In roughly half of acute ischemic strokes, the emergency non-contrast CT scan looks normal to expert radiologists, even when an MRI taken minutes later shows the infarct clearly (APIS). The early tissue damage changes CT density by only a few Hounsfield units, about the same size as the scanner's own noise. Deep segmentation networks do poorly here: the best of 41 teams on the APIS paired CT–MRI challenge reached a Dice overlap of only 0.20. This proposal treats that failure as a statistical detection problem rather than an architecture problem. A faint lesion is a patch of voxels that is slightly darker than its mirror-image patch in the other hemisphere. The question is whether that patch is darker than the noise could plausibly produce. Spatial statistics has a tool for exactly this question, the Getis-Ord $G_i^*$ hotspot statistic, which tests whether a neighborhood sum is too extreme to arise by chance. Its textbook form, however, assumes neighboring values are independent. CT noise is not independent: each reconstruction kernel smears noise into blobs whose size is specific to the scanner and protocol. A hotspot test with an independence null is therefore miscalibrated by a different amount on every scanner.
The core claim is that the useful thing spatial autocorrelation analysis offers this problem is a correct null model, not a new pooling layer or a texture feature. For each scan, we estimate how the noise is correlated in space by fitting a semi-variogram, a curve showing how fast the difference between two voxels grows with their distance, on tissue known to be healthy. We then compute hemispheric-difference hotspot z-scores whose variance accounts for that correlation, at several neighborhood sizes. The resulting maps should mean the same thing on every scanner, so "z > 3" carries the same evidential weight everywhere. We test whether these calibrated maps, used alone as a training-free detector and as an extra input channel to nnU-Net (the standard self-configuring 3D segmentation network), improve detection of small and visually invisible lesions beyond what plain smoothing or uncalibrated statistics achieve.
Core Claim
Let $d_i$ be the mirror-difference value at voxel $i$: the intensity at $i$ minus the intensity at its contralateral homologue, after symmetric registration and subtraction of the healthy-population mean asymmetry. Let $w$ be a neighborhood weight vector at scale $s$. The calibrated statistic is
$$z_i^{(s)} = \frac{\sum_j w_{ij}\, d_j}{\sqrt{w_i^\top \Sigma\, w_i}},$$
where $\Sigma$ is the covariance of $d$ under the no-lesion hypothesis. $\Sigma$ is built from the scan's own fitted variogram, which captures kernel-induced noise correlation, plus a population term for normal anatomical asymmetry estimated from control scans. Under the independence assumption of classical $G_i^*$, the denominator collapses to $\sigma\sqrt{\sum_j w_{ij}^2}$. When noise is positively correlated, that smaller denominator inflates z-scores by a factor that grows with neighborhood size and with kernel sharpness. The prediction is that this inflation, which varies from scan to scan, is a main source of false positives and of cross-scanner instability for faint lesions. Correcting it should matter most in the small-lesion regime, where detection thresholds are tight. The numerator is just a weighted local mean, and a skeptic is right to call it a box filter. What is new is the scan-specific denominator and the multi-scale significance it allows, including control of the false discovery rate (the expected share of flagged clusters that are false alarms) over clusters.
Relation to Prior Work
Two lines of work motivate the gap. APIS (Gómez et al.) established that half of its acute cases are invisible on CT and that networks collapse below 1.19 mL. ISLES'24 (dataset, challenge) found that even with CT angiography and perfusion, lesion-wise F1 falls to 14.94% below 5 mL. Neither treats faint signal detection as a hypothesis test with a calibrated null.
Two works supply the mirror idea we build on. The APIS winner used inter-hemispheric difference images inside nnU-Net. CNN-R (Marcus et al.) compares lesions against contralateral tissue to estimate lesion age, but it relies on human-adjudicated segmentations. We keep the mirror difference and change how its significance is judged.
The closest methodological precedent is Getis-Ord Gi* pooling. It replaces max pooling in a 2D network with a $G_i^*$-gated rule and improves cross-city transfer on overhead imagery, from 15.48% to 23.02% mIoU for Vaihingen→Potsdam. It uses the standard independence null inside a learned feature space. We instead apply the statistic to the physical image, with a null derived from scanner physics, and we port its pooling layer to 3D as a competing baseline.
Radiomics studies on admission CT, for malignant edema (AUC 0.818 radiomics alone) and hemorrhagic transformation (AUC 0.90), show that sub-visual texture carries signal. They require manual regions of interest, however, which our map-based approach avoids. These studies also constrain the design: both found that clinical variables add little or nothing, which agrees with ISLES'24. We therefore do not rely on tabular data.
The epidemiological uses of hotspot statistics (Beijing TB, Queensland dengue) contribute the method and its known weak point. Fixed, a priori weights and independence nulls are exactly what we replace.
Research Program
The pipeline has four parts.
- Registration and difference map. Each non-contrast CT is registered symmetrically to a template, the mid-sagittal plane is estimated, and the mirror-difference map $d$ is computed.
- Healthy-tissue mask. Variogram fitting needs tissue presumed healthy. We use a conservative mask that excludes the top tail of raw asymmetry and the vascular territories on the symptomatic side when that side is known. We check that the fitted noise parameters do not shift when lesion tissue is deliberately left in.
- Noise model. We fit an isotropic exponential-plus-nugget variogram to the noise per scan. CT noise correlation is roughly stationary within a slab, and anisotropy along the slice axis gets its own range parameter.
- Fast computation. Calibrated $z^{(s)}$ is computed at three to four neighborhood radii, spanning roughly 0.1 to 5 mL. Local sums are computed with fast Fourier transforms, which keeps the full-volume cost at $O(N\log N)$ and addresses the voxel-scale compute concern for brain volumes.
