Finding Silent Procedural Micro-Infarcts by Comparing Each Patient Against Their Own Pre-Procedure Scan

Catheter procedures in the brain's arteries, including diagnostic angiography, carotid stenting and clot retrieval, often knock loose tiny fragments of plaque or clot. These fragments lodge downstream and cause small strokes that produce no obvious symptoms. On diffusion MRI, the scan most sensitive to fresh stroke, such lesions turn up in roughly 28–40% of patients after a procedure. Most of them are only a few voxels in size, and on the thick slices routinely used in hospitals they are blurred into the surrounding tissue. Automated segmentation models trained on single scans reliably miss them. Yet the patients most at risk usually had an MRI before the procedure. That earlier scan shows the same brain, with the same anatomy, the same old scars and the same scanner artifacts, a few days earlier. We propose to treat the pre-procedure scan as a patient-specific "nothing new here" reference. Micro-lesion detection then becomes a change-detection problem: the question is what is new in this brain since the reference, not what looks abnormal compared with other brains.

The central bet is that this change of question moves the sensitivity limit. A single-scan detector must separate a faint, partially averaged lesion from everything else that varies between people: vessels, cerebrospinal fluid spaces, old gliosis, susceptibility artifacts near the skull base. A paired detector only has to separate it from what varies between two scans of the same person, which is acquisition noise plus registration error. We add one procedural input. The procedure log records which vessels were catheterized, and that tells the model where emboli are likely to land, including the contralateral hemisphere, because the catheter scrapes debris off the aortic arch on its way in. Together these inputs should recover lesions below the size at which single-scan models fail. They should also sharply reduce "temporal hallucinations", where a pre-existing lesion is reported as new.

Core Claim

A 3D detector that takes a registered pre-procedure scan, the post-procedure diffusion scan, and a procedure-derived vascular-territory prior can detect new punctate infarcts at clinical slice thickness (3–5 mm) substantially better than the strongest single-scan ensemble applied to both scans and differenced. The gain should appear specifically in the smallest lesion bin (under about 10 voxels) and in the rate of false "new" calls on old lesions.

The reason can be written as a rough detectability ratio. A lesion of true volume $v$ inside voxels of volume $V$ carries only a fraction $v/V$ of its intrinsic contrast $\Delta s$; this is the partial volume effect. For a single-scan detector, detectability is roughly

$$d'_{\text{single}} \approx \frac{(v/V)\,\Delta s}{\sqrt{\sigma_{\text{anat}}^2 + \sigma_n^2}}, \qquad d'_{\text{paired}} \approx \frac{(v/V)\,\Delta s}{\sqrt{2\sigma_n^2 + \sigma_{\text{reg}}^2}},$$

where $\sigma_{\text{anat}}$ is lesion-like variation across tissue and patients, $\sigma_n$ is per-scan acquisition noise, and $\sigma_{\text{reg}}$ is the residual intensity mismatch after aligning the two scans. The paired approach helps exactly when anatomical and artifact clutter dominates noise and registration is good. It helps least when the two scans come from different scanners or protocols. Every term in this relation can be measured, so it gives a testable account of when the method should and should not work, not a guarantee.

Where This Sits Relative to Prior Work

The clinical motivation comes from two imaging studies. Schinz et al. scanned 82 patients with 2 mm isotropic diffusion MRI after diagnostic angiography and found new lesions in 28% of them. Lesion count rose with the number of vessels probed, with longer procedures, and with less experienced examiners. About 20% of the lesions persisted as permanent scars at five months. All reading was manual, and the study argues that routine 5–6 mm slices miss such lesions. Kim et al. found new lesions after roughly 40% of protected carotid stenting procedures, regardless of which distal protection device was used. About one third of those lesions were in the hemisphere opposite the treated artery, which points to debris scraped loose during catheter passage through the aortic arch. Together these papers establish that the lesions are common, small, and spatially tied to catheter route. That is why a route-conditioned spatial prior is plausible. Neither paper builds an automated detector.

