19. Acute Stroke Imaging

Noncontrast computed tomography (NCCT) remains the indispensable first image in suspected acute stroke not because it is sensitive to early ischemia — it is not — but because it is exquisitely and rapidly specific for the two findings that must be excluded before any reperfusion therapy: intracranial hemorrhage and a large established infarct. Its physiologic content is governed by a single, almost embarrassingly simple relationship. Brain attenuation tracks tissue water. Normal grey matter measures approximately $37\text{--}45$ HU and white matter approximately $30\text{--}37$ HU, the difference reflecting the higher lipid (myelin) content and lower water fraction of white matter. When an arterial occlusion drops cerebral blood flow below the threshold for membrane ion-pump failure (roughly $10\text{--}15\ \mathrm{mL}\,100\mathrm{g}^{-1}\mathrm{min}^{-1}$), $\mathrm{Na^+/K^+}$-ATPase fails, sodium and water move intracellularly, and cytotoxic edema begins within minutes. Net tissue water rises, and because the effective linear attenuation coefficient of the parenchyma is a volume-weighted average of its components, attenuation falls. The classic empirical calibration is that attenuation decreases by roughly $2.6$ HU for every $1\%$ increase in tissue water content; net water in infarcting cortex rises on the order of $1\text{--}2\%$ in the first hours, predicting an early attenuation drop of only a few Hounsfield units. This is the crux of the detection problem: the signal change is of the same order as image noise.\n\nThe consequence is that windowing is not a cosmetic preference but a diagnostic instrument. Standard brain windows (width $\approx 80$, level $\approx 40$) span $0\text{--}80$ HU and compress the critical $30\text{--}40$ HU grey–white band into a handful of grey levels the eye cannot resolve. Deliberately narrowed “stroke windows” (width $\approx 30\text{--}40$, level $\approx 35\text{--}40$) expand contrast across exactly the range where infarcting tissue separates from normal, and reproducibly improve detection of early hypoattenuation. The interactive below is a true-Hounsfield head CT; toggling from the Brain to the Stroke preset demonstrates the principle directly.\n\nThe systematic search pattern is grayscale-symmetry-driven and comparison-based: read the two hemispheres side by side at matched levels, interrogating the deep grey nuclei (caudate, lentiform), the insular ribbon, the cortical mantle and sulci, and the ventricular margins for any asymmetric loss of grey–white distinction or effacement. The Bayesian posture is essential. A normal NCCT in a patient with a severe deficit does not lower the probability of stroke — sensitivity of NCCT for ischemia within three hours is only $\sim 25\text{--}60\%$ — it merely excludes hemorrhage and large completed infarct, which is precisely what is required to proceed. Conversely, NCCT is highly specific: when unequivocal hypoattenuation is present, an irreversible component almost certainly exists. The failure modes are characteristic. Beam-hardening and the inter-petrous Hounsfield artifact obscure the posterior fossa and inferior frontal/temporal cortex, hiding both ischemia and small hemorrhage; patient motion blurs the grey–white interface and can fabricate or efface asymmetry; volume averaging across an angled gantry mimics insular ribbon loss. Cognitively, the dominant error is satisfaction of search — finding a dense MCA and stopping before the contralateral subdural or the brainstem is read — and premature closure on a normal-appearing scan when the real lesion is an inconspicuous early hypodensity that stroke windows would have revealed.

