Introduction
Multispectral Photoacoustic Imaging (PAI) is a non-invasive technique that utilizes laser pulses to generate ultrasound waves, enabling the mapping of internal biological structures and functional parameters such as blood oxygenation ($sO_2$). However, the reconstruction of these properties is fundamentally "ill-posed" due to light scattering and fluence loss within tissue. Standard deep learning approaches typically provide point estimates—single "best guesses"—which can be misleading in ambiguous scenarios. If two distinct physiological states produce similar pressure measurements, a standard model may average them into a physically impossible result or arbitrarily select one. This research introduces a framework using Conditional Invertible Neural Networks (cINNs) to quantify this uncertainty by generating full posterior probability distributions.
Methodology: Conditional Invertible Neural Networks
To address the limitations of point estimation, the authors employ Conditional Invertible Neural Networks (cINNs). Unlike traditional feed-forward networks, INNs are bijective and mathematically reversible. The architecture utilized in this study consists of 20 blocks with GLOW coupling layers.
The model maps the target parameter ($sO_2$) to a latent space following a standard normal distribution, while being conditioned on the observed initial pressure spectrum. During inference, the latent space is sampled multiple times for a given pressure spectrum, allowing for the reconstruction of the full posterior probability density. To automate the detection of ambiguous results, the researchers developed a multi-mode score based on kernel density estimation, which identifies when the output distribution contains multiple peaks (modes).
Experimental Setup and Scenarios
The study was conducted in silico using the Monte Carlo eXtreme (MCX) framework to ensure access to ground truth data. Simulations involved $10^7$ photons across 26 wavelengths (700 nm to 950 nm). Three primary scenarios were tested:
- S1 (Single Vessel): A baseline scenario featuring a single blood vessel in muscle tissue.
- S2 (Multiple Vessels, Single IU): Designed to create ambiguity, where one vessel shields another from a single illumination unit (IU).
- S3 (Multiple Vessels, Multiple IUs): A complex setup with three simultaneous illumination units to test if hardware redundancy can resolve ambiguity.
Results: Quantifying and Resolving Ambiguity
A key insight from the study is that uncertainty in PAI is often multi-modal rather than merely "fuzzy." In the S2 scenario, the posterior distribution frequently exhibited two distinct peaks, indicating that the model recognized two likely but different oxygenation levels.
The researchers demonstrated that this ambiguity is not just a software challenge but a design problem. By optimizing the hardware configuration—specifically by moving from a single light source (S2) to three (S3)—the multi-modal peaks collapsed into a single, accurate peak corresponding to the ground truth. Furthermore, the model's uncertainty estimates were found to be "well-calibrated," with a median calibration error of less than 1.5 percentage points (p.p.), meaning the predicted confidence intervals accurately reflected the true probability of correctness.
Conclusion and Future Work
This work highlights the power of cINNs in providing a transparent view of uncertainty in medical imaging. By identifying multi-modal distributions, clinicians can recognize when a measurement is unreliable and adjust the acquisition process accordingly. While the current study is limited to simulations, it lays the groundwork for "smart" photoacoustic probes. Future research will focus on translating these findings to real-world experimental settings and potentially integrating reinforcement learning to enable probes to automatically find the optimal "pose" for maximum certainty.
Summary updated