arxiv: 2605.00971 · v1 · submitted 2026-05-01 · 📡 eess.IV · cs.CV

Recognition: unknown

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

Dan Soliman

Authors on Pith no claims yet

Pith reviewed 2026-05-09 18:05 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords lung nodule detectionAI sensitivityCT reconstructionz-phasereconstruction intervalLIDC-IDRIpartial volumedeep learning

0 comments

The pith

AI detection sensitivity for lung nodules depends on the ratio of reconstruction interval to nodule diameter.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that AI systems for finding lung nodules on CT scans have detection rates that change based on where the nodule sits inside the reconstruction cycle. This z-phase effect becomes large when the slice interval is similar to or bigger than the nodule size, as with 5 mm reconstructions for 3-6 mm nodules. Sensitivity then swings by 17.6 percentage points across phase positions, and overall falls from 84.8 percent at 1 mm slices to 71.6 percent at 5 mm. The variation appears only when the interval-to-diameter ratio reaches or exceeds one and stays hidden from usual scan quality checks or AI scores. The result comes from re-examining a 154-case study on the LIDC-IDRI set, grouping cases by reconstruction interval and the d/D ratio.

Core claim

AI detection sensitivity depends on the ratio of reconstruction interval to nodule diameter. When this ratio approaches or exceeds 1.0 -- as occurs for 3-6mm nodules at 5mm reconstruction -- z-phase becomes the dominant source of per-study detection variance. This stochastic effect is invisible to protocol-level quality metrics and not reflected in AI confidence scores.

What carries the argument

z-phase, the fractional position of the nodule center within the reconstruction cycle folded to [0, 0.5], which controls detection probability once the reconstruction interval to nodule diameter ratio reaches or exceeds one.

If this is right

Sensitivity falls to 71.6% at 5 mm reconstruction from 84.8% at 1 mm.
At 5 mm reconstruction, sensitivity differs by 17.6 percentage points across z-phase bins.
Sensitivity is 92.4% when d/D is below 0.5, 78.0% when 0.5 to 1.0, and 61.4% when 1.0 or above.
Systematic z-phase dependence occurs only in the d/D greater than or equal to 1.0 group.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Screening programs could reduce missed small nodules by defaulting to thinner reconstruction intervals.
AI models might gain stability if trained on data that includes varied z-phase positions for borderline-sized nodules.
Reported performance gaps between different sites may partly trace to unmeasured differences in reconstruction settings.
Controlled experiments that shift z-phase while holding other scan parameters fixed could isolate the effect further.

Load-bearing premise

That z-phase can be accurately computed from nodule center positions in the retrospective LIDC-IDRI data and that the prior 154-case perturbation study provides an unbiased sample for stratification by reconstruction interval and d/D ratio.

What would settle it

Re-analysis of the same nodules at known z-phases showing no statistically significant sensitivity differences across bins when d/D is 1.0 or greater would falsify the dependence on z-phase.

Figures

Figures reproduced from arXiv: 2605.00971 by Dan Soliman.

**Figure 1.** Figure 1: Reconstruction interval z-phase sensitivity analysis. view at source ↗

**Figure 2.** Figure 2: Interval/diameter ratio (d/D) analysis, pooling all reconstruction interval conditions. view at source ↗

read the original abstract

Background: Sensitivity of AI-assisted lung nodule detection systems is known to vary with CT acquisition parameters including radiation dose, reconstruction kernel, and slice thickness. However, the dependence of detection probability on nodule position within the reconstruction cycle -- the z-phase -- has not, to the author's knowledge, been characterized for deep learning-based detection systems. Methods: A retrospective analysis was performed using the LIDC-IDRI dataset. Detection results from a previously validated 154-case perturbation study were re-analyzed. For each consensus nodule (>=4-reader agreement), z-phase was defined as the fractional position of the nodule center within the reconstruction cycle, folded to [0, 0.5]. Detection sensitivity was stratified by z-phase bin, reconstruction interval (1mm, 3mm, 5mm), and by the ratio of reconstruction interval to nodule diameter (d/D). Results: At 5mm reconstruction interval, sensitivity was 71.6% vs 84.8% at 1mm baseline. Within the 5mm condition, sensitivity varied by 17.6 percentage points across z-phase bins. Stratified by d/D ratio, sensitivity was 92.4% for d/D < 0.5, 78.0% for 0.5 <= d/D < 1.0, and 61.4% for d/D >= 1.0, with a systematic z-phase effect present only in the d/D >= 1.0 stratum. Conclusions: AI detection sensitivity depends on the ratio of reconstruction interval to nodule diameter. When this ratio approaches or exceeds 1.0 -- as occurs for 3-6mm nodules at 5mm reconstruction -- z-phase becomes the dominant source of per-study detection variance. This stochastic effect is invisible to protocol-level quality metrics and not reflected in AI confidence scores.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This re-analysis of LIDC data flags a z-phase effect on AI nodule detection that appears mainly when reconstruction interval matches or exceeds nodule diameter, but the numbers lack supporting stats and method details.

