A Federated Three-Dimensional Deep Learning Method For Multi-Site Brain Image Data

With the development of devices and sensors,data can be collected conveniently and tent to distributed and stored into multiple sites.So that it is essential to take advantage of multi-site data for the data-driven models learning currently.Distributed learning makes it possible to train a central deep learning model with multi-site data.However,the scale and quality of multi-site data may be different,and the privacy sensitive may preclude using the central server or training central models.Federated learning as a emerging paradigm that allows multi-site data to train a federated deep learning models collaboratively.It can make full use of local computing resources and privacy data without data sharing.In this paper,a weighted guide-gradient federated deep learning approach for multi-site3 D image data is proposed.Then,a multi-modal data fusion mechanism is added in this approach.Additionally,there is also a federated gradient matching domain adpatation method with single source domain were proposed for eliminating the harmfuless of domain shift when training federated models.The main work of the paper includes:Proposed a weighted guide-gradient federated deep learning approach for multisite 3D image data.This paper added self-adapting weight parameters to the aggregation mechanism of multi-site sharing information and used its guide-gradient to train a specific federated model for each site.So that the specific federated model has better performance in local data classification tasks.This paper valid the proposed approach in the multi-site s MRI image data.The results have shown the federated deep learning models trained by weigted guide-gradinet approach have outperformed and more robust than the other methods in the patient classification of the MDD and ABIDE.Proposed a multimodal data fusion mechanism into the weighted guide-gradient federated deep learning approach.This paper added the multimodal data fusion mechanism into the weighted guidegradient federated deep learning approach,and tested on the classification of multimodal brain MRI image data.The results have shown that the average diagnostic accuracy of MDD patients achieved 83.19% when using multimodal brain MRI data,and it has better performance than training federated models with single mode MRI data.Proposed a federated gradient matching domain adaptation method with single source domain.This paper used one dataset with large scale as single source domain and multiple federated sites as multi-target domains to extract domain-invariant features for specific the site using adversarial domain dapatation.Then adding gradient matching loss as the constraint condition of objective function for federated parameters updating.The method was test on the gender classification while using HCP dataset as source domain,NKI and SLIM dataset as target domains.The results show that this approach has a stable and obvious improvement than the other methods.The main contributions of this paper are: It proposed the weighted guide-gradient federated deep learning method for multi-site 3D MRI image data with different modal.Additionally,it proposed the federated gradient matching domain adpatation method to tackle the problem of domain shift across different federated sites.The proposed approach has been valided on the multi-site 3D brain MRI image data for gender classification and psychiatry disorders diagnosis.The results show that the proposed method can protect the data privacy and the training efficiencies while achieve the better accuracy and robustness federated models for different sites on different ta sks.Moreover,it can be extended to many scenarios and provides a new method to cooperatively train high-quality federated models in a security way for multi-site image data.