Enhancing performance of a computed tomography perfusion software for improved prediction of final infarct volume in acute ischemic stroke patients

Abstract

Computed tomography perfusion (CTP) is crucial for acute ischemic stroke (AIS) patient diagnosis. To improve infarct prediction, enhanced image processing and automated parameter selection have been implemented in Vital Images’ new CTP+ software. We compared CTP+ with its previous version, commercially available software (RAPID and Sphere), and follow-up diffusion-weighted imaging (DWI). Data from 191 AIS patients between March 2019 and January 2020 was retrospectively collected and allocated into endovascular intervention (n = 81) and conservative treatment (n = 110) cohorts. Intervention patients were treated for large vessel occlusion, underwent mechanical thrombectomy, and achieved successful reperfusion of thrombolysis in cerebral infarction 2b/2c/3. Conservative treatment patients suffered large or small vessel occlusion and did not receive intravenous thrombolysis or mechanical thrombectomy. Infarct and penumbra were assessed using intervention and conservative treatment patients, respectively. Infarct and penumbra volumes were segmented from CTP+ and compared with 24-h DWI along with RAPID, Sphere, and Vitrea. Mean infarct differences (95% confidence intervals) and Spearman correlation coefficients (SCCs) between DWI and each CTP software product for intervention patients are: CTP+  = (5.8 ± 5.9 ml, 0.62), RAPID = (10.0  ± 5.2 ml, 0.73), Sphere = (3.0 ± 6.0 ml, 0.56), Vitrea = (7.2 ± 4.9 ml, 0.66). For conservative treatment patients, mean infarct differences and SCCs are: CTP+ = (–8.0 ± 5.4 ml, 0.64), RAPID = (–25.6 ± 11.5 ml, 0.60), Sphere = (–25.6 ± 8.0 ml, 0.66), Vitrea = (1.3 ± 4.0 ml, 0.72). CTP+ performed similarly to RAPID and Sphere in addition to its semi-automated predecessor, Vitrea, when assessing intervention patient infarct volumes. For conservative treatment patients, CTP+ outperformed RAPID and Sphere in assessing penumbra. Semi-automated Vitrea remains the most accurate in assessing penumbra, but CTP+ provides an improved workflow from its predecessor.

Introduction

Rapid reperfusion in acute ischemic stroke (AIS) patients is essential to reduce risk of disability and mortality.¹ Computed tomography perfusion (CTP) is an imaging modality used to diagnose ischemic tissue in AIS patients and, in combination with patient demographics such as age, time since onset of stroke symptoms, National Institute of Health Stroke Scale (NIHSS) score, and cardiac output, assess a patient’s eligibility for endovascular reperfusion procedures.^2–4 A penumbra-to-infarct ratio of 1.8 is preferred to deem a patient eligible for such procedures.⁵ Multiple studies have shown that patients whose baseline CTP imaging accurately predicted final infarct volumes (FIVs), when compared within follow-up diffusion-weighted magnetic resonance imaging (DWI) within 24 h, achieve successful reperfusion.⁶ Although CTP is relied upon to assess reperfusion eligibility, multiple CTP software products exist that utilize different perfusion parameters and thresholds for identifying infarct and penumbra tissue.⁷,⁸

The aforementioned perfusion parameters are calculated using arterial input function (AIF) and venous output function (VOF) curves. The AIF and VOF locations are automatically selected within the CTP volume as a feeding artery and a draining vein, respectively. Typically, the AIF intensity values will appear lower in intensity than they actually are due to partial volume effect from tissue. This occurs since the size of the feeding artery does not fully encompass the voxels used as the selection location. The VOF is therefore used to normalize the AIF since the draining vein is much larger, leading to the selected voxels likely being fully encompassed by the vein and because the amount of contrast that flows through the vein should equal that which flows through the artery.⁹ Following scaling of the AIF, perfusion parameters such as cerebral blood flow (CBF) and cerebral blood volume (CBV) can be calculated from the AIF using decomposition or Bayesian techniques.⁹,¹⁰

RAPID (iSchemaView, Menlo Park, CA, USA) and Sphere (Olea Medical, La Ciotat, France) CTP software are two commercially available software products that primarily rely on the CBF parameter, but utilize different contralateral hemisphere thresholds to identify infarct.⁸,¹¹ Reliance on the CBF parameter can, however, lead to false positive identification of infarct in regions of leukoaraiosis. This decrease in CBF, allowing for the segmentation of infarct, is due to white matter naturally having lower amount of blood flow than gray matter, especially in regions of leukoaraiosis.¹²,¹³ Meanwhile, Vitrea and recently developed CTP+ (Vital Images, Minnetonka, MN, USA) software primarily rely on CBV to identify infarct.⁸ Although previous studies have been conducted assessing the reliability of infarct and penumbra segmentation using RAPID, Sphere, and Vitrea, similar studies need to be conducted on the recently developed CTP+ software.⁷,⁸,¹¹,¹²,¹⁴

We aimed to assess the agreement of FIVs and penumbra between DWI and the newly developed CTP+ software. Additionally, infarct and penumbra volumes from CTP+ were compared with commercially available RAPID and Sphere, along with CTP+’s predecessor, Vitrea.

