Deep learning-based identification of acute ischemic core and deficit from non-contrast CT and CTA

Abstract

The accurate identification of irreversible infarction and salvageable tissue is important in planning the treatments for acute ischemic stroke (AIS) patients. Computed tomographic perfusion (CTP) can be used to evaluate the ischemic core and deficit, covering most of the territories of anterior circulation, but many community hospitals and primary stroke centers do not have the capability to perform CTP scan in emergency situation. This study aimed to identify AIS lesions from widely available non-contrast computed tomography (NCCT) and CT angiography (CTA) using deep learning. A total of 345AIS patients from our emergency department were included. A multi-scale 3D convolutional neural network (CNN) was used as the predictive model with inputs of NCCT, CTA, and CTA+ (8 s delay after CTA) images. An external cohort with 108 patients was included to further validate the generalization performance of the proposed model. Strong correlations with CTP-RAPID segmentations (r = 0.84 for core, r = 0.83 for deficit) were observed when NCCT, CTA, and CTA+ images were all used in the model. The diagnostic decisions according to DEFUSE3 showed high accuracy when using NCCT, CTA, and CTA+ (0.90±0.04), followed by the combination of NCCT and CTA (0.87±0.04), CTA-alone (0.76±0.06), and NCCT-alone (0.53±0.09).

Introduction

The fast and accurate identification of irreversible infarction and salvageable tissue plays a key role in planning the treatments for acute ischemic stroke (AIS) patients.^1,2 The identification of ischemic stroke lesions (i.e., core and deficit lesions) is essential for the assessment of salvageable brain tissue and prediction of clinical outcome following reperfusion.^3–5 The volumes of ischemic core and deficit are important image biomarkers that have been successfully applied to select the treatment option for AIS patients with prolonged last known to be well time (LTKT) in several randomized clinical trials, e.g., DAWN, ⁶ DEFUSE3, ⁷ and EXTEND. ⁸ Some studies have investigated the variations of CTP threshold selection according to the time to recanalization in stroke patients.^9–11 The parameter maps calculated from computed tomographic perfusion (CTP) images can be used to segment the ischemic core and deficit regions. According to the RAPID (iSchemaView, Stanford, CA) definition, a reduction in cerebral blood flow (CBF) to below 30% of normal brain tissue is associated with ischemic core,^12,13 while the prolonged time-to-maximum (Tmax) of the tissue residue function to more than 6 s is associated with deficit area.^14,15 Some other definitions, like Sphere (Olea Medical, La Ciotat, France) and Vitrea (Vital Images, Minnetonka, MN), are also commonly used for infarct and penumbra quantifications.^16,17

However, high-speed computed tomography (CT)is often required to perform CTP to cover most of the territories of anterior circulation. Limited by the equipment, the majority of community hospitals and primary stroke centers do not have the capability to perform CTP scan in emergency situations 24/7. ¹⁸ Currently, the most widely used routine emergency CT imaging protocol only includes non-contrast computed tomography (NCCT) and CT angiography (CTA). Therefore, AIS lesion identification from widely available NCCT and CTA is urgently needed.

Recently, several studies have reported the applications of deep learning for stroke management.^19–25 In particular, deep learning has been applied to predict the outcomes of stroke lesions using acute-phase imaging before treatment.^21–28 Qiu et al. ²⁶ proposed a U-net-based deep learning approach to quantify early infarction using NCCT scans in AIS patients. Sheth et al. ²⁷ developed a symmetry-sensitive model to determine the acute ischemic outcome based on CTA. Olli Öman et al. ²⁸ applied a 3D-CNN model to detect stroke infarct lesions and to evaluate the infarction size based on CTA.

