Use of a convolutional neural network to identify infarct core using computed tomography perfusion parameters

1. INTRODUCTION

Within the United States, it was estimated that 7 million individuals over the age of 20 self-reported having a stroke within a 1-year span [1]. Of those who suffered from a stroke, approximately 87% suffered from an ischemic stroke, while 13% suffered from a hemorrhagic stroke [2]. An ischemic stroke occurs when a vessel supplying the brain with oxygenated blood becomes blocked through an embolus or stenosis [3]. Computed tomography perfusion (CTP), diffusion-weighted imaging (DWI), and fluid-attenuation inversion recovery magnetic resonance imaging (MRI) are imaging modalities commonly used to diagnose ischemic strokes which attribute to 2.7 million deaths in the United States each year [1]. Obstruction of the vessel feeding the brain with oxygenated blood leads to the formation of infarct and penumbra tissue which represent tissue that is irreversibly damaged and tissue that is deficient in blood flow but can be salvaged through reperfusion techniques, respectively [4].

Two such reperfusion techniques that are utilized to remove emboli are intravenous thrombolysis and mechanical thrombectomy. Intravenous thrombolysis produces a chemical reaction with the embolus which causes the embolus to breakdown within the vessel. Once broken down enough, the embolus will dislodge from within the vessel, due to inherent vessel flow conditions, allowing for reperfusion of the vessel. Meanwhile, the mechanical thrombectomy reperfusion technique requires an endovascular clinician to feed a catheter from the femoral or radial artery into the brain where the embolus is lodged. A stent retriever is then pushed through the clot from the catheter and the retriever is expanded to the size of the vessel wall. Following expansion, the embolus is captured and the clinician is able to pull the clot out of the vessel [5, 6]. In order to determine which patients are eligible for such reperfusion procedures, it is necessary to determine the amount of cerebral tissue classified as infarct and penumbra in each patient [7].

CTP is commonly used to identify these types of ischemic tissue utilizing perfusion parameters such as cerebral blood flow (CBF), cerebral blood volume (CBV), time-to-peak (TTP), mean-transit-time (MTT), and delay time [8]. CBF and CBV are typically used to identify infarct tissue as CBF represent the volume of blood traveling through the capillaries per unit time per unit brain tissue (Figure 1) while CBV represents the volume of blood per 100 grams of brain tissue. Meanwhile, TTP, MTT, and delay time can be utilized to estimate penumbra and represent: time until peak enhancement of brain tissue, average time blood takes to travel through capillaries, and arrival time of contrast to tissue, respectively [9]. Each of these parameters can be calculated utilizing a time density curve (TDC). A TDC represents the change in contrast intensity at a given voxel in the perfusion volume over time (Figure 2). From this curve, TTP is calculated as the time it takes for the curve to reach its maximum value, MTT is the time corresponding to the full width at half maximum, and delay time is the time until contrast intensity is above 0. Meanwhile, CBV represents the area under the time density curve and CBF is calculated as CBV divided by MTT [8, 10]. Although CTP is currently used to diagnose cerebral infarct and penumbra tissue, there are many flaws associated with the current processes utilized to segment these ischemic tissue from the perfusion parameter maps generated using this modality.

Currently, the issue with using CTP to estimate these ischemic tissue is it is based on comparing perfusion parameters in contralateral hemispheres and setting thresholds for each parameter to identify tissue as healthy, infarct, or penumbra. Various studies have been conducted to determine the optimal thresholds for each perfusion parameter to detect ischemic tissue, but these thresholds have varied based on the CTP software utilized and patient baseline hemodynamics [7, 9]. CTP software such as RAPID (iSchemaView, Menlo Park, CA) and Sphere (Olea Medical, La Ciotat, France) emphasize a relative CBF of 30% and 25% of the contralateral hemisphere, respectively, to identify infarct. Vitrea CTP software meanwhile emphasizes a relative CBV of 38% of the contralateral hemisphere to identify infarct [11]. These thresholds and parameters utilized to identify infarct are different between software due to the algorithms used to calculate the perfusion parameters.

Singular value decomposition and Bayesian algorithms are just two of many algorithms utilized across all CTP software to generate perfusion parameter maps [12–14]. Since the Bayesian algorithm implements smoothing filters in its generation of the perfusion maps, it leads to the optimal threshold utilized for ischemic tissue segmentation being less than those used for singular value decomposition map generation. Additionally, baseline hemodynamics for each patient influence what thresholds should be utilized. Some patients may naturally have decreased flow in one hemisphere of the brain, and although they have been neurologically intact for their entire lives, the optimally set CTP thresholds could label perfectly functioning tissue as infarct or penumbra. This indicates that an alternative to thresholds may be necessary to accurately detect ischemic tissue in ischemic stroke patients.