The Alberta Stroke Program Early CT Score (ASPECTS), introduced by Barber and colleagues in the Lancet in 2000, converts the inherently subjective recognition of early ischemic change into a reproducible ordinal metric, and in doing so created the quantitative vocabulary in which modern stroke trials are written. The middle cerebral artery territory is partitioned into ten standardized regions assessed on two representative axial levels: at the level of the basal ganglia and thalamus, the caudate (C), lentiform nucleus (L), internal capsule (IC), insular ribbon (I), and the cortical regions M1 (anterior), M2 (lateral to insula), and M3 (posterior); and at the level just above the basal ganglia, M4, M5, and M6 representing the anterior, lateral, and posterior cortex respectively. A normal scan scores 10. One point is subtracted for each region demonstrating early ischemic hypoattenuation or focal swelling, so the score is a $0\text{--}10$ integer that is, by construction, inversely proportional to the volume of visibly compromised tissue. The crucial conceptual move is that ASPECTS is region-weighted rather than purely volumetric: loss of the insular ribbon or a single deep nucleus carries the same single-point penalty as a cortical subregion, embedding clinical knowledge that strategic territories matter.\n\nThe score’s value is overwhelmingly prognostic and decision-theoretic. In the original cohort, lower ASPECTS predicted worse functional outcome and symptomatic hemorrhage after intravenous thrombolysis, with an inflection around a score of 7. It subsequently became the anatomic gatekeeper of the thrombectomy era: the pivotal trials and the HERMES individual-patient meta-analysis (Goyal and colleagues, Lancet 2016) demonstrated benefit predominantly in patients with ASPECTS $\geq 6$, and ESCAPE (Goyal and colleagues, NEJM 2015) used multiphase imaging and a small-core/good-ASPECTS selection to achieve one of the largest treatment effects in stroke history. Accordingly the 2019 AHA/ASA guideline (Powers and colleagues) assigns the strongest recommendation for mechanical thrombectomy within 6 hours to patients with ASPECTS $\geq 6$, causative internal carotid or M1 occlusion, and NIHSS $\geq 6$.\n\nBayesian use of ASPECTS requires holding two facts simultaneously. First, the score is a noisy estimator of an underlying continuous quantity (core volume); inter-rater reliability is moderate and improves materially with structured training, stroke windows, and side-by-side reading. Second, its discriminating power is time-dependent — a low ASPECTS at 90 minutes implies a far more aggressive infarct (and worse collaterals) than the same score at 8 hours, and should lower the posterior probability of a favorable thrombectomy outcome more steeply. The mapping between ASPECTS and absolute core volume is imperfect: an ASPECTS of 6 corresponds on average to roughly $30\text{--}50\ \mathrm{mL}$ of DWI lesion but with wide scatter, which is exactly why CT perfusion or DWI is preferred when the clock is long. The dominant failure modes are technical and cognitive in equal measure. Old lacunar infarcts and chronic small-vessel hypodensity are misread as acute, spuriously lowering the score (mitigated by demanding focal swelling or comparing with priors); leukoaraiosis blurs grey–white margins symmetrically; and observers anchor to the first abnormal region and over-deduct. Automated ASPECTS software reduces variance but introduces automation bias — the algorithm’s number must be reconciled against the human read, not substituted for it.

Early ischemic changes (EIC) are the constellation of NCCT findings that appear before frank, well-demarcated hypodensity, and recognizing them is the single highest-yield perceptual skill in acute stroke CT. Mechanistically every sign is an expression of the same underlying event — cytotoxic edema raising regional water and lowering attenuation — manifesting first where grey–white contrast is intrinsically highest or where the affected nucleus is most water-sensitive. The insular ribbon sign, blurring and hypoattenuation of the thin band of insular cortex between the extreme capsule and the Sylvian cistern, is among the earliest because the insula sits at a watershed perfused by short, fragile insular perforators of the M2 segment with poor collateral redundancy. The obscured lentiform nucleus sign, loss of the normally sharp margin and slight hypoattenuation of the putamen and globus pallidus, reflects the high metabolic rate and end-arterial supply of the lenticulostriate territory, which infarcts rapidly and visibly when the proximal M1 is occluded. Cortical signs follow: loss of grey–white differentiation, effacement of cortical sulci from focal parenchymal swelling, and subtle hypoattenuation of the cortical ribbon itself.