uploaded by multiple institutions without extra data or discriminator assistance.Additionally,an AD module is designed in p-AMFC to provide adversarial feedback for training.Unlike traditional adversarial networks,the proposed AD module does not compromise the classifier’s performance,making it more suitable for training federated dermoscopy image classification models without assistance data.In the highly heterogeneous dermoscopy image classification dataset ISIC2019,p-AMFC consistently outperforms the latest federated learning methods in multiple experiments.This validates the effectiveness of the proposed method in federated dermoscopy image classification tasks.