Abstract
Deep Probabilistic Imaging (DPI) addresses the limitations of traditional point-estimate algorithms in underdetermined imaging problems. By leveraging untrained, flow-based generative models, DPI characterizes the entire posterior distribution of images consistent with sparse or noisy data. This approach enables researchers to quantify uncertainty and identify multiple valid scientific interpretations, a critical capability for high-stakes fields such as black hole imaging and medical diagnostics.
Introduction
In computational imaging, many systems are "underdetermined," meaning the collected data is insufficient to uniquely define a single image. Traditional methods typically output a single "best guess," which can mask the existence of alternative, equally valid solutions. For unique targets like the M87* black hole, where massive training datasets are unavailable, standard deep learning methods are often inapplicable. DPI bridges this gap by combining variational inference with deep generative models to provide a probabilistic map of possible solutions.
Methodology
DPI utilizes an untrained generative model that optimizes network weights to fit a specific dataset, rather than learning from a large external corpus. The core of the framework is a flow-based model (Real-NVP), consisting of invertible neural networks that allow for the exact calculation of probability density.
The model is trained using a variational loss function that minimizes the Kullback–Leibler (KL) divergence between the generated distribution and the true posterior. This loss function balances data-fitting, regularization (image priors), and an entropy term. A tunable β parameter controls the weight of the entropy term, allowing researchers to adjust the diversity of the generated samples.
Experiments and Results
The framework was evaluated on toy potentials, radio interferometry data (including synthetic and real EHT observations), and medical MRI data from the fastMRI dataset.
Uncertainty Quantification and Multi-modality
A key insight of DPI is its ability to capture "multi-modality"—the existence of distinct types of solutions that fit the data. In black hole imaging simulations, DPI identified multiple valid modes, such as 180-degree rotations of the primary image, which traditional methods might overlook.
Validation on EHT and MRI Data
When applied to real M87 data, DPI successfully recovered the characteristic "orange donut" shape. Crucially, it produced a standard deviation map* highlighting regions of high uncertainty, such as the southern portion of the ring. In MRI tests, DPI demonstrated that as data becomes sparser (higher acceleration speeds), the uncertainty maps grow logically, pinpointing areas where reconstruction is most likely to fail.
Discussion and Limitations
DPI provides a robust framework for automated probabilistic imaging without the need for large training sets. However, challenges remain. Current implementations are limited to lower resolutions (e.g., 64 x 64 pixels) due to the computational demands of flow-based models. Additionally, the model can sometimes struggle to transition between highly disconnected modes in the probability space. Future work will focus on scaling these models to higher resolutions and incorporating more sophisticated learned priors to further refine the posterior distributions.
Summary updated