Rain retrieval from spaceborne microwave radiometers remains a challenge despite de progress made since TRMM and GPM. These two missions both combine a passive microwave radiometer and one (TRMM) or two (GPM) rain radar with a common swath. Most rain retrieval algorithms use these co-located data to build a database of individuals made of rain rates and their associated brightness temperature vector. Then, the retrieval process itself relies mostly on a Bayesian approach to select the relevant individuals in the database and offer a solution to a measured input brightness temperature vector. Eventually, in the database a series of ancillary data can also be stored and used for the retrieval: information on the temperature profile, on the surface characteristics, on the humidity profile etc... After almost 25 years of development, these Bayesian approaches offer good performances but they do not exploit fully the topology of the brightness temperature fields or the rain field. Using deep-learning techniques on co-located DPR and GMI data, we were able to develop an retrieval algorithm that rely exclusively on brightness temperatures and offers performances close to those of Gprof. An assessment of these performances will be presented with a focus over continental France using the Météo-France radar mosaic.