\n\nThe hyperdense vessel sign is mechanistically distinct — it is the thrombus, not the ischemia. Acute intraluminal clot, rich in retracted red cells and dense fibrin, measures higher than flowing blood; a hyperdense MCA sign typically shows the occluded segment at $\sim 50\text{--}70$ HU against $\sim 35\text{--}40$ HU contralaterally, and the analogous “dot sign” marks an M2/M3 branch thrombus in the Sylvian fissure. Its diagnostic utility is high specificity for occlusion when unequivocal and asymmetric, but it is insensitive and must be quantitatively confirmed, because the principal mimic — a high-hematocrit state or dense atherosclerotic calcification — raises attenuation bilaterally and along the vessel rather than focally. The search pattern proceeds from deep to superficial: insula, then lentiform, then cortical ribbon and sulci, then a deliberate trace of the proximal MCA, basilar, and the carotid terminus for hyperdensity. Bayesian weighting is direction-specific: in a patient with a dense hemispheric deficit, an insular ribbon sign or obscured lentiform substantially raises the posterior probability of a proximal MCA occlusion and predicts a larger early core, nudging toward urgent vessel imaging; conversely their absence at three hours is reassuring but non-exclusionary.\n\nThe failure modes are unforgiving. Volume averaging of the angled insular cortex against CSF fabricates a false insular ribbon sign; symmetric chronic small-vessel disease mimics grey–white loss bilaterally; and a unilaterally calcified or high-hematocrit vessel is over-called as thrombus — a $> 90$ HU “thrombus” is calcium until proven otherwise, and dual-energy or comparison with the contralateral vessel resolves it. Cognitively, EIC are the textbook substrate for both false-negative (the changes are by definition subtle and are missed on brain windows or under time pressure) and false-positive errors (an eager reader over-interprets normal asymmetry or artifact). The discipline that defends against both is invariant: narrow stroke windows, rigorous left–right comparison at matched levels, demanding that a putative hyperdense vessel be both focal and quantitatively denser than its mate, and correlating every candidate finding with the clinical territory and, when available, the CTA.

CT angiography (CTA) is the pivotal study that transforms a syndromic suspicion of stroke into a vascular diagnosis with a treatment target. A timed bolus of iodinated contrast (typically $\sim 50\text{--}70\ \mathrm{mL}$ at $4\text{--}6\ \mathrm{mL/s}$) is imaged during peak arterial opacification, ideally from the aortic arch through the vertex so that the entire path of a potential thrombectomy catheter — arch configuration, cervical carotid and vertebral origins, tandem cervical disease, and the intracranial circulation — is depicted in one acquisition. The physical content is straightforward: contrast raises intraluminal attenuation to several hundred HU, so an abrupt cutoff, a meniscus, or a non-opacified segment localizes occlusion with high conspicuity. Interpretation, however, demands a disciplined multiplanar method rather than reliance on a single maximum-intensity projection, because thin-section source images preserve the small branch occlusions and the precise clot margins that MIPs and volume renderings can obscure or exaggerate.\n\nThe systematic search follows the blood: from the arch and great-vessel origins, up each cervical carotid and vertebral artery noting stenosis or dissection, through the carotid siphon and vertebrobasilar junction, to the circle of Willis and the named intracranial segments — ICA terminus, M1 and M2/M3 of the MCA, A1/A2 of the ACA, P1/P2 of the PCA, and the basilar trunk and tip. Each named segment is interrogated for abrupt termination, intraluminal filling defect, and downstream paucity of branches. Beyond simple occlusion detection, two clot characteristics carry therapeutic weight: clot burden (the length and density of thrombus, formalized in clot-burden scores) and clot composition, since longer and more proximal thrombi are less likely to recanalize with intravenous thrombolysis alone and predict the need for, and difficulty of, mechanical retrieval. CTA source images additionally encode a perfusion surrogate — tissue that fails to opacify reflects severely reduced cerebral blood volume and correlates with core — a property exploited before formal CT perfusion was widespread.\n\nThe Bayesian logic is that CTA powerfully updates both diagnosis and management. In a patient with NIHSS $\geq 6$ and an early ischemic NCCT, the demonstration of an ICA-terminus or M1 occlusion raises the post-test probability of benefit from thrombectomy to near-certainty within the validated windows, whereas a patent large vessel shifts the differential toward distal/branch occlusion, lacune, or stroke mimic and away from large-vessel intervention. CTA is the proximate evidence the 2019 AHA/ASA guideline and the HERMES trials require to identify the treatable lesion. The failure modes are instructive. Mistimed acquisition — scanning before peak arterial enhancement or capturing the venous phase — produces pseudo-occlusion through inadequate opacification, the most dangerous false positive; severe cardiac output failure or proximal tandem stenosis delays bolus arrival and mimics distal occlusion. Beam-hardening from dense skull base or dental amalgam degrades the petrous and cavernous ICA; venous contamination and motion blur small branches. Cognitively, the reader who anchors on a hemispheric NCCT may overlook a basilar occlusion, and automation bias toward an AI vessel-detection flag can both manufacture and mask occlusions — the source images, read multiplanar, remain the arbiter.

