A score can rank when CT systems fail without warning

A learned inference system may report a narrow uncertainty interval and pass its own checks even when the true value is outside that interval. Can this kind of silent failure be anticipated before the system sees the target measurements, especially when some regions cannot be measured?

The paper develops a pretrial score that estimates which image errors may escape checks and fall outside a CT system’s uncertainty interval.

The score was computed before any trial and compared with the realised silent-failure rates of 240 designed errors. Each error was tested on five fixed CT slices with two noise draws per slice.

The score ranked the outcomes for both learned systems, showing that a pretrial estimate can order errors by their observed risk in this bounded-reach setting.

That supports prospective ranking, but not a distinct advantage from learned-prior geometry. In the main and robustness studies, a physics-based predictor ranked errors about as well.

Under attenuated access, results differed by condition: at the weakest attenuation, a Gaussian predictor ranked at least as well as the learned score; at the pre-specified F=128,000 condition, VAE-EKI’s learned score beat the Gaussian on rankings; at F=512,000, their rankings were alike.

Trial-level comparisons found no detectable gain beyond physics in bounded-reach studies or at interpretable attenuation strengths.

The ranking evidence comes from two learned systems, selected errors and simulated measurements on slices from one CT data source. The study does not show that the score gives calibrated probabilities or works on independently acquired objects or real restricted-access measurements.