| In the process of medical image analysis,deep neural networks are used to assist diagnosis and reduce the probability of misdiagnosis.However,in reality,due to policy regulations and patient privacy issues,multiple medical institutions are unable to share data for training neural networks.Federated learning can be used to solve this problem by training neural networks while protecting privacy.Predefined neural networks are often used in federated learning,and when the dataset changes,customized networks often need to be redesigned.Neural Architecture Search(NAS)can be combined with federated learning to automatically search for networks under federated learning settings.In addition,the deployment scenarios for medical image segmentation are diverse,and researchers hope to obtain multiple neural networks with different parameters,inference speed,segmentation performance,and other objectives through a single search.Therefore,NAS can be regarded as a multi-objective optimization problem.There is a common problem of expensive subnet evaluation in NAS,and the commonly used methods for alleviating the problem of expensive subnet evaluation in centralized NAS cannot be directly applied to federated NAS at present.Moreover,federated learning faces the challenge of high communication cost,which is even more severe under the influence of expensive subnet evaluation issues.In response to the above issues,this thesis first designs a medical image segmentation network Federated Searchable U-Net++(FSU-Net++)suitable for federated NAS,which has lower parameter numbers and similar segmentation performance compared to U-Net++.In order to accelerate the search process of federated NAS,this thesis designs three accelerated evaluation methods suitable for federated neural architecture search.The weight sharing method proposed in this thesis for federated NAS enables subnets to directly obtain supernet weights without training,reducing computational and communication costs in federated NAS.In order to improve the accuracy of subnet evaluation using weight sharing methods,this thesis proposes a low fidelity evaluation method suitable for federated NAS.In order to use the historical information of individuals with low fidelity evaluation,this thesis proposes an online surrogate model prediction method for federated NAS,which is used to build a high-quality surrogate model to predict individual performance and further alleviate the problem of expensive evaluation of subnets.Finally,combining FSU-Net++ with the three accelerated evaluation methods mentioned above,this thesis proposes the federated multi-objective evolutionary neural architecture search algorithm Federated Searchable U-NAS(FSU-NAS)to achieve automatic search of subnet architectures with low parameter count and high segmentation accuracy under federated learning settings.Experiments on retinal vascular segmentation dataset and brain tumor segmentation dataset show that the algorithm proposed in this thesis can effectively solve the problem of expensive subnet evaluation and communication cost in federated NAS.In addition,the experimental results also indicate that the algorithm proposed in this thesis can search for subnets with similar performance to manually designed networks under federated NAS,and these subnets have relatively fewer parameter quantities. |
