Automated Measurement of Computed Tomography Acute Ischemic Core in Stroke
We are two excellent tailors, who, after many years of research, have managed to invent an extraordinary method of weaving a cloth so light and fine that it looks invisible. As a matter of fact, it is invisible to anyone who is too stupid and incompetent to appreciate its quality.
—Hans Christian Anderson, The Emperor’s New Clothes
See related article, pp 634
Bouslama et al1 have performed a very nice study comparing automated computed tomography (CT) ischemic core lesion volumes from acute noncontrast CT and perfusion CT (CTP) in a large number of patients with acute anterior circulation large vessel occlusion. The reference standard for the acute automated CT core lesion volumes was final infarct volume (FIV), predominantly measured on magnetic resonance imaging, within 72 hours of onset. To avoid that there would be infarct growth between acute CT and the FIV measurement the authors sensibly restricted analysis to patients who had complete reperfusion after endovascular thrombectomy. This should mean that if acute ischemic core measurements on CT are accurate, they should be very close to the FIV. Indeed, the authors concluded that automated NCCT and CTP ischemic core volumes were similar in estimating FIV. It would be more accurate (but less polite) to say that automated NCCT ischemic core volume and CTP core volume were similarly poor in estimating FIV.
Indeed, there were very large differences between CTP core and NCCT core volumes, with FIV somewhere in between. Median CTP core volume was 5 mL (interquartile range, 0–17.7 mL). Median NCCT core was 38.4 mL (interquartile range, 21.8–58 mL). Median FIV was 22.2 mL (interquartile range, 9.1–56.2 mL). The median individual difference in volumes between each modality is not presented, but if one examines the box plots and Bland-Altman plots, the majority of NCCT core volumes overestimated FIV. However, there are also a considerable minority of cases where the NCCT core volume underestimated FIV by >50 mL. The substantial underestimation of core in some cases is not so surprising given that in some patients with truly large acute core there is simply no, or minimal, ischemic change on acute NCCT (no matter how good the machine, or the human, is). For CTP core, however, the differences between FIV were even greater than seen with NCCT core but in contrast showed predominant underestimation of FIV. This may be partially related to CTP coverage of 8 cm, compared with whole-brain CT and magnetic resonance imaging measurements. However, there were a large number of cases where the undercall of FIV by CTP core was >50 mL. It is also worth bearing in mind that the majority of patients in this study had relatively small FIVs, so the large volumetric differences in acute core estimates, particularly with CTP, are quite alarming.
Because of this, it is highly likely that these data will be used in yet another oversimplified argument by the no need for perfusion imaging before thrombectomy proponents. I have no time for such arguments (nor for similar guideline recommendations). Suffice to say that in most cases it is not about using CTP to deny therapy. It is about making an accurate diagnosis, defining the pathophysiology, and the likely response to reperfusion therapy. Finally, it is about communicating with patients and their families. Indeed, showing a patient’s family the CTP images is an incredibly helpful communication tool. Reassuringly, in this study, both CTP and NCCT acute ischemic core volumes independently predicted 3-month outcome in multivariate analysis, yet again validating the clinical value of acute ischemic core estimation in predicting treatment response. However, this does not hide the fact that there are considerable issues with the accuracy of both NCCT- and CTP-based acute core estimates in this study. In particular, we do need to ask, why are the two acute core estimates used in this study volumetrically so far apart from each other?
First, the NCCT acute ischemic core volume was estimated by Brainomix software (Oxford, United Kingdom). Dr Nogueira and team are to be congratulated for focusing on the volume of acute ischemic change as a measure of core on NCCT, rather than rely on an oversimplified score, the electronic Alberta Stroke Program Early CT Score (eASPECTS). It was apparent to us Australians that ASPECTS was becoming extinct some time ago, first with diffusion-weighted imaging core, and then, we thought the final death blow came with the advent of volumetric core estimates on CTP.2 However, the availability of automated ASPECTS has temporarily prolonged its use-by date (contributed to by effective marketing from the various software companies that now produce an eASPECTS output). The irony is that eASPECTS relies on automated measures of early ischemic change on NCCT. The question is, how often can early ischemic change on acute NCCT be reliably detected? And, if so, how reliably does it estimate the ischemic core? To complicate matters, we know that there are 2 types of early ischemic change: hypodensity, which clearly reflects irreversible injury and is due to ischemic edema; and focal swelling, which, when not associated with hypodensity, might actually be salvageable.3 Thus, training a machine, as Brainomix have done, using a machine learning classifier trained on a large dataset to distinguish between ischemic and nonischemic regions is easier said (and marketed) than done. This still means human experts have had to classify regions on NCCT as acute ischemia versus nonischemia in the first place. Thus, the process is still quite prone to bias and human error, no matter how good the machine learning algorithm is. For example, if one teaches the algorithm to be more sensitive to subtle changes in Hounsfield Units (hypodensity), one might expect an overcall of the true ischemic core. This is likely what has occurred in many cases in the current study. Nonetheless, this study is a good first step on the pathway of automating volumetric acute ischemic core measurement on NCCT. However, NCCT core estimation will always be limited by the fact that the majority of hypoperfused regions appear normal on NCCT.3 This is, of course, where CTP assists. It is also apparent that estimates of acute ischemic core and NCCT and CTP are complimentary. For example, it is not uncommon to see hypodensity on NCCT that is outside the CTP core and sometimes even outside the entire perfusion lesion. The most obvious example of such is when spontaneous reperfusion is occurring during the imaging process. I still find that even very astute clinicians fail to recognize that if reperfusion (partial or full) has occurred, then CTP core will not be an accurate estimate of the true core. Thus, I believe that the future will be to use raw data from concurrent NCCT and CTP to more reliably measure acute ischemic core. And it is inevitable that this future will involve artificial intelligence/machine learning/deep learning analysis.