Large-vessel occlusion (LVO) is the specific anatomic diagnosis that defines candidacy for mechanical thrombectomy and is therefore the organizing target of the entire acute imaging pathway. Operationally, LVO denotes occlusion of the intracranial internal carotid artery (including the carotid terminus or “T”), the M1 and proximal M2 segments of the middle cerebral artery, and — with distinct urgency — the basilar artery; occlusions of A1, P1, and the vertebral arteries are sometimes included by extension. The pathophysiologic significance is volumetric and temporal: occluding a proximal trunk simultaneously deprives a large territory and its deep, collateral-poor perforator beds, so the lenticulostriate-supplied basal ganglia infarct early and the cortex survives only to the degree that leptomeningeal collaterals sustain it. This is why LVO produces severe deficits (NIHSS typically $\geq 6$, and often much higher with a T-occlusion), and why the relationship between time, collateral grade, and core growth dominates outcome.\n\nThe imaging signatures are layered across the three studies. On NCCT, a hyperdense MCA or basilar sign and an obscured lentiform nucleus raise suspicion; on CTA, the diagnosis is made directly by the non-opacified segment and abrupt cutoff; on perfusion, the territory shows a large region of prolonged $T_{max}$ with a variably sized rCBF core. Quantitatively, the hyperdense segment exceeds $\sim 50\ \mathrm{HU}$ and is denser than the contralateral vessel; the M1 measures roughly $14\text{--}16\ \mathrm{mm}$ to its genu, so a clot spanning the proximal M1 implies a large captured perforator territory. Basilar occlusion deserves separate emphasis: it is frequently missed because the deficit (fluctuating consciousness, crossed signs, locked-in syndrome) is misattributed, NCCT shows only a subtly dense basilar, and the prognosis untreated is catastrophic — a low threshold for CTA of the posterior circulation is mandatory.\n\nThe Bayesian and evidentiary architecture is unusually firm. The 2015 thrombectomy trials — MR CLEAN (Berkhemer and colleagues), ESCAPE (Goyal and colleagues), EXTEND-IA (Campbell and colleagues), SWIFT PRIME, and REVASCAT — each randomized predominantly anterior LVO and were halted or reported with overwhelming benefit; the HERMES pooled analysis (Goyal and colleagues, Lancet 2016) reported a number-needed-to-treat of approximately $2.6$ for reduced disability, among the largest effects in vascular medicine. The DAWN (Nogueira and colleagues, NEJM 2018) and DEFUSE 3 (Albers and colleagues, NEJM 2018) trials extended the treatable window to 24 and 16 hours respectively for selected LVO with clinical–core or perfusion mismatch. Demonstrating an LVO in an eligible patient therefore moves the posterior probability of meaningful benefit from thrombectomy to near its ceiling. The failure modes mirror those of CTA: bolus mistiming or low cardiac output fabricates pseudo-occlusion; distal M2/M3 and basilar clots are under-detected on MIP-only review; and the cardinal cognitive error is failing to image the posterior circulation in a patient whose decreased consciousness is ascribed to metabolic causes. Confirming a candidate occlusion on thin multiplanar source images, and reconciling NCCT, CTA, and CTP into a single coherent territory, is the discipline that prevents both missed and overcalled LVO.