Methods

Patient inclusion

Institutional review board approval was obtained and informed consent was waived for this Health Insurance and Accountability retrospective study. This study included AIS patients who underwent initial non-contrast computed tomography (CT) to rule out hemorrhage, baseline CTP imaging upon comprehensive stroke center arrival, and DWI 24 h following CTP imaging. One hundred and ninety-one consecutive patients between March 2019 and January 2020 met the inclusion criteria and were allocated into endovascular intervention and conservative treatment cohorts. Intervention patients (n = 81) were included for the assessment of infarct, were required to have emergent large vessel occlusion, and have undergone successful reperfusion (thrombolysis in cerebral infarction (TICI) 2b/2c/3) through mechanical thrombectomy. Patients who achieve recanalization of TICI 1 or 2a were not included since it is unclear what combination of infarct and penumbra would correspond to infarct on follow-up DWI as it is expected that some penumbra would be salvaged. Therefore, an estimation would need to be made of how much penumbra converts to infarct for each of these recanalization scores. Additionally, TICI 0 patients were not included due to a small sample size and conservative treatment patients were used as an alternative. TICI scores were measured as a consensus by an attending physician and two endovascular fellows not involved in data collection. Conservative treatment patients (n = 110) were included for the assessment of penumbra, included large and small vessel occlusion patients, and did not receive mechanical thrombectomy or intravenous thrombolysis. Since reperfusion techniques were not utilized for conservative treatment patients, it was assumed all CTP imaging penumbra converted to infarct on follow-up DWI allowing for penumbra assessment.¹⁵

CTP analysis

CTP data were collected as a series of 19 volumes using the Neuro ONE protocol. Volumes containing 320 slices, with a slice thickness of 0.5 mm, and an in-plane resolution of 0.42 mm were obtained using 512 row by 512 column detector Aquilion ONE CT units (Canon Medical Systems Corporation, Otawara, Japan). During the scan, Omnipaque 350 was injected at a volume of 50 ml and a rate of 5 ml/s. Scan parameters (mean  ±  95% confidence interval) included a tube voltage of 80 kV peak, tube current of 150 mA, CT dose index 34.3 ± 4.4 mGy, and dose length product 346.5 ± 103.6 mGy·cm. Scan acquisition started 7 s after contrast administration, duration time was 49.3 s, and reconstructed CTP volumes were available within 5 min from the start of scan for post-processing.

Automated analysis was conducted for segmentation of infarct and penumbra using RAPID, Sphere, and CTP+. Vitrea ischemic tissue segmentation was conducted using Vital Images’ recommended workflow of manually segmenting the top, middle, and bottom of the ischemic region and interpolating within the region of interest following automated segmentation. Manual segmentation was conducted by an operator with 2.5 years of experience of examining CTP scans. Recommended Vitrea workflow was compared with CTP+ to observe potential improvements in clinical workflow and tissue segmentation for Vital Images’ automated CTP+ software.

During CTP processing, RAPID, Sphere, Vitrea, and CTP+ resample the input CTP volumes to have slice thicknesses of 10 mm, 5 mm, 2 mm, and 2 mm, respectively, to reduce noise. Selection of the AIF and VOF were conducted automatically by each software. Raw contrast intensity data from the selected AIFs and VOFs were exported from both CTP+ and Vitrea for later comparison. iSchemaView conducted RAPID analysis offsite and segmented tissue with CBF  < 30% relative to the contralateral hemisphere as infarct and tissue with time to reach maximum residue function (Tmax) >6 s as penumbra. Sphere (SP20-3.0) segmented tissue with CBF  < 25% and time-to-peak (TTP) >5 s relative to the contralateral hemisphere as infarct and tissue with TTP>5 s compared with the contralateral side as penumbra. Vitrea (7.10) segmented infarct as tissue with either CBV  < 38% and TTP >5.3 s or mean-transit-time (MTT)  < 55% relative to the contralateral hemisphere and tissue with CBV >38% and either TTP >5.3 s or CBF  < 58% relative to the contralateral hemisphere as penumbra. CTP+ segmented infarct as CBV  < 40% and Tmax >2.2 s relative to the contralateral hemisphere and penumbra as Tmax >2.2 s compared with the contralateral side. CTP+ includes improvements from Vitrea by offering: a fully automated solution, improved AIF and VOF auto-selection, improved ventricle segmentation (Figure 1), improved summary map delineation, and use of the Tmax parameter instead of TTP to segment penumbra. Figure 2 compares infarct segmentation for a conservative treatment patient using all four tested CTP software products with DWI.

Figure 1.

The improved ventricle segmentation (purple) of computed tomography perfusion plus (CTP+) (right) compared with Vitrea (left). Notice in Vitrea, segmented ventricles would sometimes appear as infarct (red) and the entire ventricle would not be completely segmented as compared with CTP+.

Figure 2.

Diffusion-weighted imaging (DWI) and computed tomography perfusion (CTP) software images corresponding to a conservative treatment patient with a left middle cerebral artery occlusion. Hyper-intensified regions in the DWI correspond to infarct while pink and green regions correspond to infarct and penumbra, respectively, in the RAPID perfusion map. For CTP+, Sphere, and Vitrea, red and yellow regions correspond to infarct and penumbra, respectively. Note the strong agreement of ischemic tissue location between all CTP predictions and final DWI infarct.

