Dynamic Inverse Problems in Time-Lapse Microscopy

Microscopy technology has undergone significant development in recent years. Light optical microscopy, for instance, now offers superresolution techniques that produce images with unprecedented resolution beyond the diffraction limit. 
The acquisition speed of new technologies has increased up to hundreds of frames per second, allowing time-lapse imaging of events as they unfold in real time. Modern microscopy techniques are increasingly being adopted by the biomedical research community and are becoming an indispensable tool for understanding biological processes at the molecular level. However, quantitatively analyzing processes with these techniques presents new challenges. Complex biological processes are inherently dynamic as they evolve in space and time. A short acquisition time, necessary to minimize the effects of motion and harmful radiation (photobleaching and phototoxicity), combined with high spatial resolution, results in a low signal-to-noise ratio (SNR < 1dB), which makes any image processing highly unstable. Quantitative analysis of time-lapse imaging in biomedical experiments requires a solution of various inverse problems, of which the most important are to automatically detect objects of interest and track their behavior in time. Employing machine learning techniques that are capable of generalization is indispensable, as it enables adaptation to new challenges without extensive manual input, a task that proves arduous in noisy, ever-changing environments.

The goal of the project is to propose a robust and universal solution of dynamic inverse problems with temporal structural models to disentangle complex biological processes at the molecular level.
We identify four objectives that jointly lead to the goal:

1.  Robust Reconstruction of High-Resolution Data: Tackling very low SNR scenarios and time-dependent measurement processes using model-driven approaches with Deep Unrolling
2.  Optimal Detection and Tracking: Achieving simultaneous and globally optimal detection and tracking of numerous textureless objects using the concept of dense hypotheses.
3. Generalizable Models: Creating models that can adapt to new challenges without extensive manual annotation using self-supervised learned embeddings.
4. Model Validation: Validating the developed models against biological problems, in collaboration with cell biologists from Cambridge Institute for Medical Research.

Structured Illumination Microscopy and particularly the total internal reflection fluorescence variant (TIRF-SIM) with high spatial and temporal resolution, low laser illumination power, and reduced phototoxicity is ideal for live-cell imaging. We develop an analysis pipeline for TIRF-SIM time-lapse movies that allows to investigate new biological questions related to, e.g., the formation and dynamic behavior of endocytic and exocytic vesicles on the plasma membrane.