Collateral circulation is the physiologic variable that, more than the clock alone, determines how fast core grows behind a large-vessel occlusion and how much penumbra remains to be rescued. When a proximal trunk occludes, perfusion to the distal territory is maintained retrogradely through the leptomeningeal (pial) anastomoses that connect the terminal cortical branches of the ACA, MCA, and PCA, supplemented proximally by the circle of Willis. Robust pial collaterals sustain cerebral blood flow above the infarction threshold in the cortex even while the deep perforator territory infarcts, decoupling tissue fate from elapsed time; poor collaterals permit rapid, near-complete territorial infarction within an hour. Collateral status is thus the mechanistic explanation for the wide observed variance between time-from-onset and core volume, and it is an independent predictor of recanalization success, final infarct size, hemorrhagic transformation, and functional outcome.\n\nThe imaging assessment is intrinsically dynamic, which is the central technical insight. Single-phase CTA captures one time point in the contrast transit; because retrograde pial filling is delayed relative to antegrade arterial flow, a single early-phase image systematically under-represents collaterals and over-calls poor filling in slow-collateral patients — a timing artifact, not a true deficit. Multiphase CTA (Menon and colleagues, Radiology 2015) solves this by acquiring three time-resolved cranial passes (peak arterial, peak venous, and late venous) and grading the extent and delay of pial filling in the affected hemisphere relative to the normal side on an ordinal scale. The most influential validation came from ESCAPE (Goyal and colleagues, NEJM 2015), which used multiphase CTA collateral grading to select patients with moderate-to-good collaterals and achieved one of the largest thrombectomy treatment effects reported, establishing collateral imaging as a legitimate selection tool alongside core volume. Perfusion-derived surrogates — the hypoperfusion intensity ratio (the ratio of severely delayed $T_{max} > 10\,\mathrm{s}$ volume to $T_{max} > 6\,\mathrm{s}$ volume) — provide an automated correlate, with a low ratio indicating good collaterals.\n\nBayesian integration is where collateral grade earns its keep. In a late-window patient with an LVO, good collaterals raise the prior probability that a small core has been preserved despite elapsed time, increasing the expected mismatch and the posterior probability of benefit; poor collaterals predict a large or rapidly completing core and a higher hemorrhage risk, shifting the calculus away from intervention even when the clock is favorable. Collateral grade thus acts as a tissue-clock that partially substitutes for the wall-clock. The failure modes are dominated by acquisition timing: a single-phase study read as “poor collaterals” may simply be early; conversely, contamination by venous phase can over-estimate filling. Reduced cardiac output and proximal tandem stenosis globally delay contrast and degrade the assessment. Cognitively, the reader who evaluates collaterals on one phase, or who conflates slow filling with absent filling, will systematically under-treat exactly the slow-but-viable patients in whom collaterals are protecting salvageable cortex — the population thrombectomy most helps.

CT perfusion (CTP) interrogates the brain at the level of capillary hemodynamics, converting a dynamic first-pass contrast acquisition into quantitative maps of cerebral blood flow, cerebral blood volume, mean transit time, and time-to-maximum of the residue function. The physics rests on tracer-kinetic theory. After a tight bolus, repeated low-dose scans of a tissue slab capture the time–attenuation curve of contrast as it transits each voxel; using an arterial input function sampled from a large artery and a venous output function for normalization, deconvolution recovers the tissue residue function $R(t)$, whose properties yield the parametric maps. Cerebral blood volume (CBV) is the area under the tissue curve normalized to the arterial input, $\mathrm{CBV} = \dfrac{\int C_{tissue}(t)\,dt}{\int C_{art}(t)\,dt}$ (units $\mathrm{mL}/100\,\mathrm{g}$); cerebral blood flow (CBF) is the peak of the deconvolved residue ($\mathrm{mL}\,100\mathrm{g}^{-1}\mathrm{min}^{-1}$); and by the central volume principle the mean transit time is their ratio, $\mathrm{MTT} = \mathrm{CBV}/\mathrm{CBF}$. $T_{max}$, the time to the peak of the residue function, captures bolus-delay and dispersion and has proven the most robust and reproducible marker of critical hypoperfusion across vendors.