DWI analysis

Collection of DWI data was conducted using IsoDWI protocol with a Vantage Titan 1.5 Tesla MRI unit (Canon Medical Systems Corporation, Otawara, Japan). IsoDWI protocol utilizes a repetition time of 8700 ms, an echo time of 100 ms, a diffusion b-value of 1000 s/mm², a slice thickness of 5 mm, and an in-plane resolution of 0.81 mm. Infarct was segmented from DWIs as regions with a 162% increase in intensity compared with the contralateral hemisphere, similar to a previously conducted method.¹⁶

Statistical analysis

Frequency distributions for categorical variables and summary statistical for continuous variables were tabulated for analyzed data. Raw data from AIFs and VOFs were analyzed by determining the number of curves which were simply noise and the number that had the improper bell curve shape. A curve was considered to be noise if all or all but one of the intensity values were less than 33% of the mean of maximum values from all AIFs or VOFs in the study. AIFs and VOFs were considered to have the incorrect shape if there were multiple peak intensities above 50% of the maximum intensity of the curve. Area under the curve (AUC) values were additionally calculated for each AIF and VOF curve from Vitrea and CTP+ along with AIF scalars, which are the VOF AUC divided by the AIF AUC.

Mean differences between DWI FIVs and predicted CTP infarct volumes were calculated for each software product for intervention and conservative treatment patients. Predicted infarct for conservative treatment patients was determined as a summation of infarct and penumbra on baseline CTP imaging. Analysis of variance (ANOVA) and Student t-test analysis were conducted on mean infarct differences comparing tested CTP software. Distributions of infarct differences for each CTP software product were included as mountain plots. Median absolute deviations (MADs) were calculated between final and predicted infarct volumes for each software. The percentage of patients with infarct overestimation was calculated for each category. Subgroup analysis was conducted determining all of the aforementioned metrics based on reperfusion status in the intervention category. Spearman correlation coefficients (SCCs) were determined between DWI FIVs and predicted infarct volume and regression analysis was conducted along with Bland–Altman plot generation. Inter-reader variability was conducted across five Vitrea users to assess recommended Vitrea workflow infarct segmentation through mean infarct difference and MAD calculations. Processing times to determine ischemic tissue were determined for all software. RAPID processing time was defined as time from completion of CTP scan until RAPID perfusion maps were uploaded to the hospital’s picture achieving and communications system. Sphere and CTP+ processing times were defined as time from loading CTP volumes into each software until perfusion maps were generated. Vitrea processing time was defined as time from loading CTP volumes into Vitrea until recommended workflow segmentation was completed.

Results

FIVs, predicted infarct volumes, and patient demographics for all patients are represented in Table 1, along with the intervention and conservative treatment cohorts. Similar metrics are detailed in Table 2 based on intervention reperfusion status. A significant difference (p < 0.05) is seen between the two cohorts for the age, NIHSS, and time from onset of symptoms to perfusion imaging characteristics. Intravenous thrombolysis was administered at a median of 115.0 min (interquartile range, 85.0–180.0 min) following symptom onset to 33/81 (40.7%) intervention patients.

Table 1.

Characteristics and outcomes of ischemic stroke patients.

Characteristic	AllN = 191	Endovascular interventionn = 81	Conservative treatmentn = 110
Male sex, %	46.6% (89/191)	51.9% (42/81)	42.7% (47/110)
Age, years, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	68.5$\hspace{0.17em}\pm \hspace{0.17em}$14.5 [68.0] (60.0–79.0)	66.0$\hspace{0.17em}\pm \hspace{0.17em}$16.2 [65.0] (57.0–79.0)	70.4$\hspace{0.17em}\pm \hspace{0.17em}$12.8 [70.0] (63.3–80.0)
NIHSS score, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	11.7$\hspace{0.17em}\pm \hspace{0.17em}$7.6 [10.0] (6.0–16.0)	14.3$\hspace{0.17em}\pm \hspace{0.17em}$7.5 [13.0] (9.0–20.0)	9.8$\hspace{0.17em}\pm \hspace{0.17em}$7.2 [7.0] (5.0–14.8)
Site of occlusion
Middle cerebral artery	70.2% (134/191)	74.1% (60/81)	67.3% (74/110)
Anterior cerebral artery	1.6% (3/191)	0.0% (0/81)	2.7% (3/110)
Posterior cerebral artery	14.1% (27/191)	4.9% (4/81)	20.9% (23/110)
Internal carotid artery	10.5% (20/191)	14.8% (12/81)	7.3% (8/110)
Basilar	3.7% (7/191)	6.2% (5/81)	1.8% (2/110)
Time from onset of stroke to perfusion imaging, minutes, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	968.8$\hspace{0.17em}\pm \hspace{0.17em}$373.5 [318.0] (122.5–877.5)	598.1$\hspace{0.17em}\pm \hspace{0.17em}$1235.6 [186.0] (85.0–552.0)	1241.8$\hspace{0.17em}\pm \hspace{0.17em}$3261.9 [567.0] (172.3–1027.5)
Time from onset of stroke to reperfusion, minutes, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	–	728.4$\hspace{0.17em}\pm \hspace{0.17em}$1276.3 [266.0] (161.0–717.0)	–
Time from perfusion imaging to recanalization, minutes, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	–	130.3$\hspace{0.17em}\pm \hspace{0.17em}$376.4 [64.0] (55.0–83.0)	–
CTP+ infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	28.7$\hspace{0.17em}\pm \hspace{0.17em}$57.6 [7.3] (2.4–30.0)	26.0$\hspace{0.17em}\pm \hspace{0.17em}$57.9 [7.6] (1.8–23.2)	30.7$\hspace{0.17em}\pm \hspace{0.17em}$57.6 [6.7] (2.6–34.2)
RAPID infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	37.0$\hspace{0.17em}\pm \hspace{0.17em}$72.5 [6.0] (0.0–39.5)	21.8$\hspace{0.17em}\pm \hspace{0.17em}$41.6 [4.0] (0.0–30.0)	48.3$\hspace{0.17em}\pm \hspace{0.17em}$87.1 [10.0] (0.0–54.3)
Sphere infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	40.1$\hspace{0.17em}\pm \hspace{0.17em}$55.2 [22.7] (7.5–44.1)	28.8$\hspace{0.17em}\pm \hspace{0.17em}$42.8 [14.9] (5.1–33.1)	48.3$\hspace{0.17em}\pm \hspace{0.17em}$61.7 [23.8] (10.0–58.6)
Vitrea infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	22.8$\hspace{0.17em}\pm \hspace{0.17em}$50.1 [5.2] (1.3–19.8)	24.6$\hspace{0.17em}\pm \hspace{0.17em}$52.0 [7.1] (1.4–19.5)	21.4$\hspace{0.17em}\pm \hspace{0.17em}$48.8 [4.5] (1.1–19.7)
Post-treatment reperfusion
TICI 2b	–	37.0% (30/81)	–
TICI 2c	–	27.2% (22/81)	–
TICI 3	–	35.8% (29/81)	–
Final DWI MRI infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	26.6$\hspace{0.17em}\pm \hspace{0.17em}$52.1 [7.9] (1.6–27.0)	31.8$\hspace{0.17em}\pm \hspace{0.17em}$52.8 [18.3] (4.5–41.2)	22.7$\hspace{0.17em}\pm \hspace{0.17em}$51.5 [5.1] (0.7–16.5)