Although deep learning has shown its value in various stroke managements, there are no studies focusing on the prediction of deficit regions from baseline images, which provides additional information to select the patients for endovascular therapy beyond 6 h from symptom onset according to DEFUSE3. ⁷ A robust and automatic model is needed to quantify the ischemic core and deficit volumes from widely available NCCT and CTA to aid in the diagnosis and evaluation of the extent of infarction. Additionally, there is an increasing use of multiphase CTA (mCTA) in recent clinical trials and many clinical centers.^29,30 The American Heart Association guidelines suggest that collateral assessment may be useful in selecting patients for mechanical thrombectomy, citing the use of mCTA collateral assessment for patient selection rather than single-phase CTA (sCTA) or CTP.^31,32 The improved vessel occlusion detection rate, higher interrater reliability, and relatively low costs of mCTA support its use in lieu of sCTA and CTP as the standard initial vessel imaging test for AIS patients. Besides, mCTA can be acquired with widely available, low-end 32-row CT scanners. This study aimed to a) investigate the feasibility of using a deep learning model to identify both ischemic core and deficit regions based on NCCT and CTA images with similar diagnostic accuracy to CTP; and b) find the predictive values of using a post-angiography CT image (CTA+) for the stroke lesion identification.

In this study, the deep learning model was an end-to-end 3D convolutional neural network (CNN) architecture. NCCT, CTA, and CTA+ images were treated as independent channels for the network. The CTA+ images were acquired after CTA with 8 s delays. Model validation was performed on internal (N = 345) and external test cohorts (N = 108).

Material and methods

Patient population

This retrospective study was conducted in accordance with the Helsinki declaration and was approved by the Institutional Review Board of Changhai Hospital and the Tianjin First Central Hospital, with a waiver of written informed consent. For the internal cohort, we included a consecutive cohort of AIS patients (N = 345; mean age, 67 ± 2 years; interquartile range, 59-73 years; 188 men) from our emergency department between September 2017 and June 2019. The vessel occlusion locations included M1 (52%), internal carotid artery (40%), and M2 (8%). Another consecutive cohort including 108 patients (mean age, 61 ± 4 years; interquartile range, 50-73 years; 71 men) from another medical center (Tianjin First Central Hospital) between October 2018 and April 2021, was included to further validate the generalization performance of the proposed model. The vessel occlusion locations included M1 (40%), internal carotid artery (42%), and M2 (18%). All patients underwent the same neuroimaging protocol, which consisted of NCCT, followed by CTP. Patient inclusion and exclusion criteria are shown in Supplemental Figure 1. The inclusion criteria is: (1) last known well time (LKWT) of between 6 to 16 h; (2) National Institutes of Health Stroke Scale (NIHSS) score > 10; (3) the occlusion locations included M1, internal carotid artery and M2; and (4) adequate image quality for analysis according to the degree of noise, vessel sharpness, and overall quality. The characteristics of patients are listed in Table 1.

Table 1.

Characteristics of patients.

Patient information	Internal cohort (N = 345)	External cohort (N = 108)
Age (years)	67 ± 2 (59–73)	61 ± 4 (50–73)
Male	188	71
Onset to CT time (min)	567 (389–812)	521 (396–626)
NIHSS	19 ± 2 (15–20)	18 ± 3 (15–19)
Patents with ASPECTS < 6	84	36
Occlusion locations
MCA-M1 segment	52%	40%
Internal carotid artery	40%	42%
MCA-M2 segment	8%	18%

NIHSS = National Institutes of Health Stroke Scale, ASPECTS = Alberta Stroke Program Early CT Score, MCA = middle cerebral artery.

