Smart healthcare relies on artificial intelligence (AI) functions for learning and analysis of patient data. Since large and diverse datasets for training of Machine Learning (ML) models can rarely be found in individual medical centers, classical centralized AI requires moving privacy-sensitive data from medical institutions to data centers that process the fused information. Training on data centers thus requires higher communication resource/energy demands while violating privacy. This is considered today as a significant bottleneck in pursuing scientific collaboration across trans-national clinical medical research centers. Recently, federated learning (FL) has emerged as a distributed AI approach that enables the cooperative training of ML models, without the need of sharing patient data. This paper dives into the analysis of different FL methods and proposes a real-time distributed networking framework based on the Message Queuing Telemetry Transport (MQTT) protocol. In particular, we design a number of solutions for ML over networks, based on FL tools relying on a parameter server (PS) and fully decentralized paradigms driven by consensus methods. The proposed approach is validated in the context of brain tumor segmentation, using a modified version of the popular U-NET model with representative clinical datasets obtained from the daily clinical workflow. The FL process is implemented on multiple physically separated machines located in different countries and communicating over the Internet. The real-time test-bed is used to obtain measurements of training accuracy vs. latency trade-offs, and to highlight key operational conditions that affect the performance in real deployments.

No abstract

In mission-critical verticals such as automated driving, 5G-advanced networks must provide centimeter-level dynamic positioning along with ultra-reliable low-latency communication services. Massive Multiple-Input Multiple-Output (mMIMO) and millimeter waves (mmWave) are the key enablers, allowing high accuracy angle and delay estimation. Still, extracting such information from highly-dimensional Channel Impulse Responses (CIRs) results in a complex task, due to channel sparsity and intermittent blockage. In this paper we focus on non-lineof-sight (NLOS) identification from CIR data, proposing a Deep Autoencoding Kernel Density Model (DAKDM) to characterize the statistics of the channel latent features. We formulate the problem as a semi-supervised anomaly detection task in which only LOS samples, i.e., normal data, are adopted for training. DAKDM is a single-stage training model that takes as input the full CIR thanks to an AutoEncoder (AE) structure. The proposed method is able to learn the latent distribution by means of a Kernel Density Estimator (KDE) in combination with a deep learning likelihood network. We validate the proposed solution in a 5G Urban micro (UMi) vehicular scenario. Results show that the proposed model can significantly outperform conventional algorithms and obtain similar performances to variational Bayes algorithms at one tenth of the inference time.

No abstract

No abstract