Convolutional neural networks (CNN) have the potential to be the new automated method capable of identifying infarct tissue. CNNs utilize various convolutions to detect patterns and features in images which allow for pixel-by-pixel classification, or semantic segmentation [15]. Although the network can be considered as a “black-box,” it is capable of detecting more features of the inputs images than standard manual analysis. During this segmentation, the CNN determines which features of the image are most important and weights these more predominantly during the classification process. Combining this pattern detection method with CTP parameter maps may provide a method to segment infarct tissue that is no longer based solely on contralateral hemisphere thresholds.

In this study, we propose the utilization of 5 separately trained CNNs, using the 5 perfusion parameter maps provided by CTP, to automatically segment infarct tissue from two-dimensional perfusion map images. Additionally, we compared the spatial location of the segmented infarct with 48-hour follow-up DWI infarct in both two-dimensional and three-dimensional spaces. Three-dimensional volume comparison was conducted after infarct volume reconstruction from the CTP infarct predictions. Furthermore, we compared the predicted CTP infarct volumes with 48-hour follow-up DWI infarct to assess the ability of each CNN to accurately quantify infarct. Successful infarct segmentation using this method could provide a novel method to improve clinical decision making regarding patient eligibility for reperfusion procedures without the use of contralateral hemisphere perfusion parameter thresholds.

2. METHODS

Collection of CTP data for 63 emergent large vessel occlusion acute ischemic stroke patients was conducted retrospectively following Institutional Review Board approval and Health Insurance and Portability and Accountability Act requirements. All patients were required to have undergone baseline CTP imaging along with 48-hour follow-up DWI. Patients were required to have undergone successful reperfusion classified as thrombolysis in cerebral infarction of 2b, 2c, or 3, or patients were required to have minimal penumbra present on baseline CTP imaging in addition to symptom onset time occurring 48 hours prior to initial CTP imaging. The combination of minimal penumbra present on initial CTP imaging and symptom onset time greater than 48 hours indicate all penumbra has likely converted to infarct [16, 17]. CTP data was acquired using 2 2012 Aquilion ONE (Canon Medical Systems Corporation, Otawara, Japan) CT units using the Neuro ONE protocol. Nineteen volumes were obtained for each patient containing 320 slices with a thickness of 0.5 mm and an in-plane resolution of 0.42 mm. Iohexol was injected throughout the patient’s vasculature at a volume of 50 mL and an injection rate of 5 mL/second. Scan parameters utilized included a tube voltage of 80 kVp, CT does indices ranging from 15.3 to 44.3 mGy, and dose length products ranging from 244.5 to 709.8 mGycm. Scan acquisition time was 49.3 seconds and reconstructed CTP volumes were available for perfusion processing within 5 minutes from the start of the scan.

DWI data was collected for all 63 patients using a Vantage 1.5 Tesla MRI unit (Canon Medical Systems Corporation, Otawara, Japan) using the IsoDWI protocol. This protocol includes a repetition time of 8,700 ms, an echo time of 100 ms, a slice thickness of 5 mm, and an in-plane resolution of 0.81 mm. DWI was utilized for ground truth infarct labels since this modality is commonly used for estimating final infarct volumes [18]. Ground truth infarct labels were generated by segmenting infarct from DWI images based on a 162% difference in intensities across contralateral hemispheres as shown in a previous study [19]. Figure 3 indicates the segmentation of infarct from a DWI along with the corresponding CTP slice.

CTP volumes were registered with DWI using MATLAB ‘ s intensity and geometric based registration tool which has been utilized within previous studies [20–23]. Linear interpolation, affine transformation, and multimodal Mattes method similarity metric were utilized with a one-plus-one evolutionary optimizer. The multimodal Mattes similarity metric was created specifically to register images from different modalities through joint probability distribution calculation in each image [24]. Registration accuracy was assessed through expert user manual segmentation of the ventricles in DWI and CTP followed by dice coefficient calculation to determine the degree of ventricle overlap [25]. Figure 4 indicates the overlap of ventricles following registration of the CTP and DWI volumes.

CTP volumes were loaded into Vitrea 7.10 (Vital Images, Minnetonka, MN) software for CTP analysis and generation of CBF, CBV, TTP, MTT, and delay time perfusion volumes. Volumes for each parameter were exported as 320 two-dimensional binary images resulting in 20,160 total images for each parameter for the 63 patients. CTP images determined not to contain infarct based on 48-hour follow-up DWI were removed from the dataset resulting in a total of 8,352 images for each CTP parameter. All 8,352 images were preprocessed by dividing all intensity values in the image by the maximum intensity value in the image.