Image acquisition

All patients were imaged using a 320-Row Brilliance iCT scanner (Philips Healthcare, Eindhoven, Netherlands). NCCT (120 kVp, 320 mA) and CTP (80 kVp, 190 mA) were conducted to cover the intracranial range. A perfusion jog method was used to cover the whole intracranial range with a field-of-view (FOV) of 30 × 30 × 16 cm³ for NCCT scan. Transversal reconstruction was performed with a slice thickness of 5.0 mm. The in-plane resolution was 0.5 × 0.5 mm². All patients underwent CTP directly after NCCT imaging using the same scanner, and the CTP data were of sufficient quality to perform automated quantification. The FOV and spatial resolution of CTP images were identical to NCCT. Iodinated contrast (50 mL) was injected at 5.0 mL/s, and 14 frames with time interval of 4 s were acquired. For the external cohort, the following parameters were used for NCCT, CTA, and CTA+ (with 8 s delay after CTA): 100 kVp, 350 mA, spatial resolution of 0.48 × 0.48 mm², FOV of 26 × 26 × 15 cm³, and slice thickness of 0.6 mm. CTA scan was autotriggered by the appearance of contrast material in a region of interest (ROI) manually placed in the ascending aorta. CTP scan was followed by CTA with a second injection of contrast agent. CTP scan (70 kVp, 180 mA) was initiated 3-5 s after the injection of 0.5 mL of an iodinated contrast agent with spatial resolution of 0.625 × 0.625 mm² and slice thickness of 1.2 mm. All data were anonymized and stored on a server running the Extensible Neuroimaging Archive Toolkit (XNAT, version 1.1.6).

Image processing

The CTP images were post-processed by RAPID (iSchemaView, Stanford, CA) software to calculate the ischemic core (relative CBF < 30%) and deficit (Tmax > 6 s) volumes. For the internal test cohort, three-dimensional arterial images were reconstructed by IntelliSpace Portal (Philips Healthcare, Eindhoven, Netherlands, v6) from CTP arterial phase and pre-contrast phase to identify the occlusion site.

During the training stage, the images from three different CTP phases were selected for analysis: (1) the phase before contrast agent injection (NCCT); (2) the arterial phase (CTA); and (3) the phase immediately after CTA (CTA+ image, which was obtained 8 s after the arterial phase). These phases offer similar contrast and are good mimics of authentic NCCT and multiphase CTA scans. During the test stage, the acquired NCCT and mCTA (CTA and CTA+) images were input into the network pre-trained using the CTP-derived images. FSL Brain Extraction Tool was used to remove the skull and non-brain tissue from all these images. The matrix sizes across the dataset were resampled to 256 × 256 × 32 before putting them into the CNN model. F-STROKE ³³ Software (NeuroBlem, Ltd. Co., version 1.0.7) was used to compute the parameter maps from the corresponding CTP images and generate Tmax > 6 s and rCBF < 30% labels used for the CNN model.

CNN model

A 3D Unet architecture was used for ischemic core/deficit predictions (Figure 1). The NCCT, CTA and CTA+ images were treated as independent channels for the network. The architecture contained a compression path and synthesis path. In the compression path, each layer contained two 3 × 3 × 3 convolutions followed by a rectified linear unit (ReLu) function. Batch normalization layer was added before each ReLu. A 2 × 2 × 2 max pooling layer with strides of two was attached to each convolutional layer. In the synthesis path, each layer consisted of an upconvolution of 2 × 2 × 2 by strides of two, followed by two 3 × 3 × 3 convolutions each followed by a ReLu function. A multi-scale feature extraction scheme was performed in the architecture. Concatenate paths from the layers of equal resolutions in the compression paths were added to the network to provide high-resolution features to the synthesis paths. In the last layer, a 1 × 1×1 convolution was implemented to reduce the number of output channels to the number of labels.

Figure 1.

Diagram of the data processing workflow.

The network output and reference labels were compared using softmax with weighted cross-entropy loss. A batch size of 50 was chosen for the training. An automatically segmented intracranial mask was used to limit the investigative volume. The output consisted of two labels, i.e., the infarct core and deficit. Five-fold cross validation was done for evaluation. For the internal cohort, the training and validation cohort included 259 patients (75%), while the test cohort included the remaining 86 patients (25%) during each cross validation. For the external test cohort, all images were used for testing. Image pre-processing was performed on a Linux (Ubantu 3.5) workstation with 32-core Intel-Xeon Gold-6130-CPU at 2.10 GHz processor. The network was implemented using the TensorFlow library in Python 3.7 and trained using the GPUs of NVIDIA Tesla V100 (4 cores, each with 32 GB memory). An Adam optimization was used based on the TensorFlow library. The training procedure took 12 h for CTA-only or NCCT-only networks and 14 h for NCCT and CTA/CTA+ networks.

