Per-Bundle, Pre-Reconstruction Inversion for Temporal CT: Flux-Compensation Design and Training-Free Empirical-Bayes Estimation
Abstract
In Temporal computed tomography (TCT) each detector reading sums the photons of up to three rays. Earlier work set the resulting per-bundle limits and tested a learned prior on one patient at one incident count $N_0$; this paper lifts both. N$_0$FC ($N_0$ flux compensation), a bow-tie and tube-current-modulation co-design, equalizes each bundle's detected flux. SNN2 is a per-bundle Poisson maximum-likelihood estimator. V8c is a neural-network corrector trained on 19 patients stratified by water-equivalent diameter (WED). T2bp, a training-free empirical-Bayes post-processor, combines per-bundle Poisson likelihoods with a prior built from the patient's own bundles. Results use that work's Dose Inflation Factor (DIF), charging bias as error. On 23 non-test volumes, N$_0$FC lowers the per-volume median middle-path Cramer-Rao floor, as a DIF, from 23-48 under a flat-fluence (no bow-tie) baseline to 3.3-6.3 at 35-62% of its incident photons, near the equal-attenuation value 13/3 and 1.4-1.7 times below conventional bow-ties; the bow-tie alone achieves this median, and tube-current modulation trims the tail. For bundles with $X_2<6$ (83%), SNN2 attains this floor (median mean-squared-error ratio 1.02); training-free T2bp$\circ$SNN2 lowers the ratio to 0.89 (0.74-1.26). It is best or tied on three of four N$_0$FC test cells (WED 207-342 mm) and trails T2bp$\circ$V8c by 3% on the fourth. V8c beats SNN2 in distribution under baseline acquisition, gains at most 9% under N$_0$FC (losing 14% on the largest in-distribution test patient), and fails on a patient acquired below its training count level. Acquisition co-design removes most of the attenuation-contrast penalty, and training-free estimation attains the resulting limit. The remaining three- to six-fold gap to single-source CT at equal fluence, and the high-attenuation tail, are targets for inter-bundle priors.
- Comment
- v2: title changed from v1 (was "...: Flux-Compensation Design and Multi-Patient Generalization"); corpus results recomputed with the truth-free T2bp configuration; conventional bow-tie comparison and acquisition and bias-structure tests added. 72 pages (30 main text, 42 supplementary material), 8 figures, 19 tables. Paper 2 of a series; companion to arXiv:2604.10934