Measuring Organ Drift Across Serial Screening CTs Before Clinical Markers Move

Many clinical warning signs only change after the body has run out of ways to compensate. Ejection fraction, spirometry ratios and serum creatinine stay normal while the underlying tissue is already being lost. Millions of people, however, already get repeated chest CT scans for lung-cancer screening. Each low-dose scan also captures the heart, aorta, lungs, spine and chest-wall muscle. This proposal treats the change in those structures between annual screens as an early warning signal. Automated volumetric segmentation, meaning software that outlines each 3D structure voxel by voxel so its volume and density can be measured, would supply the measurements. The main case is a person whose single scan looks normal but whose vertebral bone density, muscle mass, emphysema fraction or cardiac size is moving the wrong way.

The closest prior work reads one scan per person. It trains a 3D convolutional classifier on a single National Lung Screening Trial (NLST) scan to predict non-cancer death. It reaches an AUC of 0.73, above radiologists at 0.64 (Lu et al.). Its saliency maps point at ribs and vertebrae, which suggests frailty signals the model never measures explicitly. We ask a narrower and more testable question. Does the year-over-year rate of change of segmented, named structures predict non-cancer mortality beyond what any single scan contains? The test group that matters most is people whose single-scan values sit inside normal ranges.

Core Claim

Organ-level volumetric trajectories, taken from routine serial low-dose CT, carry prognostic information that static snapshots do not. This is the "pseudo-normal" compensatory window described across the lagging-indicator literature. Capturing that information requires one technical ingredient that current pipelines lack: segmentations that are consistent between timepoints. If each scan is segmented independently, the error in the measured change is roughly $\sqrt{2}$ times the error of a single measurement, because the two errors are independent and add in variance. For small or low-contrast structures, that noise can swamp a real 1–3% annual change. Small structures are exactly where current segmenters are weakest; adrenal glands and thin veins score around 0.56–0.64 Dice (He et al.; Ji et al.). We therefore segment each person's scans jointly, anchored by deformable registration. The claim is that this is what makes trajectory features usable.

Positioning Against Prior Work

Three studies motivate the idea. The COVID-19 volumetry studies showed that automated opacity volumes appear days before symptoms (Varble et al.) and follow standard temporal reference curves (Kassin et al.). They are limited in three ways: they studied acute infection, the cohorts were 29–74 patients, and one pipeline needed manual refinement. We move to chronic decline, cohorts of thousands, and a fully automatic pipeline. The third study, Feng et al., fused a CT lesion ratio with laboratory values and reached AUC 0.959 for inpatient deterioration. It is a late-stage predictor built on a single scan. We borrow its lesson that explicit volume ratios beat qualitative scores, and apply it to change over time.

Lu et al. also constrains the design. Adding clinical variables to the imaging model by late fusion did not help, so we do not expect bolting on tabular data to rescue a weak imaging signal. We therefore judge the imaging features by what they add over the imaging-only 3D-ResNet.

The segmentation literature supplies the method. On AMOS, a well-tuned plain 3D U-Net beat transformer backbones on CT (89.04% versus 86.32% mean Dice; Ji et al.), and nnU-Net-style pipelines won both KiTS challenges (Heller et al.; Heller et al.). For that reason we build on an nnU-Net segmenter trained on the TotalSegmentator CT data rather than a transformer. The file reports no benchmark number for the TotalSegmentator CT model itself.

Non-imaging work shows lagging markers can be bypassed without imaging. Examples are ICU time-series models for kidney injury (Rank et al.) and spirogram shape for COPD (Mei et al.). Those models need dense monitoring or dedicated tests. Our setting reuses scans that are already acquired.

Research Program

The first component is a longitudinally consistent segmenter. For each participant with two or three NLST screening rounds, we register later scans to the first. We then segment the first scan with nnU-Net. Masks for later scans are refined from the propagated mask by a network conditioned on both images. That network is trained with a consistency penalty: it should not change voxels whose intensity-normalized appearance has not changed.

The target structures are chosen for known links to cardiovascular, respiratory and frailty-related death, and for being measurable on low-dose non-contrast CT:

  • vertebral bodies, for bone density in Hounsfield units;
  • paraspinal and pectoral muscle, for area and density;
  • lung lobes, for volume and the fraction below −950 HU, a standard emphysema measure;
  • the cardiac silhouette and pericardial fat;
  • the thoracic aorta, for diameter and calcified volume;
  • coronary calcium.