Statistical analysis

The data are presented as mean ± standard deviation if normally distributed, while median and interquartile ranges were used if non-normally distributed. The analyses were performed using the MATLAB libraries (Mathworks, Natick).

To compare the performance of the proposed model in the identification of ischemic core and deficit, four different groups of combinations were tested: (1) inputs with NCCT, CTA, and CTA+; (2) inputs with NCCT and CTA; (3) input with CTA alone; and (4) input with NCCT alone. The four CNN prediction models were evaluated using the same test data.

The spatial overlaps between CNN predictions and Tmax > 6 s/rCBF < 30% labels were evaluated using the DICE (percentage of spatial overlap) coefficient. The agreement between the volume measurements derived from CTP and different CNN models was assessed by Pearson correlation. Bland-Altman plots were used to illustrate the differences in lesion volumes predicted by CNN versus RAPID calculated volume. Concordance correlation coefficient (r) for the two volumes was assessed. Kolmogorov–Smirnov tests were used to determine whether the data sets were well modeled by a normal distribution. We also evaluated the agreement on binary study eligibility (yes or no) for patients based on the DEFUSE3 criteria. ⁷ As such, the patients with deficit volume > 15 cm³, core volume < 70 cm³, and ratio of deficit volumes to core volumes > 1.8 were identified to be suitable for surgery. Receiver operating characteristic (ROC) analysis was performed to validate the diagnostic consistency between the CTP-based decisions and proposed model-based decisions. The area under the ROC curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated for model evaluations.

Results

Evaluation on internal cohort

Figure 2 and Figure 3 show the CNN-predicted core/deficit lesions and the corresponding CTP-based segmentations in two representative subjects, each containing three slices. Columns 1 and 2 show the CBF and Tmax maps calculated from CTP. Columns 3 shows the rCBF < 30% and Tmax > 6 s regions generated from CTP, which were considered as the labels. Columns 3 to 6 show the predictions from different prediction models corresponding to four groups of input combinations: (1) with NCCT, CTA, and CTA+; (2) with NCCT and CTA; (3) with CTA-alone; and (4) with NCCT-alone as inputs. With the usage of NCCT, CTA, and CTA+ images, the CNN model predicted both the core and deficit volumes with high accuracy vs CTP (core: 47.0 mL vs 48.6 mL; deficit: 166.6 mL vs 170.7 mL). In both cases, CNN predictions that used the NCCT, CTA and CTA+ images provided the best volume estimations compared to the other models. As shown in Table 2, it also demonstrated the highest spatial overlap of core and deficit predictions in terms of DICE coefficient (0.63 ± 0.04 and 0.71 ± 0.03, respectively). In contrast, the prediction from NCCT-only model did not align well with the reference image compared to the other three models, and its DICE coefficient was the lowest (0.38 ± 0.04 for core and 0.51 ± 0.05 for deficit).

Figure 2.

Representative results from an internal test cohort using the CNN predictive models (patient #1). Four separate CNNs were trained: (1) with NCCT, CTA and CTA+; (2) with NCCT and CTA; (3) with CTA-alone; and (4) with NCCT-alone as inputs. The CTP based segmentations of core and deficit are shown in yellow and blue, respectively.

Figure 3.

Representative results from an internal test cohort using the CNN predictive models (patient #2). Four separate CNNs were trained: (1) with NCCT, CTA and CTA+; (2) with NCCT and CTA; (3) with CTA-alone; and (4) with NCCT-alone as inputs. The CTP based segmentations of core and deficit are shown in yellow and blue, respectively.

Table 2.

Statistical analysis results of the agreements on core/deficit volume calculations against CTP-RAPID definitions.