\n\nThe clinical payload of CTP is the physiologic separation of tissue that is already dead from tissue that is merely threatened, and this separation is operationalized through validated thresholds derived against diffusion MRI and final infarct. Severely reduced relative CBF — the benchmark established by Cereda and colleagues (Journal of Cerebral Blood Flow and Metabolism, 2016) is $\mathrm{rCBF} < 30\%$ of the contralateral normal — best approximates the irreversibly infarcted core, while prolonged $T_{max} > 6\,\mathrm{s}$ (validated by Olivot, Lansberg and colleagues in the DEFUSE program) best approximates the total volume of critically hypoperfused tissue, core plus penumbra. The penumbra is then the arithmetic mismatch between these two volumes. This thresholding scheme is not arbitrary curve-fitting; it is the very logic that powered the late-window thrombectomy trials, in which automated software (most prominently RAPID) computed core and $T_{max}$ volumes at the bedside to enroll patients.\n\nBayesian use of CTP demands respect for its noise. The maps are quantitatively sensitive to the arterial input selection, motion, truncation of the time–attenuation curve if the acquisition is too short, and post-processing algorithm — different software can yield materially different core volumes from identical raw data. Therefore CTP updates the management probability strongly only when it is technically clean and concordant with NCCT, CTA, and the clinical deficit; a discordant or artifact-laden study should lower, not raise, one’s confidence. The failure modes are specific and consequential. A delayed or truncated bolus inflates $T_{max}$ and over-estimates penumbra; a poorly chosen arterial input function systematically biases all maps; chronic small-vessel disease and contralateral stenosis (which violate the assumption of a normal reference hemisphere) distort the relative thresholds; and motion produces uninterpretable striping. The cardinal cognitive error is automation bias — accepting the software’s color core and penumbra without verifying the input function placement, the bolus adequacy, and anatomic plausibility against the source data and the rest of the examination.

The distinction between infarct core and ischemic penumbra is the central organizing principle of acute stroke therapeutics: every reperfusion decision is, at bottom, a wager that a worthwhile volume of penumbra remains and that the core is small enough to make rescue safe. The concept is hemodynamic and threshold-defined. As cerebral blood flow falls, tissue passes through graded states. Above roughly $20\ \mathrm{mL}\,100\mathrm{g}^{-1}\mathrm{min}^{-1}$ function is preserved; between approximately $10$ and $20$ electrical activity ceases but membrane integrity is maintained — this is the penumbra, electrically silent yet structurally alive and salvageable if flow is restored; below approximately $10\ \mathrm{mL}\,100\mathrm{g}^{-1}\mathrm{min}^{-1}$ ion pumps fail, cytotoxic edema ensues, and the tissue becomes the core, irreversibly destined for infarction. The penumbra is therefore a spatial annulus of metabolically threatened brain surrounding the core, sustained by collateral flow and recruitable for minutes to many hours depending on collateral robustness — the tissue-clock that explains why fixed time windows are biologically crude.\n\nThe imaging operationalization translates these flow thresholds into measurable map volumes. On CTP, core is estimated as the volume with $\mathrm{rCBF} < 30\%$ and the total hypoperfused (core plus penumbra) volume as $T_{max} > 6\,\mathrm{s}$; the penumbra is the difference, and the therapeutic target is a favorable mismatch. The late-window trials made these numbers explicit and clinically binding. DEFUSE 3 (Albers and colleagues, NEJM 2018) required, for 6–16 hours, a core $< 70\ \mathrm{mL}$, a mismatch ratio (total hypoperfusion volume divided by core volume) $\geq 1.8$, and an absolute mismatch volume $\geq 15\ \mathrm{mL}$. DAWN (Nogueira and colleagues, NEJM 2018) used a clinical–core mismatch for 6–24 hours, stratifying allowable core by age and NIHSS — for example, a patient under 80 with NIHSS $\geq 10$ qualified with a core $< 31\ \mathrm{mL}$, and with NIHSS $\geq 20$ a core up to $51\ \mathrm{mL}$. The mismatch ratio can be written $MR = V_{T_{max}>6s}/V_{rCBF<30\%}$, and the absolute mismatch $V_{T_{max}>6s} - V_{rCBF<30\%}$; both must clear threshold.\n\nBayesian interpretation requires treating these volumes as estimates with measurement error rather than ground truth. A large mismatch in a technically excellent study with good collaterals raises the posterior probability of benefit steeply; a borderline mismatch, a discordant clinical picture, or a study degraded by motion or a poor arterial input function should temper enthusiasm, because the very algorithms that compute core can over- or under-estimate it by tens of milliliters. The penumbra is also temporally unstable — it converts to core continuously, so the measured mismatch is a snapshot whose validity decays with door-to-reperfusion time. The dominant failure modes are the “ghost core” (CTP over-estimating core in the very early window because reduced CBF in still-viable tissue has not yet become infarction, leading to inappropriate exclusion) and its converse, under-estimation of core when collaterals transiently support flow. Reconciling the CTP core against ASPECTS on NCCT, the CTA occlusion site, the collateral grade, and the clinical deficit — rather than trusting any single number — is the safeguard that turns the core–penumbra model into sound, individualized decisions.