Empty cells correspond to non-applicable data based on category or combination of categories.

IQR: interquartile range; NIHSS: National Institute of Health Stroke Scale; CTP+: computed tomography perfusion plus; TICI: thrombolysis in cerebral infarction, DWI: diffusion-weighted imaging; MRI: magnetic resonance imaging.

Table 2.

Characteristics and outcomes of ischemic stroke patients based on reperfusion status.

Characteristic	TICI 2b n = 30	TICI 2c n = 22	TICI 3 n = 29
Male sex, %	60.0% (18/30)	54.5% (12/22)	41.4% (12/29)
Age, years, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	66.3$\hspace{0.17em}\pm \hspace{0.17em}$15.0 [65.0] (58.3–76.3)	70.1$\hspace{0.17em}\pm \hspace{0.17em}$14.0 [73.0] (61.0–80.5)	62.5$\hspace{0.17em}\pm \hspace{0.17em}$18.6 [65.0] (52.0–73.0)
NIHSS score, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	14.0$\hspace{0.17em}\pm \hspace{0.17em}$6.7 [14.0] (9.3–19.0)	14.3$\hspace{0.17em}\pm \hspace{0.17em}$7.7 [12.0] (8.3–19.5)	14.6$\hspace{0.17em}\pm \hspace{0.17em}$8.3 [14.0] (10.0–20.0)
Site of occlusion
Middle cerebral artery	90.0% (27/30)	77.3% (17/22)	55.2% (16/29)
Posterior cerebral artery	0.0% (0/30)	0.0% (0/22)	13.8% (4/29)
Internal carotid artery	10.0% (3/30)	22.7% (5/22)	13.8% (4/29)
Basilar	0.0% (0/30)	0.0% (0/22)	17.2% (5/29)
Time from onset of stroke to perfusion imaging, minutes, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	344.5$\hspace{0.17em}\pm \hspace{0.17em}$359.7 [183.0] (86.8–499.5)	756.3$\hspace{0.17em}\pm \hspace{0.17em}$1871.8 [147.0] (80.0–462.0)	740.4$\hspace{0.17em}\pm \hspace{0.17em}$1221.0 [278.0] (104.0–647.0)
Time from onset of stroke to reperfusion, minutes, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	410.6$\hspace{0.17em}\pm \hspace{0.17em}$361.7 [243.0] (150.5–559.8)	858.0$\hspace{0.17em}\pm \hspace{0.17em}$1866.2 [248.5] (143.5–634.5)	958.9.4$\hspace{0.17em}\pm \hspace{0.17em}$1315.4 [397.0] (232.0–975.0)
Time from perfusion imaging to recanalization, minutes, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	66.0$\hspace{0.17em}\pm \hspace{0.17em}$24.9 [61.5] (53.5–71.8)	101.7$\hspace{0.17em}\pm \hspace{0.17em}$108.4 [63.5] (58.0–82.3)	218.5$\hspace{0.17em}\pm \hspace{0.17em}$618.2 [69.0] (48.0–99.0)
CTP+ infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	32.8$\hspace{0.17em}\pm \hspace{0.17em}$81.2 [7.5] (2.7–25.3)	24.8$\hspace{0.17em}\pm \hspace{0.17em}$30.1 [13.0] (6.7–34.4)	19.9$\hspace{0.17em}\pm \hspace{0.17em}$16.4 [6.3] (1.0–17.7)
RAPID infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	26.2$\hspace{0.17em}\pm \hspace{0.17em}$47.5.5 [12.0] (0.0–36.0)	27.2$\hspace{0.17em}\pm \hspace{0.17em}$39.0 [6.5] (0.0–42.3)	13.1$\hspace{0.17em}\pm \hspace{0.17em}$36.7 [0.0] (0.0–14.0)
Sphere infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	32.2$\hspace{0.17em}\pm \hspace{0.17em}$50.5 [22.3] (6.8–40.7)	32.4$\hspace{0.17em}\pm \hspace{0.17em}$38.0 [16.9] (5.3–34.8)	22.6$\hspace{0.17em}\pm \hspace{0.17em}$38.1 [10.1] (4.4–29.9)
Vitrea infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	31.8$\hspace{0.17em}\pm \hspace{0.17em}$69.3 [8.4] (3.1–29.8)	23.4$\hspace{0.17em}\pm \hspace{0.17em}$36.3 [12.2] (3.2–21.9)	18.1$\hspace{0.17em}\pm \hspace{0.17em}$40.8 [2.4] (1.3–16.9)
Final DWI MRI infarct volume, mL, mean$\hspace{0.17em}\pm \hspace{0.17em}$standard deviation, [median] (IQR)	37.3$\hspace{0.17em}\pm \hspace{0.17em}$68.6 [25.5] (5.8–37.4)	31.1$\hspace{0.17em}\pm \hspace{0.17em}$30.9 [24.3] (6.9–43.3)	26.8$\hspace{0.17em}\pm \hspace{0.17em}$48.1 [14.2] (3.3–27.4)

