A hallucination is image content produced by the prior rather than supported by the data: a plausible-looking feature that is not there, or a missing feature that is.
Why EIT is especially exposed. Boundary data determine the interior only weakly (see Stability of the Calderón Problem and Decay of Boundary Measurements). Many very different interiors fit the data equally well within the noise, and a strong generative prior will “fill in” the interior with typical training content.
Mitigations.
- Data consistency checks: the final image must reproduce the measured data to within the noise (discrepancy principle). Hard projection or prox steps onto the data-consistent set are stricter than soft penalties.
- Physics-aware priors: constrain the prior’s updates to directions the data cannot see, or penalise changes that alter the predicted currents. One option is an operator-weighted metric in the prox, with built from the forward operator.
- Uncertainty quantification: draw multiple posterior samples (different noise seeds in Diffusion Posterior Sampling or DiffPIR) and report the pixel-wise mean and variance. High-variance regions are where the prior dominates. Calibrated methods, such as conformal prediction, give coverage guarantees.
- Prior mismatch testing: evaluate on out-of-distribution conductivities.
References
- S. Bhadra, V. A. Kelkar, F. J. Brooks, M. A. Anastasio (2021). On Hallucinations in Tomographic Image Reconstruction. IEEE Trans. Med. Imaging 40(11), 3249–3260. doi:10.1109/TMI.2021.3077857
- V. Antun, F. Renna, C. Poon, B. Adcock, A. C. Hansen (2020). On instabilities of deep learning in image reconstruction and the potential costs of AI. PNAS 117(48), 30088–30095. doi:10.1073/pnas.1907377117