Reconstruction Interval Z-Phase Dependence of AI Detection Sensitivity in CT Lung Nodule Screening

A new study from Dan Soliman, published on arXiv, reveals a critical blind spot in AI-assisted lung nodule detection: the sensitivity of these systems depends not just on standard acquisition parameters like slice thickness or radiation dose, but on the exact position of the nodule within the CT reconstruction cycle—termed the "z-phase." Analyzing 154 cases from the LIDC-IDRI dataset, Soliman stratified detection sensitivity by reconstruction interval (1mm, 3mm, 5mm) and by the ratio of reconstruction interval to nodule diameter (d/D). At a 5mm reconstruction interval, overall sensitivity dropped to 71.6% compared to 84.8% at 1mm baseline. But more striking, within the 5mm condition, sensitivity varied by a full 17.6 percentage points across z-phase bins. This stochastic effect is invisible to protocol-level quality metrics and not reflected in AI confidence scores.

When stratifying by the d/D ratio, the z-phase effect was most pronounced for nodules where the reconstruction interval was equal to or larger than the nodule diameter (d/D >= 1.0). In this stratum, sensitivity plunged to 61.4%, and the z-phase became the dominant source of per-study detection variance. For smaller nodules (3-6mm) at typical 5mm slice thickness, this means the AI's ability to detect a nodule may be essentially random based on where it falls in the reconstruction cycle. The study underscores that for small nodules, AI-assisted screening may be unreliable if scans use thick slices, and suggests that radiologists and AI developers must account for this hidden variance to avoid missed diagnoses.