IQR: interquartile range; NIHSS: National Institute of Health Stroke Scale; CTP+: computed tomography perfusion plus; TICI: thrombolysis in cerebral infarction; DWI: diffusion-weighted imaging; MRI: magnetic resonance imaging.

AIF and VOF analysis

Table 3 compares the number of AIFs and VOFs from Vitrea and CTP+ that were determined to be noisy curves or have the improper bell curve shape. Additionally, Table 3 indicates the median AUC values for the AIFs and VOFs included in this study along with the median AIF scalars for CTP+ and Vitrea. Figure 3 indicates the AIF selection location for what was a noisy curve in Vitrea but has improved based on the selection location in CTP+. The AIF from Vitrea in Figure 3 is considered noisy because it has one high spike in contrast intensity and the rest of the intensities are background noise. Figure 4 indicates the VOF selection locations for what was a noisy curve in Vitrea and is now a smoother, bell shaped curve in CTP+. Figures 5 and 6 indicate the AIF and VOF selection locations, respectively, within CTP+ and Vitrea with CTP+ having more properly shaped curves without multiple contrast intensity peaks.

Table 3.

Comparison of AIFs and VOFs between CTP+ and Vitrea perfusion software.

	Vitrea	CTP+
AIF of noise, %	3.14% (6/191)	2.09% (4/191)
VOF of noise, %	6.81% (13/191)	2.09% (4/191)
Improper AIF shape, %	4.19% (8/191)	3.66% (7/191)
Improper VOF shape, %	1.57% (3/191)	0.52% (1/191)
AIF AUC, dHU$·$s, median (IQR)	8170 (6860–10300)	8620 (7190–10400)
VOF AUC, dHU$·$s, median (IQR)	8240 (6430–9880)	9680 (8000–11200)
AIF scalar, median (IQR)	0.95 (0.85–1.08)	1.12 (1.03–1.23)

AIF: arterial input function; VOF: venous output function; CTP+: computed tomography perfusion plus; AUC: area under the curve; dHU: derived Hounsfield unit; IQR: interquartile range.

Figure 3.

A comparison of the arterial input function (AIF) selection locations (red dot) for computed tomography perfusion plus (CTP+) (upper left) and Vitrea (upper right) along with the corresponding AIF intensity curves for each software. Note that the selected location within CTP+ is on a main artery while the Vitrea location is in a significantly anterior location. This leads to the noisy generated AIF (bottom right) which contains one spike in contrast intensity.

Figure 4.

Venous output function (VOF) curves and selection locations (blue dot) for computed tomography perfusion plus (CTP+) (left) and Vitrea (right). Note the location of the VOF selection in CTP+ is along the sagittal sinus vein while selection in Vitrea is erroneously placed in a region containing no vein. This erroneous selection leads to the noisy curve indicated in the lower right.

Figure 5.

A comparison of the arterial input function (AIF) shapes when using the AIF selection (red dot) from computed tomography perfusion plus (CTP+) (top left) and Vitrea (top right). The CTP+ AIF location is slightly inferior, leading to it being located on top of a feeding artery. This location leads to the curve having a smoother shape (bottom left) instead of multiple contrast peaks as seen in the Vitrea AIF (bottom right).

Figure 6.

The venous output function (VOF) selection locations (blue dot) for computed tomography perfusion plus (CTP+) (top left) and Vitrea (top right) along with the corresponding VOF curve for each software. When using Vitrea, the intensity values of the VOF decrease, likely due to partial volume effect from surrounding tissue, and there are multiple intensity peaks compared with CTP+. Multiple peaks and a drop in intensity from partial volume effects will hinder the calculation of perfusion parameters using Vitrea.

