Skip to main content
Go to the home page of the European Commission (opens in new window)
English en
CORDIS - EU research results
CORDIS

Model-aware learning for imaging inverse problems in fluorescence microscopy

Project description

Imaging inverse problems in fluorescence microscopy

Fluorescence microscopy imaging involves a wide range of inverse problems in which meaningful representations of biological samples must be reconstructed from incomplete and noisy measurements. These problems are inherently ill-posed, leading to solutions that are highly sensitive to small perturbations. Existing model-based approaches rely on expert knowledge of the underlying reconstruction procedures to ensure stability, as well as hand-crafted optimisation methods, which limits their practical applicability. The ERC-funded MALIN project addresses these limitations by developing an integrative framework that combines model-based methods with data-driven deep-learning techniques. Methodologically, the project focuses on optimisation-driven approaches that incorporate available physical knowledge while accounting for modelling inaccuracies. The resulting methods will enhance efficiency, accuracy and cost-effectiveness, and will be disseminated through open-source software tools.

Objective

This project will develop model-aware, i.e. physics-informed, learning methods for solving imaging inverse problems (IIPs) in fluorescence microscopy imaging (FMI). IIPs are frequently encountered in FMI whenever a visual representation of a biological sample needs to be reconstructed from incomplete and noisy input measurements. Such IIPs are typically ill-posed: their solution (if it exists) is unstable to perturbations. Classical model-based approaches reformulate the IIP at hand as an energy minimisation task. Such approaches rely both on the (approximate) knowledge of the complex physical processes involved and on the mathematical design of hand-crafted optimisation methods whose tuning is often very time-consuming. Concurrently, the impressive development of machine and deep learning methods has enabled the applied imaging community with new data-driven methodologies providing unprecedented results in tasks such as image classification. The performance of data-driven methods for solving IIPs in FMI, however, is halted by their intrinsic unstable behaviour. In MALIN, I propose an integrative paradigm where the stable performance of model-based approaches is combined with the effectiveness of data-driven techniques by means of shallow model-constrained learning and deep physics-informed generative approaches. The reliability of the model-aware methods proposed will be justified by theoretical results providing reconstruction and convergence guarantees. The study will further account for possible geometric invariances and imperfect physical modelling, showing robustness to modelling errors which are frequent when standard (low-cost) equipment is used. Algorithmic acceleration strategies and inexact/stochastic algorithms will be devised to guarantee efficient performance also under limited computational resources and training data. The methodologies will be deployed on several IIPs in FMI and democratised through the release of open software and plug-ins.

Fields of science (EuroSciVoc)

Programme(s)

Topic(s)

Funding Scheme

HORIZON-ERC - HORIZON ERC Grants

See all projects funded under this funding scheme

Call for proposal

(opens in new window) ERC-2023-STG

See all projects funded under this call

Host institution

UNIVERSITA DEGLI STUDI DI GENOVA
Net EU contribution
€ 1 432 734,00
Address
VIA BALBI 5
16126 GENOVA
Italy
Region
Centro (IT) Lazio Roma
Activity type
Higher or Secondary Education Establishments
Links
Total cost
€ 1 432 734,00

Beneficiaries (1)