Reperfusion injury and its most feared expression, hemorrhagic transformation, are the iatrogenic shadow of successful recanalization and the principal safety constraint that bounds every thrombolysis and thrombectomy decision. The pathophysiology is a paradox: restoring flow to ischemic tissue, while necessary to rescue the penumbra, can itself damage tissue whose microvasculature has been injured by the preceding ischemia. During occlusion, energy failure degrades the endothelial tight junctions and basal lamina of the blood–brain barrier, upregulates matrix metalloproteinases (notably MMP-9), and primes the parenchyma for edema and bleeding. When flow returns — particularly an abrupt, high-pressure return into a maximally vasodilated, autoregulation-lost bed — the compromised barrier leaks. The result spans a spectrum from benign contrast extravasation and vasogenic edema, through hemorrhagic infarction (petechial blood within the infarct), to parenchymal hematoma that exerts mass effect and worsens outcome. The probability of this cascade scales with the size and density of the established core, the duration of ischemia, hyperglycemia, and the intensity of reperfusion, which is precisely why a large rCBF core and a low ASPECTS are contraindications to aggressive reperfusion.\n\nThe imaging signatures require careful temporal and physical disambiguation, because the most common post-procedure finding — hyperdensity within the infarcted territory on the immediate post-thrombectomy CT — is frequently not hemorrhage at all but iodinated contrast staining extravasated through the leaky barrier. The two are difficult to separate on conventional single-energy CT because both are hyperdense, yet the distinction is decisive: contrast staining is benign and resolves, whereas hematoma is dangerous. Three discriminators are used. First, attenuation: contrast staining commonly exceeds the density of acute blood, often measuring $> 90\text{--}100\ \mathrm{HU}$ and frequently higher, whereas acute hematoma typically measures $\sim 50\text{--}70\ \mathrm{HU}$. Second, temporal evolution: contrast washes out and largely resolves on follow-up CT at 24 hours, while true hemorrhage persists or expands. Third, and most powerfully, dual-energy CT: by exploiting the energy dependence of iodine’s attenuation, material decomposition separates an iodine map from a virtual non-contrast image, definitively identifying staining versus blood on a single immediate scan. The hemorrhagic transformation grades themselves — hemorrhagic infarction type 1 and 2, and parenchymal hematoma type 1 and 2 — derive from the ECASS classification and carry escalating prognostic weight, with PH2 (hematoma exceeding $30\%$ of the infarct with significant mass effect) strongly associated with neurologic deterioration and death.\n\nBayesian and management integration is direct. The pretest probability of symptomatic hemorrhage was quantified in the foundational thrombolysis trials — NINDS reported $\sim 6\%$ symptomatic intracranial hemorrhage with intravenous tissue plasminogen activator within three hours, and ECASS III (Hacke and colleagues, NEJM 2008) reported $\sim 2.4\%$ in the 3–4.5 hour window — numbers that define the risk side of the risk–benefit ledger and underlie blood-pressure, glucose, and antithrombotic management after reperfusion. On encountering post-treatment hyperdensity, the reader updates toward benign staining if the density is very high and the patient is neurologically stable, and toward hematoma if there is mass effect, the patient deteriorates, or follow-up shows persistence — with dual-energy resolving the ambiguity when available. The failure modes are the mirror image of one another: calling benign contrast staining a hemorrhage triggers inappropriate reversal of antithrombotics and withholding of needed therapy, while dismissing a true early hematoma as staining delays life-saving management. Beam-hardening adjacent to a stented vessel and partial-volume averaging at the cortex compound the difficulty, and the cognitive trap is anchoring on the expected (“post-thrombectomy hyperdensity is always staining”) rather than applying the density, temporal, and dual-energy discriminators systematically.