Infarct volume analysis

Table 4 indicates mean differences, MADs, and SCCs between DWI FIVs and CTP predicted infarct for each software along with percentage of patients who had infarct overestimated for the intervention and conservative treatment category. Table 4 also includes penumbra-to-infarct ratios for each CTP software for the intervention cohort. Table 5 includes similar metrics for intervention cohort subgroup analysis based on reperfusion status. No significant difference (p > 0.05) was seen between mean infarct differences for the four CTP softwares in the intervention cohort based on ANOVA analysis. Student-t test analysis indicated no difference between the CTP+ and Vitrea mean infarct differences (p = 0.72). Within the conservative treatment category, significant difference was evident (p < 0.05) between all CTP software for mean infarct differences based on ANOVA analysis. A p = 0.007 indicates similar results between CTP+ and Vitrea infarct differences for the conservative treatment cohort. Ninety-five percent confidence intervals for each CTP software processing times are: CTP+ = 192.6 ± 3.4 s, RAPID = 831.0 ± 34.1 s, Sphere = 146.1 ± 1.1 s, and Vitrea = 242.5 ± 3.8 s. ANOVA (p < 0.05) indicates significant differences between the processing times of the software.

Table 4.

Mean differences (with 95% confidence intervals), MADs, SCCs, and percentage of patients with infarct overestimation for calculated infarct between each CTP software and DWI.

Intervention patients
	Mean difference, mL	MAD, mL		SCC		Patients with infarct overestimation		Penumbra-to-infarct ratio
CTP+	5.8 ± 5.9	7.6		0.62		35.8% (29/81)		13.1 ± 9.5
RAPID	10.0 ± 5.2	6.9		0.73		21.0% (17/81)		45.8 ± 15.7
Sphere	3.0 ± 6.0	8.1		0.56		58.0% (47/81)		10.3 ± 6.5
Vitrea	7.2 ± 4.9	7.0		0.66		29.6% (24/81)		9.2 ± 7.1
Conservative treatment patients
	Mean difference, mL		MAD, mL		SCC		Patients with infarct overestimation
CTP+	–8.0 ± 5.4		5.1		0.64		67.3% (74/110)
RAPID	–25.6 ± 11.5		7.2		0.60		55.5% (61/110)
Sphere	–25.6 ± 8.0		12.5		0.66		90.1% (100/110)
Vitrea	1.3 ± 4.0		3.3		0.72		51.8% (57/110)

MAD: median absolute deviation; SCC: Spearman correlation coefficient; CTP: computed tomography perfusion; DWI: diffusion-weighted imaging; CTP+: computed tomography perfusion plus.

Table 5.

Mean differences (with 95% confidence intervals), MADs, SCCs, and percentage of patients with infarct overestimation for calculated infarct between each CTP software and DWI based on reperfusion status.

	Mean difference, mL	MAD, mL	SCC	Patients with infarct overestimation	Penumbra-to-infarct ratio
TICI 3
CTP+	6.9 ± 6.9	5.2	0.69	31.0% (9/29)	25.6 ± 26.7
RAPID	–8.4 ± 10.8	8.6	0.75	3.4% (1/29)	47.1 ± 22.0
Sphere	10.5 ± 5.0	9.2	0.57	48.3% (14/29)	5.3 ± 1.7
Vitrea	2.9 ± 10.7	5.6	0.76	27.6% (8/29)	5.4 ± 4.4
TICI 2c
CTP+	6.3 ± 9.2	6.4	0.43	45.5% (10/22)	5.5 ± 4.2
RAPID	3.9 ± 9.9	10.4	0.65	36.4% (8/22)	66.6 ± 48.0
Sphere	–1.3 ± 10.3	16.0	0.45	68.2% (15/22)	7.5 ± 4.8
Vitrea	7.7 ± 9.1	11.3	0.46	40.9% (9/22)	17.3 ± 25.3
TICI 2b
CTP+	4.5 ± 13.8	10.3	0.74	33.3% (10/30)	6.5 ± 2.8
RAPID	11.1 ± 10.6	6.1	0.80	26.7% (8/30)	29.3 ± 14.6
Sphere	5.1 ± 13.1	4.8	0.60	60.0% (18/30)	17.1 ± 17.6
Vitrea	5.5 ± 10.5	7.2	0.74	23.3% (7/30)	6.8 ± 6.6

MAD: median absolute deviation; SCC: Spearman correlation coefficient; CTP: computed tomography perfusion; DWI: diffusion-weighted imaging; TICI: thrombolysis in cerebral infarction; CTP+: computed tomography perfusion plus.

Figure 7 represents mountain plots showing the distribution of infarct differences for each CTP software with DWI for the intervention cohort. Dashed horizontal lines and vertical solid lines indicate 90th percentile of volume differences and ideal median volume difference of 0 ml, respectively. Figure 8 represents mountain plots for the conservative treatment cohort showing distributions of mean infarct differences between CTP software and DWI. Figure 9 shows scatter plots with regression analysis and SCCs for intervention patients comparing CTP software predicted infarct with DWI infarct. Figure 10 shows similar analysis for the conservative treatment cohort. Figures 11 and 12 represent Bland–Altman plots comparing DWI FIVs and CTP predicted infarct volumes for the intervention and conservative treatment cohorts respectively. Dashed lines represent limits of agreement and solid lines represent mean infarct differences.