Moving on to CTP, the core estimates were calculated with RAPID (Rapid Processing of Perfusion and Diffusion) software using a CBF <30% threshold (IschemaView, Menlo Park, CA). This is probably the best validated CTP threshold for core estimation across different software/algorithms.4–6 So, it is hard to explain why RAPID CTP core so drastically underestimated FIV in this large dataset. To calculate the various perfusion measures, RAPID uses a variant of deconvolution utilizing a Fourier transformation that is said to be equivalent to a block circulant single value deconvolution and has been postulated to be delay insensitive.7 Even if that is true, simply correcting for the delay of contrast arriving to the ischemic region does not take into account that the contrast bolus is dispersed (via collateral pathways) by the time it reaches the ischemic region. This means that the real tissue arterial concentration/time curve is shorter and fatter than in the proximal arteries. Failure to correct for this leads to underestimation of CBF in ischemic tissue and hence overestimation of the ischemic core volume. This was certainly the case in a previous study where we tried to approximate the RAPID perfusion algorithm.7 However, we do not know what postprocessing goes on in the RAPID box after the perfusion calculation to generate the end product, visually appealing, smoothed maps. In the current study, the lower interquartile range for RAPID core was zero. It is hard to imagine that >25% of large vessel occlusion patients had no core and suggests considerable smoothing of the data by RAPID. To further add to the perplexing CTP results of the current study, recent data suggest that if complete and reasonably fast reperfusion (within 90 minutes of CTP) occurs with thrombectomy, the 30% CBF threshold will overestimate the FIV.6 The times between CT and reperfusion are not presented, but if there were delays, and hence infarct growth during that period, this, combined with limited CTP brain coverage, might paradoxically explain underestimation of core by RAPID in some cases. However, it seems unlikely long delays between CT and reperfusion would have systematically occurred in a high-quality center.
The authors have concluded that automated core volumes from NCCT and CTP could be used alongside each other for patient assessment.1 I totally agree. There is great potential to combine their information, which is clearly complimentary. However, we have a lot of work to do with getting the most out of both modalities, particularly CTP. I believe the data from the current study highlight that, as in many walks of life, sometimes the product does not live up to marketing hype. Does the CTP Emperor have no clothes? Well, perhaps not quite, but he certainly has been caught with his pants down, revealing pink and green designer underwear. Forget the marketing, let’s focus on the science and getting the most out of the raw imaging data. We can (and should) do better.
Sources of Funding
Disclosures Dr Parsons reports research collaborations (discounted software) with Siemens, Canon, and Apollo Medical Imaging (MIStar).
Bouslama M, Ravindran K, Harston G, Rodrigues G, Pisani L, Haussen D, Frankel MR, Nogueira R. Noncontrast computed tomography e-Stroke infarct volume is similar to RAPID computed tomography perfusion in estimating postreperfusion infarct volumes.Stroke. 2021; 52:634–641. doi: 10.1161/STROKEAHA.120.031651LinkGoogle Scholar
Bivard A, Parsons M. ASPECTaSaurus (a dinosaur)?Int J Stroke. 2012; 7:564. doi: 10.1111/j.1747-4949.2012.00854.xCrossrefMedlineGoogle Scholar
Parsons MW, Pepper EM, Bateman GA, Wang Y, Levi CR. Identification of the penumbra and infarct core on hyperacute noncontrast and perfusion CT.Neurology. 2007; 68:730–736. doi: 10.1212/01.wnl.0000256366.86353.ffCrossrefMedlineGoogle Scholar
Campbell BC, Christensen S, Levi CR, Desmond PM, Donnan GA, Davis SM, Parsons MW. Cerebral blood flow is the optimal CT perfusion parameter for assessing infarct core.Stroke. 2011; 42:3435–3440. doi: 10.1161/STROKEAHA.111.618355LinkGoogle Scholar
Mokin M, Levy EI, Saver JL, Siddiqui AH, Goyal M, Bonafé A, Cognard C, Jahan R, Albers GW; SWIFT PRIME Investigators. Predictive value of RAPID assessed perfusion thresholds on final infarct volume in SWIFT PRIME (Solitaire With the Intention for Thrombectomy as Primary Endovascular Treatment).Stroke. 2017; 48:932–938. doi: 10.1161/STROKEAHA.116.015472LinkGoogle Scholar
Bivard A, Kleinig T, Miteff F, Butcher K, Lin L, Levi C, Parsons M. Ischemic core thresholds change with time to reperfusion: a case control study.Ann Neurol. 2017; 82:995–1003. doi: 10.1002/ana.25109CrossrefMedlineGoogle Scholar
Lin L, Bivard A, Kleinig T, Spratt NJ, Levi CR, Yang Q, Parsons MW. Correction for delay and dispersion results in more accurate cerebral blood flow ischemic core measurement in acute stroke.Stroke. 2018; 49:924–930. doi: 10.1161/STROKEAHA.117.019562LinkGoogle Scholar