Which Monoenergetic Energy Can See a Faint Lesion? A Detectability Ceiling for Spectral CT

Spectral CT scanners can produce "virtual monoenergetic images" (VMIs): pictures computed to look as if the patient had been imaged with X-rays of a single energy $E_0$ instead of the usual broad spectrum. Radiologists choose $E_0$ largely by habit. They use about 40 keV to make iodine contrast agent glow, because iodine absorbs strongly just above its 33 keV absorption edge, and about 70 keV for quiet, artifact-free images. Every VMI is the same weighted sum of two underlying "basis material" images, such as a water map and an iodine map, with weights that depend on $E_0$. Picking an energy therefore amounts to picking one direction in a two-dimensional space. For any detection task, statistics sets a best possible direction and a hard limit on how detectable the lesion can be. This proposal computes that limit, predicts which energy comes closest to it for a given task and dose, and tests whether the monoenergetic family can reach it at all.

The second half of the idea comes from rare-event physics. There, signals are scarce, counts follow Poisson statistics, and a real discovery must be separated from chance fluctuations that multiply when you search many locations. We carry two of those tools into CT. The first is the ideal likelihood-ratio detector as the reference ceiling. The second is an explicit correction for searching a whole organ rather than one known spot, which physicists call the "look-elsewhere" or trial-factor effect. The result is a calculator and a validation study. The calculator says which $E_0$ to reconstruct for a given task and dose, how much detectability any single VMI gives up relative to the ideal detector, and the photon count below which the whole Gaussian picture fails.

Core Claim

Suppose a lesion changes the basis coefficients by a vector $\Delta c$ (for example, a small iodine increase with little water change). Suppose the basis-map noise at that location has covariance $\Sigma$, whose large negative off-diagonal term is the well-known anti-correlated noise. A VMI at energy $E_0$ applies the weight vector $a(E_0) = (\mu_1(E_0), \mu_2(E_0))$, where $\mu_m(E_0)$ is the attenuation of basis material $m$ at that energy. For a lesion at a known location, the squared detectability index of that VMI is

$$d'^2(E_0) = \frac{(a(E_0)^\top \Delta c)^2}{a(E_0)^\top \Sigma\, a(E_0)} \;\le\; \Delta c^\top \Sigma^{-1} \Delta c = d'^2_{\text{ideal}},$$

where the bound follows from the Cauchy–Schwarz inequality. Equality holds only if some energy makes $a(E_0)$ parallel to $\Sigma^{-1}\Delta c$. The attainable directions form a limited arc as $E_0$ runs over roughly 40–140 keV, so the optimum may lie outside it. In that case no VMI is optimal and a fused two-channel reading is strictly better.

This gives three falsifiable predictions.

  1. Task dependence. The best energy depends on the direction of $\Delta c$. Detecting an iodine-enhancing lesion and telling a cyst from a hypodense metastasis point in different directions. This offers a quantitative reading of the unexplained split in Liver lesion localisation and classification with convolutional neural networks, which reported 40 keV as best for binary detection and localization but 70 keV as best for three-class classification.
  2. Dose invariance. Lowering tube current scales $\Sigma$ by a constant, and in the formula above that constant cancels. So in the Gaussian, linear regime the optimal energy should not depend on dose. Any shift in the optimal energy that we observe at low dose then measures a failure of that regime. The likely causes are Poisson counts too sparse for Gaussian approximations, the nonlinear log transform of near-zero counts, or vendor noise suppression.
  3. Search tasks. When the lesion must be found anywhere in an organ, the threshold for a fixed false-positive rate per scan grows with the number of independent noise "cells" searched. That number depends on noise texture, and texture differs across energies. The best energy for search can therefore differ from the best energy for a known location.

Relation to Prior Work

The closest statistical work optimizes hardware rather than image synthesis. High-Sensitivity Iodine Imaging by Combining Spectral CT Technologies uses the Fisher information matrix (a measure of how precisely data pin down parameters) to choose source spectra and K-edge filters. It raised a Material Separability Index from 2.54% to 12.52%. Design of Spatial-Spectral Filters for CT Material Decomposition applies the same kind of index ($s^2 = 0.174$) to tiled filter design. Both supply our method: noise propagation from spectra to basis-map covariance. Both stop at the basis maps. Neither asks which monoenergetic projection of those maps a reader or network should see, and neither accounts for search over a volume. The second paper explicitly leaves open the non-Gaussian breakdown at very low dose, which is our third regime.