read the letter

The main point to take away is that this work shows AI-based lung nodule detection sensitivity can shift by as much as 17.6 percentage points depending on the nodule's z-phase at 5 mm reconstruction intervals, and that this effect kicks in mainly when the reconstruction interval is at least as large as the nodule diameter. What is new here is the focus on z-phase for these deep learning detectors. Earlier studies looked at dose, kernel, and thickness, but the dependence on position within the reconstruction cycle had not been broken out this way. The re-analysis of the LIDC-IDRI perturbation study does a good job of stratifying the results by reconstruction interval and by the d/D ratio, highlighting that the z-phase variation is isolated to the d/D >= 1.0 group where sensitivity drops to 61.4%. The paper is straightforward in its approach and uses an existing public dataset, which is a plus for reproducibility. The numbers are presented clearly enough to see the pattern. That said, there are some gaps. The abstract does not include any statistical tests or confidence intervals around those sensitivity figures, and there are no sample sizes listed per bin, so it is difficult to know how reliable the 17.6 pp difference really is. Details on exactly how z-phase was calculated from the nodule centers are missing too. If the original perturbation study relied on post-processing shifts rather than actual re-reconstruction from projections, the physical effects might not be captured accurately, which would affect how much we can trust the causal interpretation. This paper would be of interest to researchers in medical imaging AI who are thinking about sources of variability in detection performance, or to those involved in standardizing CT protocols for screening. It is not a complete story on its own, but it flags something worth checking in larger studies. I would send it to peer review. The core observation is worth a closer look from referees who can ask for the missing stats and method clarifications.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a retrospective re-analysis of detection results from a previously validated 154-case perturbation study on the public LIDC-IDRI dataset. It claims that AI lung nodule detection sensitivity depends on the ratio of reconstruction interval to nodule diameter (d/D); when d/D approaches or exceeds 1.0 (as for 3-6 mm nodules at 5 mm reconstruction), z-phase (fractional nodule-center offset within the reconstruction interval, folded to [0, 0.5]) becomes the dominant source of per-study variance, producing a 17.6 percentage-point sensitivity swing at 5 mm reconstruction (overall 71.6% vs. 84.8% at 1 mm baseline) that is absent in lower d/D strata and invisible to protocol metrics or AI scores.

Significance. If the central stratification result holds after clarification, the work identifies a previously uncharacterized, stochastic, and protocol-independent source of detection variability in deep-learning nodule detectors. The grounding in a public dataset and re-use of an existing perturbation study is a clear strength for reproducibility and falsifiability; the d/D-threshold observation supplies a concrete, testable prediction for future protocol design and AI evaluation.

major comments (2)

[Methods] Methods: The description of z-phase computation from LIDC-IDRI nodule-center annotations and DICOM metadata does not include validation steps (e.g., precision of center coordinates, handling of reconstruction-interval metadata, or inter-observer variability in consensus annotations). Because the headline 17.6 pp z-phase effect and the d/D-stratum interaction rest entirely on correct bin assignment, this omission is load-bearing for the causal claim.
[Methods] Methods: The 154-case perturbation study is not characterized as to whether z-phase shifts were generated by true re-reconstruction from raw projections or by post-hoc image-domain operations (slice shifting or interpolation). Only the former captures the partial-volume and aliasing physics that actually couple z-phase to detection; the latter would misrepresent the mechanism and undermine the interpretation that the effect appears only for d/D >= 1.0.

minor comments (2)

[Abstract] Abstract and Results: Sensitivity percentages (71.6%, 84.8%, 92.4%, 78.0%, 61.4%) and the 17.6 pp difference are reported without per-stratum sample sizes, confidence intervals, or any statistical test, preventing immediate assessment of whether the z-phase variation is robust or driven by small bins.
[Abstract] Abstract: The background statement that sensitivity varies with dose, kernel, and slice thickness cites no references, which is a minor but unnecessary omission in an otherwise concise summary.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment of our work's significance and for the constructive major comments on the methods. We address each point below and will revise the manuscript to incorporate clarifications and additional details.

read point-by-point responses

Referee: [Methods] Methods: The description of z-phase computation from LIDC-IDRI nodule-center annotations and DICOM metadata does not include validation steps (e.g., precision of center coordinates, handling of reconstruction-interval metadata, or inter-observer variability in consensus annotations). Because the headline 17.6 pp z-phase effect and the d/D-stratum interaction rest entirely on correct bin assignment, this omission is load-bearing for the causal claim.

Authors: We agree that explicit validation details would strengthen the methods and support the causal claims. In the revised manuscript, we will expand the Methods section with a new subsection on z-phase computation. This will describe: extraction of nodule centers from LIDC-IDRI consensus annotations (≥4 readers, which inherently limits inter-observer variability); retrieval of reconstruction interval from DICOM metadata; the folding operation to [0, 0.5]; and internal validation steps including cross-checks against known slice positions and sensitivity tests on bin boundaries. These additions will directly address the load-bearing concern for bin assignment accuracy. revision: yes
Referee: [Methods] Methods: The 154-case perturbation study is not characterized as to whether z-phase shifts were generated by true re-reconstruction from raw projections or by post-hoc image-domain operations (slice shifting or interpolation). Only the former captures the partial-volume and aliasing physics that actually couple z-phase to detection; the latter would misrepresent the mechanism and undermine the interpretation that the effect appears only for d/D >= 1.0.