	Internal test cohort (N=345)				External test cohort (N=108)
	Core R	Deficit R	Core DICE	Deficit DICE	Core R	Deficit R	Core DICE	Deficit DICE
NCCT, CTA, CTA+	0.84	0.83	0.63 ± 0.04	0.71 ± 0.03	0.80	0.67	0.61 ± 0.04	0.68 ± 0.02
NCCT, CTA	0.71	0.83	0.45 ± 0.04	0.68 ± 0.03	0.78	0.67	0.57 ± 0.01	0.67 ± 0.02
CTA	0.70	0.72	0.43 ± 0.03	0.62 ± 0.04	0.59	0.59	0.58 ± 0.03	0.65 ± 0.03
NCCT	0.59	0.20	0.38 ± 0.04	0.51 ± 0.05	0.63	0.33	0.32 ± 0.02	0.47 ± 0.02

NCCT = non-contrast computed tomography, CTA = CT angiography, CTA+ = image acquired immediately after CTA with 8-s delay.

The overall ischemic core and deficit volume estimations are plotted against the RAPID results in Supplemental Figure 2(a-d) and Supplemental Figure 3(a-d). The highest correlations (r = 0.84 for core and r = 0.83 for deficit) were observed when NCCT, CTA and CTA+ images were all used for predictions, followed by NCCT+CTA (r = 0.71 for core, and r = 0.83 for deficit), CTA-alone (r = 0.70 for core, and r = 0.72 for deficit) and NCCT-alone (r = 0.59 for core, and r = 0.20 for deficit). The CTA-based model performed better than the NCCT-based model in terms of DICE coefficients in core (0.43 ± 0.03 vs 0.38 ± 0.04, respectively) and deficit (0.63 ± 0.04 vs 0.51 ± 0.05, respectively). Compared to the model using CTA images alone, certain improvements in DICE were observed for the identification of core and deficit volumes when NCCT image was included as an additional input channel (from 0.43 ± 0.03 to 0.45 ± 0.04 for core and from 0.62 ± 0.04 to 0.68 ± 0.03 for deficit), possibly because NCCT images offered useful baseline perfusion information for lesion identifications when combined with CTA.

Evaluation on external cohort

A representative case in the external test cohort is shown in Figure 4. Similar to the findings in the internal test cohort, CNN predictions that used NCCT, CTA, and CTA+ images as inputs provided the best core and deficit predictions based on the DICE coefficient (0.61 ± 0.04 for core and 0.68 ± 0.02 for deficit), followed by NCCT+CTA (0.57 ± 0.01 for core and 0.67 ± 0.02 for deficit), CTA-alone (0.58 ± 0.03 for core and 0.65 ± 0.03 for deficit), and NCCT-alone (0.32 ± 0.02 for core and 0.47 ± 0.02 for deficit). The spatial overlaps in core/deficit predictions were improved by adding CTA+ in the CNN model because the perfusion effect is more obvious with a certain delay in contrast bolus arrival time.

Figure 4.

A representative case from an external test cohort using the CNN predictive models. Four separate CNNs were trained: (1) with NCCT, CTA and CTA+; (2) with NCCT and CTA; (3) with CTA-alone; and (4) with NCCT-alone as inputs. The CTP based segmentations of core and deficit are shown in yellow and blue, respectively.

Similarly, the highest correlations were observed in CNN predictions when NCCT, CTA, and CTA+ images were all included (r = 0.80 for core and r = 0.67 for deficit), followed by NCCT+CTA (r = 0.78 for core and r = 0.67 for deficit), CTA-alone (r = 0.59 for core and r = 0.59 for deficit), and NCCT-alone (r = 0.63 for core and r = 0.33 for deficit) (Supplemental Figure 2(e-h) and Supplemental Figure 3(e-h)).