Spectral CT reconstruction with an explicit photon-counting detector model: a one-step approach fits basis maps directly to Poisson counts through a detector response matrix. We reuse it as the low-count reference: it provides the Poisson likelihood against which the Gaussian ceiling is checked when counts are sparse. Its authors did not synthesize VMIs, and they evaluated only a cylindrical phantom.

Two lines of work motivate the proposal empirically. On the clinical side, the liver study above and Bildgebung des Pankreaskarzinoms mittels Dual-layer Spektral-CT show large contrast gains at 40 keV: 30.3 HU at 70 keV rising to 102.4 HU at 40 keV. Neither measures detectability against noise. On the synthesis side, Jointly Optimized Deep Neural Networks to Synthesize Monoenergetic Images from Single-Energy CT Angiography and Prediction of Low-Kev Monochromatic Images From Polyenergetic CT Scans generate 40 keV-like images from single-energy scans. Our framework constrains that work: a network given one polyenergetic channel cannot exceed that channel's own detectability, so any apparent gain must come from learned priors, not from new physical information.

From the physics literature we borrow two ideas. The first is joint modeling of all event attributes rather than sequential cuts, argued in Multi-variable maximum likelihood analysis and Event-level spatial-energy likelihoods. The second is the discipline of post-trial significance illustrated by the Telescope Array anisotropy analysis, where apparent hotspots did not survive correction for the number of locations searched.

Research Program

We will build an open calculator. Its inputs are a source spectrum, detector type (dual-layer or photon-counting with a response matrix), patient water-equivalent diameter, dose, and a task given as a lesion size and contrast vector. It first propagates Poisson noise to the basis-map covariance, including its spatial correlation. From that it returns the $d'(E_0)$ curve, the optimal energy, the ideal ceiling, and the fraction of the ceiling the best VMI reaches. For search tasks it also returns a threshold corrected for the number of independent search cells, estimated from the noise autocorrelation of each energy. Finally, it reports a breakdown count: the photon count per detector bin at which Monte Carlo draws from the Poisson projection likelihood depart from the Gaussian prediction by a chosen tolerance.

We then validate the calculator in three settings of increasing realism. First, simulated photon-counting phantoms run through the one-step Poisson model, where ground truth is exact. Second, a physical water/iodine phantom with 0–20 mg/mL inserts, the same design as the Tivnan test bench, scanned at several dose levels on a dual-layer scanner. Third, a retrospective dual-layer liver cohort assembled at our site, with lesions inserted at controlled contrast and noise added to emulate reduced dose. We cannot reuse the Shapira cohort directly, so it serves as an external reference point only.

Baselines

Naive reference. Conventional polyenergetic images, plus the 70 keV VMI that is routine in practice. In the liver study, conventional CT gave 85.4% binary accuracy and 10.58 mm localization error. Any energy-selection rule must beat this reference on our cohort, run through the same network and training protocol.

Fixed 40 keV rule. The clinical low-energy default. The cited liver study reports it as strongest for binary classification: 89.9% accuracy, AUC 0.960. We run it on identical data splits, with the same classifier architecture and the same hyperparameter search budget as every other arm. Beating it means one of two things: a predicted energy gives a higher AUC or lower localization error with a paired bootstrap interval excluding zero, or it matches 40 keV at a stated fraction of the dose.

Iodine map. The iodine map is itself a particular linear combination of the basis images. It gave the best localization in the liver study (8.22 mm). We treat it as one more point on the same weight-space diagram, so the theory must correctly predict when it wins.

Empirical sweep. The strongest competitor is an exhaustive sweep that trains or evaluates a model observer at every 10 keV and picks the winner. Our calculator must choose within the sweep's confidence interval of the best energy without looking at labels. It must also correctly predict how the winner moves when the task changes.