The maps are then used two ways. First, as a training-free detector: clusters surviving false discovery rate control across scales become lesion candidates. Second, as extra input channels to nnU-Net.
A secondary hypothesis tests local spatial dependency directly. Cytotoxic edema may make tissue more homogeneous, which would show up as a change in local variogram range between homologous neighborhoods even before mean density shifts. This variance-structure map enters only as an ablation channel. If it adds nothing, we drop it.
Baselines
The naive reference point thresholds the raw mirror-difference map at a fixed HU value tuned on training folds. It shows how much signal plain asymmetry carries with no spatial reasoning. For patient-level detection, the do-nothing point is calling every scan normal.
The strongest existing method on APIS is the challenge winner, an nnU-Net with an inter-hemispheric difference input, which reports Dice 0.20 on the 36-patient test set (APIS). We reimplement it within the same nnU-Net version, with identical preprocessing, registration, folds, epochs and GPU budget, and add our channels only as extra inputs. Any gain is then attributable to the maps.
On ISLES'24, the reported winner (Kurtlab, Dice 0.285, with small-lesion lesion-wise F1 of 14.94%) used all modalities (ISLES'24). Our experiments are restricted to non-contrast CT, and we found no reported result on Lacuna for a non-contrast-CT-only ISLES'24 model. We will therefore train our own nnU-Net baseline on that channel and report the full-modality numbers only as an upper reference.
A 3D port of Getis-Ord Gi* pooling in the same nnU-Net serves as the "statistic inside the network" competitor.
Two ablations isolate the key ingredient. The first feeds Gaussian-smoothed difference maps at the same scales; this matches the numerator and drops the statistics. The second feeds uncalibrated $G_i^*$ maps with an independence null. Each gets the same tuning budget, a small grid over scales and thresholds selected on validation folds only.
Beating the strongest baseline means four things on APIS:
- higher lesion-wise F1 in the below-1.19 mL bin and on the radiologist-labeled invisible subset;
- higher patient-level detection AUC, using the 10 controls and the invisible cases;
- a confidence interval that excludes the smoothed-difference ablation, not merely the vanilla model;
- volumetric Dice that is no worse.
Why not just…?
Why not just smooth the difference image and let thresholding do the rest? This is the most serious threat, because the numerator of our statistic is a smoothed difference. Smoothing gives the same ranking of voxels within a single scan. It cannot give a threshold that means the same thing across scans whose noise correlation differs. The smoothed-difference ablation tests this directly. If it matches the calibrated maps on pooled-scanner lesion F1 and patient-level AUC, the calibration claim is false and we will report that.
Why not just let nnU-Net learn the filter from the difference image? Convolutions can learn local averaging. Learning the per-scan noise covariance, though, requires many examples of each reconstruction kernel, and APIS offers well under a hundred training scans. The winner's Dice of 0.20 suggests the network is not discovering this structure on its own. The channel-addition design tests it: equal compute, one extra input.
Why not just use radiomics texture features, which already work for edema and hemorrhage prediction? Those studies (edema, hemorrhagic transformation) summarize a lesion that a radiologist has already outlined, so they answer "what will this lesion do," not "where is it." Their manual regions of interest are precisely what is unavailable when the lesion is invisible. Texture descriptors also carry no calibrated notion of significance.
Why not just port Gi* pooling to 3D? We do, as a baseline. Its independence null is applied to learned activations, whose correlation structure is set by training rather than by physics. We expect it to help mainly with noisy speckle output, not sub-threshold detection. Its 2D result was on domain transfer, not faint-target detection.
Proposed Experiments and Decisive Checks
The primary benchmark is APIS. Its test set is small (36 patients) and its ground truth is noisy (inter-rater Dice of 0.63 between the two MRI readers), so we pool five-fold cross-validation over all 96 studies with the official test split. We report patient-bootstrap confidence intervals and treat 0.63 as a practical ceiling on Dice. ISLES'24 non-contrast CT supplies a second, multi-vendor cohort for size-binned lesion F1, with the below-5 mL bin as the target.
The first decisive check is calibration itself. On APIS controls and on the healthy hemisphere of ISLES'24 scans, the empirical false-cluster rate at nominal z thresholds should match the nominal rate for calibrated maps. For uncalibrated maps it should be inflated, and more so for sharper kernels. If calibrated maps are not closer to nominal, the mechanism fails before any segmentation result matters.
The second check is the ablation ordering on the invisible subset and small-lesion bins: calibrated above uncalibrated $G_i^*$ above smoothed difference above the APIS winner's configuration.
The third check is cross-scanner transfer. We train on Munich and test on Zurich within ISLES'24, where scanner-dependent calibration should matter most.
A negative but informative outcome would be calibrated statistics improving the training-free detector's false discovery control while adding nothing inside nnU-Net. That would indicate the network does absorb calibration once it has enough data, and the contribution would narrow to a training-free triage flag.
References
- APIS: a paired CT-MRI dataset for ischemic stroke segmentation - methods and challenges
- ISLES'24 -- A Real-World Longitudinal Multimodal Stroke Dataset
- ISLES'24: Improving final infarct prediction in ischemic stroke using multimodal imaging and clinical data
- Deep learning biomarker of chronometric and biological ischemic stroke lesion age from unenhanced CT
- Generalizing Deep Models for Overhead Image Segmentation Through Getis-Ord Gi* Pooling
- A CT-based machine learning model for using clinical-radiomics to predict malignant cerebral edema after stroke: a two-center study
- The value of CT-based radiomics in predicting hemorrhagic transformation in acute ischemic stroke patients without recanalization therapy
- Spatial Analysis of Tuberculosis Cases in Migrants and Permanent Residents, Beijing, 2000–2006
- Spatial and Temporal Patterns of Locally-Acquired Dengue Transmission in Northern Queensland, Australia, 1993–2012