Addressing Conflict between Prior and Data Fidelity in PET Image Reconstruction

I gave an oral talk at the IEEE NSS-MIC (Nuclear Science Symposium-Medical Imaging Conference) covering the broad topic of resolving conflict between data fidelity and regularization (prior) terms in unsupervised PET image reconstruction.

In the first section of the talk, I demonstrated Steerable Conditional Diffusion for PET Image Reconstruction (paper link and details here). This work builds on the work of Barbano et al. (2025), demonstrating that the diffusion prior can be adapted to new domains during reconstruction (at inference-time) in a PET image reconstruction context.

In the second section of the talk, I introduced the Distributional Consistency (DC) loss, a new data fidelity objective for noisy inverse problems. In the PET image reconstruction context, DC loss has a number of advantages over traditional likelihood maximization:

• DC loss avoids overfitting to noisy measurement data,
• DC loss specifies an “ideal” level of data consistency (i.e. DC loss = 0).

The DC loss can be viewed as generalizing likelihood-maximization from a pointwise process to a distributional process. Correspondingly, it only works when we have many independent measurements and know the noise distribution on each of them. These assumptions are met in PET image reconstruction, as we have large sinograms of independently noisy measurements and the Poisson noise model on each measurement.

This work is currently under review, and is available as a preprint (paper link and details here).

We are currently undertaking further work to demonstrate the efficacy of this work with additional regularization strategies, such as diffusion models.

References:

• R. Barbano et al., “Steerable Conditional Diffusion for Out-of-Distribution Adaptation in Medical Image Reconstruction,” in IEEE Transactions on Medical Imaging, vol. 44, no. 5, pp. 2093-2104, May 2025, doi: 10.1109/TMI.2024.3524797.