Ideal-direction ablation. This isolates the key ingredient. We compare the best VMI against a single image reconstructed with the ideal weights $\Sigma^{-1}\Delta c$, and against a network given both basis maps. If the ideal-weight image beats the best VMI by roughly the predicted gap, the "VMI family can fall short" claim holds.

Search-task reference. For search over a volume, the reference numbers are from the CAD-PE challenge: 75% per-embolus sensitivity at 2 false positives per scan. That dataset is single-energy, so it anchors how we score free-response detection rather than serving as a spectral comparison. For low-dose spectral search: no reported result found on Lacuna for that setting.

Why not just…?

Why not just sweep energies empirically and pick the best? A sweep answers one task at one dose on one cohort. It cannot say whether the winner is near the ceiling or how it will move for a different lesion type or patient size. It is still a real threat on its home turf, and we keep it as the strongest baseline. If the calculator's predicted energy does no better than the sweep, and its predicted gap to the ceiling fails to match the ablation, the contribution shrinks to an explanatory model. We would report exactly that.

Why not just give a network both basis maps and let it learn the best combination? Given enough data, a network should approach the two-channel ideal. That would confirm the ceiling rather than refute it, and it would make the "best VMI" question moot for automated reading. The experiment measures how close two-channel networks get at the cohort sizes typical in this field (172 patients in the liver study, 27 in the paired pulmonary-angiography set). Human readers still read single images, and the theory tells us when a single VMI loses little.

Why not just use the iodine map for every enhancing lesion? The iodine map removes the water component but inherits its full share of the anti-correlated noise. It is optimal only when $\Sigma^{-1}\Delta c$ happens to align with it. The liver study itself found it best for localization, yet 70 keV was best for three-class classification. The theory predicts which of the two wins for each contrast vector, and the phantom experiments test that prediction directly.

Proposed Experiments

Test 1: dose invariance. In the phantom, we measure $d'(E_0)$ with a channelized Hotelling observer, a standard linear model observer, at six dose levels spanning a factor of 16. Gaussian theory predicts an optimal energy that does not move. We will report the dose at which it starts to move, and whether that dose matches the breakdown count predicted by the Poisson Monte Carlo from the one-step model.

Test 2: task dependence. Among inserts and inserted liver lesions, we vary $\Delta c$ between pure iodine enhancement and mixed density-plus-iodine change. We then check whether the observed best energy tracks the predicted one across tasks. This includes whether the 40-versus-70 keV split reported for detection versus classification is reproduced.

Test 3: search versus known location. We compare the best energy for a known-location task with the best energy for free-response search, scored as sensitivity at 1 and 2 false positives per volume. This tests whether trial-factor correction by noise texture predicts the difference.

Test 4: the ceiling gap. Across all settings, we report the fraction of the ideal $d'^2$ reached by the best VMI, the iodine map, the ideal-weight image, and the two-basis network.

The proposal succeeds if the predicted gap is non-trivial for at least one clinically relevant task and the ablation confirms it. It fails if the best VMI always sits within noise of the ceiling. In that case current practice is statistically sound, and the result is still worth publishing.

References

  1. Liver lesion localisation and classification with convolutional neural networks: a comparison between conventional and spectral computed tomography
  2. High-Sensitivity Iodine Imaging by Combining Spectral CT Technologies
  3. Design of Spatial-Spectral Filters for CT Material Decomposition
  4. Spectral CT reconstruction with an explicit photon-counting detector model: a one-step approach
  5. Bildgebung des Pankreaskarzinoms mittels Dual-layer Spektral-CT
  6. Jointly Optimized Deep Neural Networks to Synthesize Monoenergetic Images from Single-Energy CT Angiography for Improving Classification of Pulmonary Embolism
  7. Prediction of low-keV monochromatic images from polyenergetic CT scans for improved automatic detection of pulmonary embolism
  8. Observation of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">B</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mo>→</mml:mo><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">ω</mml:mi><mml:mi mathvariant="italic">K</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow></mml:math>and Search for Related<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="italic">B</mml:mi></mml:math>Decays Modes
  9. Signal estimation in on/off measurements including event-by-event variables
  10. INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
  11. Computer Aided Detection for Pulmonary Embolism Challenge (CAD-PE)