Evaluation of the diagnostic performance

The diagnostic performance of all models according to DEFUSE3 is shown in Table 3. As expected, in both the internal and external cohorts, the highest accuracy was obtained from the combination of NCCT, CTA, and CTA+ (0.90 ± 0.04 for internal cohort and 0.90 ± 0.03 for external cohort), followed by NCCT+ CTA (0.87 ± 0.04 for internal cohort and 0.90 ± 0.03 for external cohort), CTA-alone (0.76 ± 0.06 for internal cohort and 0.89 ± 0.03 for external cohort), and NCCT-alone (0.53 ± 0.09 for internal cohort and 0.83 ± 0.03 for external cohort) (Supplemental Figure 4). Bland-Altman plots are shown in Supplemental Figures 5-6.

Table 3.

Statistical analysis results of the agreements on binary study eligibility for patients based on DEFUSE3 criteria.

	Internal test cohort (N=345)						External test cohort (N=108)
	Sensitivity	Specificity	PPV	NPV	Accuracy	AUC	Sensitivity	Specificity	PPV	NPV	Accuracy	AUC
NCCT, CTA, CTA+	0.95±0.05 (0.86-1.00)	0.83±0.06 (0.75-0.93)	0.87±0.05 (0.78-0.95)	0.93±0.06 (0.83-1.00)	0.90±0.04 (0.82-0.94)	0.89±0.02 (0.85-0.92)	0.98±0.02 (0.95-1.00)	0.63±0.09 (0.46-0.74)	0.90±0.03 (0.83-0.94)	0.91±0.09 (0.76-0.99)	0.90±0.03 (0.86-0.95)	0.81±0.04 (0.73-0.90)
NCCT, CTA	0.96±0.03 (0.90-1.00)	0.77±0.08 (0.67-0.95)	0.84±0.07 (0.74-0.96)	0.94±0.04 (0.86-1.00)	0.87±0.04 (0.82-0.93)	0.86±0.02 (0.83-0.90)	0.96±0.02 (0.92-1.00)	0.67±0.11 (0.42-0.80)	0.91±0.03 (0.83-0.95)	0.84±0.01 (0.68-1.00)	0.90±0.03 (0.84-0.95)	0.81±0.04 (0.73-0.89)
CTA	0.93±0.05 (0.84-0.99)	0.56±0.08 (0.40-0.68)	0.71±0.07 (0.60-0.83)	0.88±0.08 (0.74-0.99)	0.76±0.06 (0.67-0.85)	0.74±0.03 (0.69-0.79)	0.99±0.01 (0.97-1.00)	0.62±0.12 (0.45-0.81)	0.88±0.04 (0.83-0.95)	0.95±0.06 (0.86-1.00)	0.89±0.03 (0.83-0.95)	0.80±0.04 (0.73-0.89)
NCCT	0.97±0.04 (0.89-1.00)	0.01±0.02 (0.00-0.05)	0.54±0.08 (0.38-0.68)	0.10±0.09 (0.02-0.36)	0.53±0.09 (0.36-0.68)	0.49±0.03 (0.41-0.56)	1.00±0.01 (0.99-1.00)	0.49±0.10 (0.33-0.69)	0.86±0.03 (0.80-0.90)	0.89±0.01 (0.87-0.92)	0.83±0.03 (0.77-0.87)	0.73±0.05 (0.64-0.82)

The data in parentheses are 95% confidence intervals.

NCCT = non-contrast computed tomography, CTA = CT angiography, CTA+ = image acquired immediately after CTA with 8-s delay, PPV = positive predictive value, NPV = negative predictive value, AUC = area under the curve.

Discussion

In this study, we used a CNN model to identify the infarct core and deficit using NCCT and/or multiphase CTA and compared it to CTP-RAPID segmentations. Quantitative evaluations showed a strong consistency with CTP-based definitions in terms of locational and volumetric agreement. Good agreements on the diagnostic decisions were obtained using the proposed deep learning method according to the DEFUSE3 trial criteria. Even though the CT doses and imaging parameters were different in the two test cohorts, the pre-trained model performed well in the external cohort, which proved the generalization capability of the proposed model. The results indicated the potential of using widely available imaging modalities (NCCT and CTA) for ischemic stroke evaluation using a deep learning approach. Moreover, this study observed a significant improvement in identifying core from deficit areas with the incorporation of CTA+ image acquired immediately after conventional CTA, which is also available for low-speed CT scanner. ³⁴

