AI-Powered CT Perfusion Analysis for Stroke

In acute ischemic stroke, every minute counts. An arterial occlusion creates a core of irreversibly damaged tissue surrounded by hypoperfused but salvageable penumbra, and treatment decisions—particularly whether a patient qualifies for endovascular thrombectomy in extended time windows—hinge on distinguishing the two. CT perfusion imaging is the workhorse for making that call: serial CT images acquired during contrast passage yield 4D datasets from which maps of cerebral blood flow, blood volume, and transit time quantify the tissue at stake. But traditional CT perfusion analysis has real operational problems: it requires manual selection of arterial input and venous output functions, it struggles with patient motion during the one-minute acquisition, and it applies population-based tissue thresholds that do not generalize across scanners, protocols, and individual patients. The result is inter-operator variability, workflow delays in exactly the situations where delay costs brain tissue—and even the commercial packages themselves show significant disagreement in the core and penumbra volumes they report from identical data.

The client needed an automated system covering the complete workflow—from raw DICOM data to clinical tissue classification—that removes the manual steps, tolerates real-world image quality, and holds up against the commercial reference software already trusted in landmark thrombectomy trials. The accuracy bar was explicit and clinical: correlation above 0.90 for ischemic core volume and above 0.85 for perfusion parameters, because those volumes feed directly into patient-selection criteria for treatment. And it had to be fast enough to fit an emergency stroke workup, where an expert currently spends 5–10 minutes per case on analysis alone.

We built a seven-stage pipeline that combines deep learning where data is rich with classical optimization where geometry is well-defined:

The design principle throughout: preserve explicit, inspectable perfusion maps at every stage. Clinicians can verify intermediate results instead of trusting a black box, which matters for both adoption and regulatory review.

We validated on two complementary fronts, because neither alone is sufficient. Digital phantoms provide absolute ground truth: simulated CT perfusion acquisitions with programmed blood flow, blood volume, and transit time distributions, including realistic noise, partial volume effects, and contrast kinetics. Processing phantoms through the full pipeline measures quantitative accuracy directly, with no confounding from clinical variability. Acceptance criteria were defined up front: correlation above 0.85 for all primary parameters, systematic bias below ten percent of mean values, and confidence intervals reported rather than point estimates alone.

Clinical regression testing provides real-world concordance: the same source data processed through both our pipeline and a widely adopted commercial package with extensive validation in landmark thrombectomy trials. The regression dataset spanned multiple institutions, all three major scanner manufacturers (GE, Siemens, Philips), varied acquisition protocols including 5 mm and 10 mm slice thicknesses, and stroke severities from transient ischemic attack to large vessel occlusion.

Both validation modes run continuously, not once. The framework integrates with experiment tracking infrastructure so every pipeline modification triggers automated regression testing against historical baselines, with statistical process control flagging performance drift. Each test execution records the exact code revision, dependencies, and configuration—the reproducibility discipline that regulatory submissions and clinical deployments demand. Cases with quality warnings (severe motion, contrast timing abnormalities, truncated acquisitions) are flagged and analyzed separately, and every case with poor correlation against the reference standard was manually reviewed to characterize why—so aggregate metrics never hide a failure mode, and the system's limits are documented as carefully as its strengths.

The pipeline met or exceeded every clinical performance target:

The phantom results confirmed minimal systematic bias—mean errors within a few percent—so the pipeline neither over- nor under-estimates in a way that would skew clinical thresholds. Core volume correlation above 0.90 means reliable identification of the irreversibly injured tissue that drives patient selection, and penumbra correlation above 0.85 supports the mismatch-ratio calculations used in extended-time-window thrombectomy decisions. Mismatch volume (penumbra minus core) also showed strong correlation, and Bland-Altman analysis found narrow limits of agreement relative to typical core volume ranges—narrow enough to reliably classify patients against volumetric eligibility criteria. Sensitivity for detecting core presence was high while maintaining the specificity needed to avoid unnecessary interventions.

Complete processing takes approximately 3 minutes per case on standard infrastructure, versus 5–10 minutes of expert time for manual analysis—and because it runs unattended, it eliminates the workflow bottleneck entirely rather than just shortening it. Failure modes affecting a minority of cases (extreme motion, tandem occlusions, major pre-existing anatomical asymmetry) are characterized and flagged for expert review rather than silently mishandled.

The centerpiece is how the threshold-refinement network was designed around the clinical decision, not the segmentation metric. Thrombectomy eligibility depends on ischemic core volume—criteria like core below 70 mL—so a model can score respectably on pixel-wise overlap while systematically over- or under-estimating volume, and that error lands exactly where it changes a treatment decision. Our multi-output U-Net predicts tissue below four CBF thresholds simultaneously (38%, 34%, 30%, and 20% of contralateral normal, spanning mild hypoperfusion to definitive core) and trains with a hybrid loss: pixel-wise binary cross-entropy plus a volumetric MSE term that directly penalizes volume error in milliliters. The network sees six input channels—CBF, CBV, MTT, Tmax, temporal maximum intensity projection, and the baseline fixed-threshold mask—so it learns to refine a physically grounded starting point rather than hallucinate one. Ablations proved both choices: CBF alone as input reduced accuracy, and dropping the volumetric term kept acceptable pixel-wise IoU but degraded exactly the volume correlation that clinical eligibility rests on.

Two other decisions carried the project. First, hybrid architecture: learned models where labeled data existed (vessel segmentation, threshold refinement), classical optimization where the problem was geometric (CMA-ES for symmetry), and unsupervised learning where labels were impossible (motion correction via photometric consistency). Each tool was matched to its data regime instead of forcing deep learning everywhere—and because the pipeline preserves explicit perfusion maps as its intermediate representation, it can fall back to standard thresholding whenever learned refinements are uncertain, and clinicians can always inspect the reasoning behind a classification. Second, validation as infrastructure rather than a milestone: phantom ground truth plus commercial-software regression, wired into continuous integration, meant clinical accuracy was demonstrated—and protected—on every change, which is what turns a promising model into a deployable medical system.