Lisa Cuneo: Explainable-by-design machine learning model for unmixing fluorescence signal based on fluorescence lifetime
Speaker:
Lisa Cuneo
Data dell'evento:
Lunedì, 15 April, 2024 - 10:30
Luogo:
Room B203
Contatto:
siciliano@diag.uniroma1.it
Abstract:
In fluorescence microscopy, the capability to distinguish between various fluorophores is highly desirable and essential to perform multicolor imaging. The ability to discriminate between different fluorophores based on their temporal fingerprints, which are independent of their emission spectra but related to fluorescence lifetime information, is significantly advantageous. In fact, exploiting temporal information, would be possible to multiplex/demultiplex the fluorescence signal of spectrally overlapping fluorophores. Moreover, the lifetime provides insights into the underlying biological phenomena, therefore would be useful to estimate its value.
Over the years, several methods have emerged for distinguishing fluorophores using either tempo- ral or spectral fingerprints. Notable techniques operating in the frequency domain include the phasor approach [1] and SPLIT [2]. We introduce a deep learning approach designed to separate the signal contributions coming from two fluorophores, potentially spectrally overlapping, in time-resolved flu- orescence microscopy images. The proposed method leverages a CNN-based network based on both temporal features and 2D spatial characteristics within the images. Since the purpose of the network is to separate contributions from different fluorophores and to estimate the lifetime values, the neural network is divided into two sections, each separating one of the two different fluorescence time decay components. We focus in enhancing the explainability of the neural network architecture incorporating the physical model, rendering our approach highly interpretable and enabling a deeper comprehension of the underlying mechanisms.
This research paves the way for our comprehension of sub-cellular components and macro-molecular complexes by facilitating simultaneous labeling and imaging of diverse bio-molecules. The introduced deep learning method effectively addresses spectral overlap challenges, by providing a more precise and adaptable analysis of fluorescence microscopy data using fluorescence lifetime.
Short bio:
Lisa Cuneo is a postdoctoral researcher at the Italian Institute of Technology (IIT). She holds a a Master's degree in Applied Mathematics and Ph.D. in Physics and Nanosciences from the University of Genova, Italy, where her research focused on computational analysis of biomedical images by means of machine learning algorithms and inverse problems techniques.
Lisa's academic career has included visiting positions at prestigious institutions such as the University of Cambridge and Aalto University, where she honed her skills in computational biology and MEG inverse problems respectively.