Software

Fed-BioMed

I am initiator and scientific responsible for the development of the software Fed-BioMed for federated learning in healthcare. Today, Fed-BioMed has a dedicated engineering team, and enjoys contributions from a developers community through research partnerships and bilateral contracts with industry.

GP-Progression Model

The Gaussian Process Progression Model (GPPM) is a model of disease progression estimating long-term biomarkers’ trajectories across the evolution of a disease, from the analysis of short-term individual measurements. The GPPM software is based on the gradient matching approach proposed in Lorenzi and Filippone, ICML 2018. It has been first presented in the work Lorenzi, NeuroImage 2017, and subsequently extended in the GPPM-DS presented in Garbarino and Lorenzi, IPMI 2019 and Garbarino and Lorenzi, NeuroImage 2021.

Multi-channel Variational Autoencoder

The Multi-Channel Variational Autoencoder was presented in Antelmi et al, ICML 2019. This method extends the VAE framework to work with heterogeneous, i.e. multi-modal, data by projecting all “channels” to a common latent representation. The method adopts variational dropout to learn sparse representation, which are useful to discover in an unsupervised manner the optimal number of dimensions, i.e. the number of ground truth generative factors.

Bayesian Genome-to-Phenome Sparse Regression (G2PSR)

G2PSR is a Bayesian neural network based on a biological inspired architecture to associates genetic data to multiple phenotypic features through biologically inspired constraints (gene structure). The method was publised in Deprez et al, Front Mol Med 2022.

LCC-LogDemons

LCClogDemons is an accurate and robust diffeomorphic registration framework based on stationary velocity fields (SVF) generalizing the log-Demons (Lorenzi et al, NeuroImage 2013. It implements the symmetric Local Correlation Coefficient (LCC) as a similarity measure, and thus it is unbiased with respect to local linear intensity bias of the images.