The methodological constraints come from small-lesion segmentation. All Sizes Matter showed that a per-lesion "blob loss" plus a within-exam subtraction channel raises lesion-wise Dice on brain metastases from 0.43 to 0.49. The same paper found that a specialist model trained only on small lesions made the ensemble worse. We therefore reuse the instance-level loss and train on lesions of all sizes rather than building a small-lesion specialist. Their subtraction is between two contrasts acquired in one session. Ours is between two encounters days apart, which introduces registration error and real anatomical change, and the single-session subtraction does not address either. ISLES 2022 supplies real punctiform embolic patterns, with up to 126 separate lesions per patient, and identifies them as the main failure mode of current segmenters. The robust ISLES ensemble is the strongest single-scan system in the file.

The longitudinal reasoning work supplies the framing but not the architecture. BioViL-T, DeltaNet and the image-difference-graph model show that conditioning on a prior study reduces errors such as calling a stable finding new. Rajpurkar et al. name this failure "temporal hallucination". All four works concern 2D chest radiographs and whole-image findings, not 3D lesions a few voxels in size. CRONOS models 3D longitudinal series in continuous time, and it warns that simply copying the last image forward is a hard baseline to beat. However, it targets large-scale deformation and scores image fidelity, not detection of small lesions. We borrow CRONOS's lesson that a naive "copy the prior" reference must be the default comparison. We do not try to generate future images: the task is detection, and the file's evidence (CRONOS, §1.3.3) is that good image-fidelity scores do not guarantee that clinically useful findings are recovered.

Research Program

The model is a 3D encoder that reads the post-procedure diffusion scan and its companion ADC map, which are standard diffusion-MRI companions used to confirm that a bright spot is true acute injury. It also reads the pre-procedure scan, aligned to the post-procedure scan by rigid then deformable registration. The two scans are processed by shared-weight branches, and features are exchanged by cross-attention between corresponding regions. This follows the regional alignment idea of CoCa-CXR (Chen et al.), carried into 3D.

The procedure log is converted into a soft spatial prior. The catheterized vessels are mapped to their downstream vascular territories, with an arch-transit component that also covers the opposite hemisphere. This prior enters as an extra input channel, not as a hard mask, so lesions outside the expected territories remain detectable. The output is a set of instances, each with a "new since reference" probability. Training uses the blob loss together with an explicit penalty for flagging any component that is also present in the reference scan.

Training data comes in two tiers, because no public pre/post procedural diffusion dataset exists in the file. The first tier is a controlled synthetic benchmark built from ISLES 2022. For each case we simulate a second acquisition by re-sampling noise, applying a small head pose change, and adding a mild deformation. Into the "post" copy we paste real punctiform lesion patches taken from ISLES 2022 embolic cases. Existing lesions are left in both copies as distractors. Finally we render both scans at 2, 3 and 5 mm slice thickness using a partial-volume forward model. This gives exact ground truth at sizes below one voxel, which the human-read datasets cannot provide.

The second tier is a retrospective clinical cohort that we would assemble. It would include patients who had MRI before and diffusion MRI after angiography, stenting or thrombectomy, with the procedure logs, and ideally a subset scanned with thin-slice protocols like Schinz's. In that subset the thin-slice reads serve as reference labels while the model sees only the thick-slice reconstruction. Lesion fate at follow-up (scar or resolved) is a secondary, exploratory question. The only prior evidence here is 12 completers in Schinz et al., and we do not expect to power a fate model.

Baselines

The naive reference point is the "copy the prior" detector. It runs a strong single-scan segmenter on the post-procedure scan alone and reports every detected lesion as new. This is what current automated triage effectively does. It shows how many false new calls come from old lesions and artifacts.

