Contribution of deep learning techniques to holographic microscopy for medical applications

The project lies within the framework of the biological analysis of in vitro blood smears without staining or marking (“label free”) for the early diagnosis of leukemia and malaria and the follow-up of patients. The work aims to detect and characterize cells of interest, especially when they are present in very small proportion, for example, to detect in a population of 200,000 red blood cells a single red blood cell that hosts a parasite.

The originality of the proposed approach lies in the coupling of a phase-sensitive biophotonics device associated with adapted data processing. A biophotonics prototype has been developed by Telecom SudParis (TSP). It allows the production of optical twins (Optical Twin – OT) of microscope slides particularly rich in information. Ways of improving its imaging capabilities in terms of resolution have been identified.

First, we will address them by deep learning techniques associated with the modeling of the sensor physics to further improve the imaging capabilities of the device and produce new images (in terms of resolution). In a second step, the learning approaches will be exploited in order to allow the automatic detection and classification of rare events first on labelled or chemically stained microscope slides, then on “label free” preparations.

The work is thus part of a co-design approach of the technique of acquisition of information from microscope slides (first on blood smears) and of the efficient algorithmic processing of this information by artificial neural networks (AI).

The work of the thesis co-directed by TSP and TRIBVN will be subject to clinical pre-validation.