Figure 7.

The distribution of mean infarct differences between diffusion-weighted imaging (DWI) with (a) computed tomography perfusion plus (CTP+), (b) RAPID, (c) Sphere, and (d) Vitrea for endovascular intervention patients. Mountain peaks indicate the median infarct difference. Dashed lines represent the 90th percentile of volume differences with more narrow distributions centered around 0 ml being ideal.

Figure 8.

The distribution of mean infarct differences for conservative treatment patients between final diffusion-weighted imaging (DWI) infarct with predicted (a) computed tomography perfusion plus (CTP+), (b) RAPID, (c) Sphere, and (d) Vitrea infarct. Dashed lines represent the 90th percentile of volume differences. Note the extended tail into the negative volume difference region for RAPID and Sphere, indicating overestimation of penumbra.

Figure 9.

The correlation between final diffusion-weighed imaging (DWI) infarct with predicted (a) computed tomography perfusion plus (CTP+), (b) RAPID, (c) Sphere, and (d) Vitrea infarct for endovascular intervention patients. Spearman correlation coefficients (SCCs) and linear regression equations are represented in each plot.

Figure 10.

The correlation for conservative treatment patients between final diffusion-weighted imaging (DWI) infarct with predicted infarct volumes for (a) computed tomography perfusion plus (CTP+), (b) RAPID, (c) Sphere, and (d) Vitrea software. Spearman correlation coefficients (SCCs) and linear regression equations are represented in each plot.

Figure 11.

Bland–Altman plots for intervention patients with the x-axes demonstrating the mean infarct volume between each computed tomography perfusion (CTP) predicted infarct and final diffusion-weighted imaging (DWI) infarct. The y-axes demonstrate differences between final DWI and predicted infarct for each CTP software. Solid lines represent the mean infarct difference and dashed lines represent limits of agreement (1.96·standard deviation).

Figure 12.

Bland–Altman plots for conservative treatment patients with the x-axes demonstrating mean infarct volume between final diffusion-weighted imaging (DWI) infarct and each computed tomography perfusion (CTP) predicted infarct. The y-axes demonstrate differences between final DWI and predicted infarct for each CTP software. Solid lines represent the mean infarct difference and dashed lines represent limits of agreement (1.96·standard deviation).

Inter-reader variability analysis

Inter-reader variability analysis of segmented infarct between five Vitrea users indicated a mean infarct difference of 3.6 ml and MAD of 2.7 ml. ANOVA analysis demonstrated no significant difference (p = 0.97) between user segmented infarct.

Discussion

This study provided an assessment of the newly developed CTP+ software in assessing infarct and penumbra segmentation compared with DWI FIVs. Additionally, CTP+ was compared with ischemic tissue segmentation from commercially available RAPID and Sphere software along with Vitrea software to investigate improvements made from CTP+’s predecessor. Precise segmentation of penumbra and infarct by these software products is essential in determining endovascular intervention eligibility. Although exact perfusion thresholds for determining ischemic tissue are patient specific, software that most closely predicts infarct volumes has resulted in the best clinical outcomes.¹⁷ This involves the ability to identify AIS patients least likely to sustain reperfusion injury and who will demonstrate clinical recovery. Since it is preferred that infarct be underestimated to prevent patients from being incorrectly excluded from reperfusion procedures, CTP+ must do the same for clinical consideration.

Previous studies have indicated that accurate AIF selection is essential in accurately quantifying CTP parameters such as CBF, CBV, and MTT. This indicates that infarct quantification can be drastically affected if the AIF is not placed properly, especially if it is place distal to the clot on the ipsilateral side.¹⁸ Furthermore, penumbral flow has also been known to be affected by the placement of the AIF.¹⁹ Within the CTP+ software, improved selection of the AIF occurs by placing the AIF in the internal carotid artery or M1 segment of the middle cerebral artery using hemodynamic constraints.¹⁸ The improved selection for CTP+ compared with Vitrea can be seen with slight improvements in the number of non-noisy AIFs and VOFs along with an increase in the number of properly shaped curves. Within Figure 4, it can be seen that within Vitrea there were instances in which the VOF section was nowhere near a draining vein, which leads to a VOF with no computational value. This erroneously selected VOF greatly hinders the calculation of perfusion parameters since the AIF can no longer be normalized properly. Instead of scaling the AIF up in intensity, a noisy VOF will lead to a downscaling of the AIF, which results in an AIF even further from the true intensity values.⁹

Furthermore, the AUC values for both the AIFs and VOFs are larger when using CTP+ compared with Vitrea. This likely indicates that a larger portion of the voxels selected for the input function contain a feeding artery or draining vein in CTP+ compared with Vitrea. Since a larger portion of the selection voxels contain this artery or vein, there is less partial volume effect from surrounding tissue, which would cause a decrease in the contrast intensity, further leading to a lower AUC value. The VOF selection location for Vitrea in Figure 6 reinforces that a lower AUC value is likely due to partial volume effect since the selection location is significantly higher in the draining vein compared with CTP+ and at this high a location, the vein will have significantly decreased in size. This partial volume effect in the VOF is of great importance as it results in a lower median AIF scaling factor for Vitrea than CTP+. In fact, a median AIF scaling factor of less than 1 for Vitrea indicates the AUC of the VOF is typically less than that of the AUC of the AIF. This is counterintuitive since the draining vein is typically larger in size than the feeding artery, leading to a higher degree of contrast flow, based on partial volume effects, which is why the VOF is used to normalize the AIF. An incorrect scaling of the AIF can further lead to incorrect parameter estimation, and in the event of downscaling of the AIF, overestimation of the CBF and CBV parameters.⁹