Authors: We appreciate this distinction and its implications for mechanism interpretation. The referenced 154-case perturbation study (from our prior validated work) generated z-phase shifts via post-hoc image-domain slice shifting and interpolation applied to the already-reconstructed LIDC-IDRI images, rather than full re-reconstruction from raw projections. We will explicitly document this in the revised Methods, discuss the resulting limitations (noting that post-hoc shifting approximates partial-volume effects but may not capture all raw-data aliasing), and qualify the d/D ≥ 1.0 interpretation accordingly. We will also add a forward-looking statement that raw-data re-reconstruction studies would be a valuable next step to confirm the physics. revision: yes

Circularity Check

0 steps flagged

No circularity in observational data stratification

full rationale

The paper performs a retrospective re-analysis and stratification of AI detection outcomes on the external LIDC-IDRI public dataset using results from a prior perturbation study. No equations, derivations, fitted parameters, or predictions are present that could reduce to the inputs by construction. Z-phase is computed directly as the fractional nodule-center offset within the reconstruction interval, and sensitivities are binned empirically by reconstruction interval and d/D ratio. The claims follow from the observed stratified counts without self-definition, ansatz smuggling, or load-bearing self-citation chains. The cited perturbation study supplies the input data but does not create circularity, as the current work adds an independent stratification layer grounded in the re-analyzed empirical results.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The analysis rests on standard domain assumptions about the representativeness of the LIDC-IDRI consensus nodules and the validity of the z-phase definition; no free parameters are fitted and no new entities are postulated.

axioms (2)

domain assumption LIDC-IDRI consensus nodules (>=4-reader agreement) provide a suitable sample for measuring AI detection sensitivity.
Used to select the nodules whose detection outcomes are stratified.
domain assumption Z-phase can be defined as the fractional position of the nodule center within the reconstruction cycle and folded to [0, 0.5].
Central definition enabling the binning and sensitivity comparison.

pith-pipeline@v0.9.0 · 5635 in / 1591 out tokens · 62202 ms · 2026-05-09T18:05:42.368265+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

8 extracted references · 1 canonical work pages

[1]

Detection of pulmonary nodules by multislice computed tomography: improved detection rate with reduced slice thickness.Eur Radiol

Fischbach F, Knollmann F, Griesshaber V, Freund T, Akkol E, Felix R. Detection of pulmonary nodules by multislice computed tomography: improved detection rate with reduced slice thickness.Eur Radiol. 2003;13(10):2378–2383

2003
[2]

Automated detection of pulmonary nodules on CT images: effect of section thickness and reconstruction interval—initial results.Radiology.2005;236(1):295–299

Kim JS, Kim JH, Cho G, Bae KT. Automated detection of pulmonary nodules on CT images: effect of section thickness and reconstruction interval—initial results.Radiology.2005;236(1):295–299

2005
[3]

Marten K, Grillhösl A, Seyfarth T, Obenauer S, Rummeny EJ, Engelke C. Computer-assisted detection of pulmonary nodules: evaluation of diagnostic performance using an expert knowledge-based detection system with variable reconstruction slice thickness settings.Eur Radiol.2005;15:203–212

2005
[4]

Effect of CT reconstruction settings on the performance of a deep learning based lung nodule CAD system.Eur J Radiol.2021;136:109526

Blazis SP, Dickerscheid DBM, Linsen PVM, Martins Jarnalo CO. Effect of CT reconstruction settings on the performance of a deep learning based lung nodule CAD system.Eur J Radiol.2021;136:109526

2021
[5]

Khan S, Noor MN, Ashraf I, Masud MI, Aman M. Impact of CT intensity and contrast variability on deep-learning-based lung-nodule detection: a systematic review of preprocessing and harmonization strategies (2020–2025).Diagnostics.2026

2020
[6]

Soliman D. Beyond benchmarks: acquisition condition sensitivity of AI-assisted pulmonary nodule detection — a 154-case LIDC-IDRI perturbation study.arXiv:2603.26785(under review,Academic Radiology). 2026

work page arXiv 2026
[7]

Artificial intelligence as an independent reader of risk-dominant lung nodules: influence of CT reconstruction parameters.Eur Radiol.2025

Mao Y, Heuvelmans MA, van Tuinen M, Yu D, Yi J, Oudkerk M, Ye Z, de Bock GH, Dorrius MD. Artificial intelligence as an independent reader of risk-dominant lung nodules: influence of CT reconstruction parameters.Eur Radiol.2025

2025
[8]

The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans.Med Phys.2011;38(2):915–931

Armato SG III, McLennan G, Bidaut L, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans.Med Phys.2011;38(2):915–931. 6

2011