Thrombectomy imaging is the integrated, time-critical workflow that converts the separate modalities of this chapter — NCCT, CTA, CTP — into a single, defensible decision to take a patient to the angiography suite, and into the procedural and post-procedural surveillance that follows. The governing principle is that imaging must answer four questions in sequence, fast enough to preserve penumbra: is there hemorrhage or a completed large infarct that forbids intervention (NCCT, ASPECTS); is there a large-vessel occlusion that can be reached and removed (CTA, with arch-to-vertex coverage planning the access route); is there salvageable tissue worth the risk (collateral grade and, in the late window, CTP core and mismatch); and is the patient within an evidence-supported window for these specific findings (the temporal–physiologic gate). The 2019 AHA/ASA guideline (Powers and colleagues) crystallizes the early-window standard: for patients within 6 hours with causative ICA or M1 occlusion, NIHSS $\geq 6$, and ASPECTS $\geq 6$, mechanical thrombectomy carries the strongest recommendation, grounded in the five 2015 trials and the HERMES meta-analysis (Goyal and colleagues, Lancet 2016) with its number-needed-to-treat near $2.6$.\n\nThe late window is where imaging selection becomes not merely useful but mandatory, because elapsed time can no longer serve as a proxy for tissue viability. DAWN (Nogueira and colleagues, NEJM 2018) extended treatment to 6–24 hours using clinical–core mismatch, and DEFUSE 3 (Albers and colleagues, NEJM 2018) to 6–16 hours using a perfusion core $< 70\ \mathrm{mL}$ with a mismatch ratio $\geq 1.8$ and absolute mismatch $\geq 15\ \mathrm{mL}$; both reported large, durable benefits in their imaging-selected populations and thereby made automated CTP (or DWI) the entry criterion for late presenters and wake-up strokes. Intraprocedurally, recanalization success is graded by the modified Thrombolysis in Cerebral Infarction (mTICI) scale, with mTICI 2b–3 (reperfusion of at least half, ideally all, of the territory) the threshold associated with good outcome and the explicit procedural goal; the relationship between the number of passes, time-to-reperfusion, and outcome is steep, so first-pass complete reperfusion is the prized result.\n\nThe Bayesian and operational logic ties the chapter together: each modality conditions the interpretation of the next, and the joint posterior — not any single image — drives the decision. A high ASPECTS, a retrievable M1 occlusion, good multiphase collaterals, and a large CTP mismatch in a patient with a severe deficit push the probability of benefit to its ceiling; degrade any one component and the calculus shifts. Post-procedure imaging closes the loop: an immediate NCCT (ideally dual-energy) distinguishes benign contrast staining from hemorrhagic transformation, and a 24-hour scan confirms final infarct, detects delayed hematoma, and informs antithrombotic timing. The failure modes are systemic as well as perceptual. Workflow delay is itself a failure mode — every minute of door-to-groin time converts penumbra to core, so the imaging pathway is engineered for speed (single-session NCCT/CTA, deferring CTP when the early-window anatomy already answers the question). Technical failures propagate: a mistimed CTA pseudo-occlusion sends a patient needlessly to angiography, a truncated or poorly deconvolved CTP fabricates or erases mismatch, and beam-hardening hides a basilar clot. Cognitively, automation bias toward software core/mismatch numbers, anchoring on the anterior circulation while a basilar occlusion is missed, and premature closure once one finding is identified are the recurring threats. The defense is invariant and is the thesis of this chapter: read every modality with its own discipline, reconcile them into one coherent physiologic and anatomic story, and let that integrated, error-aware posterior — calibrated to the governing trial evidence — decide.