Intervention patient infarct volume analysis indicates CTP+ performs similarly with commercially available RAPID and Sphere software. Additionally, the fully automated CTP+ performed similarly to its semi-automated predecessor software, Vitrea, with CTP+ having a lower mean infarct difference with DWI FIVs. Although RAPID software provided the least number of patients with infarct overestimation, which is preferred so that patients are not incorrectly excluded from reperfusion procedures, CTP+ provided fewer overestimations than Sphere and a similar number to semi-automated Vitrea.¹¹ The improved ventricle segmentation likely aided in similar overestimations of infarct between Vitrea and CTP+ since without segmentation, erroneous infarcts would appear in the ventricles, as seen in Figure 1. Subgroup analysis based on reperfusion status indicate a similar trend between CTP+ and the current available software along with consistency in the mean infarct difference between DWI and CTP+ for all reperfusion statuses. Furthermore, Figure 7 indicates CTP+ to have near identical infarct difference distribution compared with the other investigated software at the 90th percentile. This similar distribution, along with the peak of the mountain plot (50% percentile) residing near 0 ml, indicates that the results from CTP+ are just as accurate and reproducible as from RAPID, Sphere, and Vitrea.

Conservative treatment patient results show a consistent overestimation of FIVs for commercially available RAPID and Sphere, similar to previous studies.⁷,⁸,²⁰ CTP+ additionally shows an overestimation of FIV for this patient group but by less than one-third of the magnitude seen by commercially available software. Figure 3 aids in indicating that overestimation of penumbra is the reason for FIVs being overestimated in conservative treatment patients. Since the mountain plots were fairly symmetrical for RAPID and Sphere in the intervention cohort, but there is an extended negative volume tail in these respective plots for conservative treatment patients, one can determine that infarct volume assessments are correct while penumbra assessments are inaccurate. Compared with RAPID and Sphere, CTP+ has much less penumbra overestimation, as indicated by the narrow mean difference distribution from DWI. Although the semi-automated Vitrea software still has the most accurate assessment of penumbra, CTP+ does provide a similar mean difference distribution and on a fully automated platform, which aids in diagnosis workflow. Processing times indicate that CTP+ can save clinicians approximately 1 min compared with manually segmenting the ischemic lesion in Vitrea. Additionally, CTP+ being available onsite allows for a more rapid assessment of the perfusion scans compared with RAPID, where transfer speeds to send the scans offsite can significantly hinder workflow.

Although the assumption that all penumbra converts to infarct allows for assessment of penumbra in conservative treatment patients, it should be noted that there is no independent way to accurately assess penumbra volumes. Alternative CTP software cannot be utilized as a penumbra validation method since it is based solely on contralateral hemisphere thresholds and not patient specific hemodynamics and characteristics. Since penumbra is incorrectly functioning neurological tissue, but is segmented on CTP using hemodynamics, patient tissue with naturally low flow velocities in one hemisphere can appear as penumbra even if the tissue is functioning correctly.¹⁰,²¹ Furthermore, DWI will not represent all penumbra as infarct in cases of spontaneous recanalization which occurs from vessel flow conditions dislodging the blood clot and restoring blood flow. Since the latter has only been shown to occur in 5% to 30% of cases, it was utilized in this study but still could have impacted the results.^22–24 However, any impact from spontaneous recanalization would have occurred in all software since the same patients were used, eliminating any potential bias.

Further limitations of the study include the potential for DWI lesion reversal in the event of rapid reperfusion.²⁵ This indicates that some of the DWI lesion may represent penumbra, which could lead to an overestimation of infarct for some of the intervention cases. Additionally, infarct growth will occur between initial image and either reperfusion or follow-up imaging, meaning predicted CTP volumes cannot perfectly represent DWI FIVs. Inclusion of only 21 patients with lesion volumes over 50 ml limits the study since the CTP+ software is not being extensively tested on a large infarct volume dataset. The comparison of CTP+ with only two commercially available software products along with its predecessor, Vitrea, is another limitation as there are other commercially available software products such as syngo.via (Siemens Healthcare, Erlangen, Germany and IntelliSpace Portal (Philips Corporation, Amsterdam, Noord Holland, Netherlands). Furthermore, AIFs and VOFs were compared only between CTP+ and Vitrea and not with Sphere or RAPID as raw data for their respective curves were not available. Additionally, spatial accuracy of infarct and penumbra was not assessed and should be considered for future studies. Last, validation cases from an outside institution were not included to assess CTP+ software.

Conclusions

Fully automated CTP+ software performed similarly to commercially available RAPID and Sphere CTP software in addition to its semi-automated predecessor software, Vitrea, when assessing infarct volumes for intervention patients. For conservative treatment patients, CTP+ outperformed both RAPID and Sphere in assessing penumbra volumes. Semi-automated Vitrea software remains the most accurate in assessing penumbra, but CTP+ provides an improved workflow from its predecessor.