CTP-generated infarct core and deficit segmentations successfully assisted the selection of AIS patients for thrombolysis and thrombectomy in multiple randomized clinical trials. CTP depicts the perfusion process with higher temporal resolution and more phases than NCCT, CTA, and multiphase CTA protocols. Several studies^35–38 reported that manual assessment on NCCT and multiphase CTA images through collateral scores, ASPECTS, or manual infarct core segmentation had similar prognostic performance compared with using infarct core and mismatch analysis by CTP software. This indicated that severe data redundancy may exist in these extra phases of CTP, and the key phases of CTP data may already contain enough information for experienced neuroradiologists to recognize the infarct and deficit areas. Deep learning is a feasible method to extract those implicit features within key phase images and relax the data redundancy of high temporal resolution scans.

The results show that the model using NCCT alone achieved a relatively low correlation relationship with CTP for core segmentation (r = 0.59), which is lower than the previous report (r = 0.76) of Nagel et al., ³⁹ possibly because the LKWT of the patients in our study was longer than 6 h, while the LKWT of patients in e-ASPECTS study was within 6 h. NCCT alone provides poor deficit segmentations (r = 0.20). The predicted lesions from CTA-alone correlate well with CTP segmentations (r = 0.70 for core and r = 0.72 for deficit), indicating that CTA is more sensitive to early ischemic changes compared to NCCT, which is consistent with previous studies.^40,41 The combined usage of NCCT and CTA slightly improved the performance of core segmentation compared to the usage of CTA-alone (r = 0.71 vs r = 0.70, respectively), which is similar to the findings of Öman et al. ²⁸ Combining NCCT with CTA significantly improved the deficit segmentation because NCCT offered a baseline image for perfusion calculation, which increased the accuracy of deficit predictions. By incorporating CTA+ into the analysis, the performance on deficit segmentation increased slightly, while its performance on but the core segmentation largely improved. This was because the contrast bolus did not reach the deficit tissue due to the delayed transit time in the arterial phase to generate the contrast between the infarct core and deficit. However, with the 8-s delay in CTA+ phase, the deficits enhanced while the enhancement on infarcted region was limited compared to the arterial phase. Therefore, CTA+ can improve the identification of infarct core from the deficit. The overall agreement on diagnostic decisions was improved by incorporating CTA+ into the analysis according to the DEFUSE3 ⁷ trial criteria.

Previous studies have found a high correlation between CTA and DWI lesion volumes in acute stroke,^42,43 which is believed to be more accurate for infarct identification. In contrast, we trained our model against CTP-RAPID definitions. Our model can be extended to predict DWI-derived ischemic core, but the main limitation is the scarcity of labeled data. It is not a routine for stroke centers to perform NCCT, multiphase CTA, and MR-DWI scans in emergency situations protocol. Our method can also be used to predict the patient outcomes and identify the patients who would benefit the most from the treatment as long as the follow-up imaging is available.^44,45

This study has several limitations. First, we did not explore the effect of adding the symmetry information on our CNN model. Other studies have reported the usability of contralateral features and anatomical context for learning.^28,46 In that way, bilateral ischemic strokes need to be carefully investigated. Second, the images used in this study came from the CT scanners of the same vendor. Even though we have validated its generalization capability on a separate dataset, further studies with images from more centers and various acquisition parameters are still needed.

In conclusion, our experimental results verified that NCCT and CTA can be used to detect ischemic stroke lesions, with an accuracy similar to CTP-RAPID definitions with the introduction of deep learning. Strong consistency with CTP-based definition of ischemic core and deficit volumes was observed in the proposed model in terms of locational and volumetric agreements. These findings demonstrate the high potential of using deep learning to assist the clinicians in detecting of acute stroke lesions with widely available NCCT and CTA images.