The strongest existing competitor is detect-then-difference with the robust ISLES ensemble. That ensemble uses majority voting over nnU-Net, SegResNet and Factorizer and reports a median Dice of 0.82 and a lesion-wise F1 of 0.86 on 1,686 external Johns Hopkins scans. We run it on both scans, match components after registration, and call unmatched post-procedure components new. It is the strongest competitor because it already generalizes across scanners and uses the reference scan, just not jointly. For fairness it is retrained on exactly our training cases, with the same augmentations, the same compute budget and an equal number of hyperparameter trials, and its matching threshold is tuned on our validation split. The file reports no result for this ensemble on lesions under 10 voxels or at 5 mm slices: no reported result found on Lacuna for small-lesion recall at clinical slice thickness. Beating it therefore means a statistically significant gain in lesion-wise recall in the smallest size bin at matched false positives per patient, plus a lower rate of false new calls on old lesions. Overall F1 alone does not count.

The third baseline is the All Sizes Matter recipe with its subtraction channel replaced by the registered post-minus-pre difference. That paper reports lesion-wise Dice of 0.49 and 0.57 false positives per patient on BraTS metastases, a different task. This baseline asks whether a simple difference image, fed to a well-tuned single network, captures most of the gain.

Two ablations isolate our ingredients. The first removes the reference scan while keeping the procedure prior. The second removes the procedure prior while keeping the reference. For the prior, we separately report performance on lesions outside the predicted territories to check that the prior does not hide unexpected emboli.

Why not just…?

Why not just run the best single-scan segmenter on both scans and subtract? This is the real threat. If the ensemble already finds most punctate lesions, joint modelling adds little. Our argument is that a lesion the single-scan model misses on the post scan cannot be recovered by comparing lists of detections. The faint lesion only stands out against the patient's own reference. The synthetic benchmark settles this directly, because we control lesion size and slice thickness and can plot recall against size for both approaches. If the curves coincide, the contribution shrinks to the benchmark and the procedure prior, and we would report it as that.

Why not just acquire thin-slice diffusion MRI? Schinz et al. show that 2 mm isotropic imaging finds more lesions, and where it is available it should be used. But routine hospital protocols remain at thick slices, and even 2 mm voxels partially average lesions of 50–500 µm. The degradation experiment below measures how much of the thin-slice yield our method recovers from thick-slice data. This makes it a complement to better acquisition rather than a rival.

Why not just super-resolve the post-procedure scan first? Learned upsampling such as LoFNO works on smooth flow fields trained on simulations. A lesion averaged below the noise floor carries little recoverable information for an upsampler to exploit. Worse, an upsampler trained on typical brains is likely to smooth away or invent point-like features, as the fidelity-versus-utility gap noted in CRONOS warns. We include "super-resolve then segment" as an extra arm on the synthetic benchmark, where invented lesions can be counted exactly.

Why not just use a small-lesion specialist? All Sizes Matter already tested this and found that the specialist degraded the ensemble. We therefore do not pursue it.

Proposed Experiments

The first decisive check is the synthetic size-by-thickness sweep. For each method we plot lesion-wise recall at fixed false positives per patient, across lesion size bins and slice thicknesses of 2, 3 and 5 mm. We also report the rate of false new calls on the pasted-in pre-existing lesions. We measure $\sigma_n$, $\sigma_{\text{anat}}$ and $\sigma_{\text{reg}}$ on the simulated pairs and check whether the gain tracks the detectability ratio. We deliberately include pairs with worse registration and cross-protocol pairs, to find where the paired advantage disappears.

The second check is resolution degradation on the clinical thin-slice subset. We reconstruct the thin acquisitions at 5 mm and ask what fraction of thin-slice-confirmed lesions each method recovers from the thick data. We report free-response curves (sensitivity against false positives per patient) and patient-level area under the curve for "any new lesion".

The third check is the territory-prior ablation, stratified into lesions ipsilateral to the catheterized vessels, contralateral (the arch-scraping pattern described by Kim et al.), and outside any predicted territory.

The approach counts as a success only if it beats the retrained ISLES ensemble on small-lesion recall and old-lesion false calls in both the synthetic sweep and the thin-slice subset. If it wins only on synthetic data, that points to a simulation gap, and we would report it as such.