In an ideal physics experiment, a signal decays simply. A single fluorescent molecule blinks out, or water molecules in a uniform tube diffuse at a predictable rate. These processes follow a mono-exponential decay: a single curve defined by a single parameter.
However, biological reality is rarely so tidy. Within a single pixel of a microscope or a single voxel of an MRI, multiple environments coexist. A placenta contains maternal blood, fetal blood, and stationary tissue; a cell contains multiple fluorescent markers or a single marker in different chemical states. When we measure these, their signals overlap, creating a multi-exponential decay.
Recent research across magnetic resonance imaging (MRI) and fluorescence lifetime imaging (FLI) is converging on a shared goal: moving beyond oversimplified "single-rate" models to accurately disentangle these overlapping signals.
The Failure of the Mono-Exponential Assumption
For decades, standard analytical models have relied on mono-exponential or bi-exponential assumptions because they are computationally easy to fit. But as our sensors improve, these models are being exposed as inadequate.
In fluorescence imaging, Marek Gersbach and Edoardo Charbon used high-resolution CMOS detectors to show that even common indicators like OGB-1, previously thought to follow a double-exponential decay, actually follow a triple-exponential model. Ignoring that third component leads to significant errors in measuring calcium concentrations.
A similar problem plagues MRI. When mapping the prostate or the placenta, researchers often assume water relaxes at a single rate ($T_2$). However, Patrick J. Bolan and colleagues have noted that mono-exponential models fail to capture the heterogeneity of biological tissues. In the placenta specifically, Paddy J. Slator and Daniel C. Alexander argue that we must transition to multi-compartment modeling to distinguish between blood flow and tissue diffusion.
Disentangling Compartments
The core challenge is "unmixing." If you have a signal that is the sum of three different decays, how do you work backward to find the original three rates?
In diffusion MRI, this is often framed as a "compartment" problem. Rémy Gardier and Jean-Philippe Thiran have shown that in "leaky" tissues like tumors, assuming cell membranes are impermeable is a fatal flaw. Their CEXI model accounts for water exchange between three or more compartments, providing a much more accurate picture of microstructure than simpler models.
In the placenta, this multi-compartment approach allows researchers to detect Fetal Growth Restriction (FGR). By decoupling the movement of water in tissue from the movement of blood in vascular networks, Slator demonstrated that restricted placentas maintain a more "ordered" (anisotropic) structure, failing to branch out into the complex, multi-directional "villous trees" seen in healthy pregnancies.
Real-Time Estimation with Deep Learning
Traditional methods for fitting multi-exponential curves are "offline"—they take hours of computation. To bring these insights into the operating room or live cell imaging, researchers are turning to neural networks.
The team of Vytautas Žičkus and Alex Turpin pioneered the use of Artificial Neural Networks (ANNs) to process fluorescence data 1,000 times faster than traditional fitting. However, even these early networks often assumed mono-exponential decay for simplicity.
The next generation of models, such as MFliNet, developed by Vikas Pandey and İsmail Erbaş, uses Transformer architectures to handle more complex signals. By incorporating pixel-wise instrument response functions (IRFs), these models can account for the "topography" of a sample—like the curves of an organ—while attempting to resolve bi-exponential and eventually multi-exponential decay patterns.
The Path Forward
The research direction is clear: the "mono-exponential" era of quantitative imaging is ending. Whether it is Andrei Ardelean implementing RNNs on FPGAs for real-time photon processing or Mengze Gao using zero-shot learning to reconstruct MRI data, the focus is on capturing the true, multi-faceted nature of biological signals.
The challenge remains in the "resolution" of these models. As Erbaş notes, while we can now process data at video rates, ensuring these models don't collapse complex multi-exponential signals into a single "average" value is the next frontier for high-fidelity biological mapping.