The second component is a feature set. It holds per-structure values at baseline and their annualized rates of change. Each rate is standardized against a noise floor estimated for that structure, so the model sees change in units of measurement uncertainty.

The third component is the outcome model. We fit a regularized Cox model and a gradient-boosted classifier on these interpretable features. We also fit a small Siamese 3D-ResNet on the paired scans as a comparison.

The noise floor comes from two sources. The first is structures that should be stable over one year, such as vertebral body volume, as distinct from vertebral density. The second is paired reconstructions of the same acquisition with different kernels, where the reconstructions are available. A subset of about 100 scan pairs will be manually corrected to measure how much error each method leaves in the change estimate.

Baselines

Naive reference. A Cox or logistic model on age, sex and pack-years. In the matched design these variables are balanced by construction, so this model should sit near chance. Lu et al. report clinical-variable models peaking at AUC 0.58 on their matched NLST cohort. We will reproduce this reference on our cohort.

Strongest existing method. The single-scan 10-layer 3D-ResNet with Med3D initialization and convex-hull preprocessing from Lu et al., reported at AUC 0.73 on 498 matched participants. It is the strongest competitor because it already extracts hidden signal end-to-end from the same kind of scan. We will reimplement it on the most recent pre-outcome scan, with the same train/test splits, the same number of hyperparameter trials and comparable GPU hours as our pipeline. We will report it both on a Lu-style age/sex/pack-year matched sample and on the larger cohort. Beating it means a statistically significant AUC gain on the matched sample (paired DeLong test) and a higher C-index on the full cohort.

Static segmented features. The same structures measured on one scan only. This separates the value of segmentation from the value of time.

Ablation of the key ingredient. The identical trajectory model, but fed changes computed from independently segmented timepoints. If joint segmentation does not improve both change reliability and prognostic performance over this ablation, the technical contribution fails.

Clinical imaging score. A manual-style coronary calcium score. Lu et al. found calcium scoring underperformed their network.

Why not just…?

Why not just give the 3D-ResNet both scans and let it learn change end-to-end? This is the most serious threat, and we run it as a Siamese baseline. With a few thousand deaths, an end-to-end change model has to learn registration, anatomy and prognosis at once. It also yields only saliency maps as explanation, which gave Lu et al. unexpected rib and vertebra signals they could not quantify. If the Siamese network matches our trajectory features, the honest conclusion is that segmentation buys interpretability and a clinically reportable number such as "vertebral density falling 2 SD per year", not accuracy. We would report that result.

Why not just use the static volumes, which are cheaper? Static volumes need one scan and already capture much of the frailty signal. The pseudo-normal subgroup analysis below decides this. If trajectories add nothing among participants whose baseline values are all within reference ranges, the lagging-indicator framing is wrong for this setting.

Why not just use coronary calcium or emphysema percentage, which radiologists can already report? Single scalar indices are the 2D-caliper equivalent here. They are cheap, but Lu et al. found calcium scoring weaker than whole-volume learning. We include both, alone and combined, so any gain from multi-structure trajectories is measured against them directly.

Proposed Experiments and Decisive Checks

We use NLST participants with at least two screening rounds before death or end of follow-up. The outcomes are cardiovascular and non-cancer respiratory death, excluding lung-cancer deaths, following Lu et al. We report results both on a matched balanced design comparable to their 498-person cohort and on the full cohort with time-to-event analysis.

Segmentation quality. We measure Dice and, more importantly, the error in change estimates against the manually corrected pairs, for joint versus independent segmentation.

Prognostic performance. We report AUC, C-index, calibration, and net reclassification over the Lu et al. 3D-ResNet.

Pseudo-normal subgroup. This is the check that decides the proposal. We restrict to participants whose baseline structure values all lie inside cohort reference ranges and ask whether trajectory features still separate future deaths from survivors.

Lead time. We stratify by years between the last scan and death. This tests whether the signal persists for deaths three or more years out, which would indicate an early rather than terminal signal.

Failure analysis. We audit performance by scanner manufacturer and reconstruction kernel, given the domain-shift sensitivity documented on AMOS, and by demographic subgroup, following the disparities KiTS21 exposed (Heller et al.).

The proposal succeeds if three things hold together. Joint segmentation must lower change error relative to independent segmentation. Trajectory features must beat both static features and the 3D-ResNet. And the gain must survive inside the pseudo-normal subgroup.