Our CNN was created using Keras, Google’s (Google LLC, Menlo Park, CA) python machine learning framework and a TensorFlow backend. A modified U-net architecture was used with input and output image sizes of 64 pixels by 64 pixels for two-dimensional image segmentation with 2 processes in the contraction path, 1 middle process, and 2 processes in the expansion path. Each contraction process contains 2 convolutional layers, 1 max pooling layer, and 1 dropout layer (30%). The middle process contains 2 convolutional layers, and each expansion process contains 1 up-sampling convolutional layer, 3 convolutional layers, and 1 dropout layer (30%). The final layer, containing 1 convolutional layer and 1 sigmoid activation function, carried out the segmentation. Figure 5 represents the network architecture utilized. To prevent overfitting, a 60:30:10 split for training:testing:validation was utilized, allocating 38 patient to training, 19 patients to testing, and 6 patients to validation. On average, each patient had 133 slices containing infarct, allocating approximately 5,038 slices to training, 2,519 slices to testing, and 795 slices to validation. Training was conducted 5 separate times using each CTP parameter. Training utilized an Adadelta optimizer, which automatically adjusts the learning rate during the training process, a batch size of 32, and early stopping after the binary cross entropy dice loss did not improve over 10 epochs. Monte Carlo cross validation was conducted by training and testing the network 20 times for each parameter, randomly allocating patients to the training, testing, and validation groups each time. Training and testing of the models was conducted on an NVIDIA (Nvidia Corporation, Santa Clara, CA) P2000 GPU.

Following network training, testing of the network for each CTP parameter was conducted on 2,519 slices (19 patients) containing infarct. Assessment of infarct segmentation accuracy on individual CTP slices was conducted using Dice coefficients and positive predictive value (PPV) metrics. Dice coefficients indicate the degree of spatial agreement between the CTP predicted infarct and final infarct from DWI. Meanwhile, PPVs represent the amount of CTP infarct contained within the DWI infarct. PPV is calculated as the number of true positive predicted infarct pixels from CTP divided by the summation of true positive and false positive infarct predicted pixels from CTP. Sensitivity, specificity, accuracy, and negative predictive values were additionally calculated to access infarct volume segmentation accuracy. Sensitivity represents the amount of true positive predicted infarct pixels from CTP divided by the number of infarct pixels in DWI. Specificity indicates the number of true negative infarct (healthy tissue or background) pixels predicted from CTP divided by the number of healthy tissue or background pixels from DWI. Meanwhile, accuracy represents the summation of true positive and true negative predicted infarct pixels from CTP divided by the total number of pixels in the image. Lastly, negative predictive value (NPV) is the number of true negative (healthy tissue or background) pixels from CTP divided by the summation of true negative and false negative infarct pixels from CTP. Average predicted infarct volumes from each CTP parameter’s two-dimensional slice (each slice has a thickness of 0.5 mm) were compared with ground truth infarct volumes from corresponding DWI. This was done through calculation of mean infarct differences between each CTP parameter and DWI along with 95% confidence intervals.

Volumes of predicted and final infarct were then reconstructed for each of the testing patients from the slices in the testing set. Spatial overlap metrics (Dice coefficients, sensitivity, specificity, accuracy, PPV, and NPV) were determined for each of the 19 patient volumes within the testing set using each of the 5 CTP parameter with DWI. Additionally, average infarct volumes were calculated from the CBF, CBV, TTP, MTT, and delay time CTP infarct predictions along with the average infarct volume from DWI. The predicted CTP volumes were compared with DWI infarct volumes through mean infarct difference calculations along with 95% confidence intervals.

Within the CTP predicted infarct volumes, visual inspection indicated small erroneous infarct lesions being predicted near the base of the perfusion maps and in the contralateral hemisphere. In an attempt to correct for this, a post-processing watershed technique was utilized. This technique isolated all lesions within each reconstructed predicted infarct volume by labeling each lesion with a unique pixel identifier. All lesions but the largest were then removed from the volume as the smaller lesions were likely erroneous segmentations of infarct tissue. Dice coefficients, sensitivity, specificity, accuracy, PPV and NPV metrics were calculated between the watershed corrected reconstructed predicted infarct volumes and the ground truth DWI infarct volumes. Infarct volume measurements from DWI and the watershed CTP predictions were calculated and average differences were determined. Additionally, spatial metrics and volume measurements were compared for the standard CNN and the watershed corrected CNN to assess the potentially improvement of implementing this watershed method.

3. RESULTS

Table 1 indicates 95% confidence intervals for the two-dimensional spatial overlap metrics between CTP predicted infarct, for each of the five perfusion parameters, with 48-hour follow-up DWI infarct. These confidence intervals were determined using Monte Carlo cross-validation method over a total of 20 iterations. Figure 6 indicates the CBF perfusion slice that was fed into the network (a), the ground truth infarct label determined from DWI (b), and the predicted infarct output from our CNN using the CBF perfusion map (c). Figure 7 represents the same slice from the same patient utilizing the CBV perfusion map as the CNN input (a) as opposed to the CBF perfusion map with (b) and (c) again representing the ground truth infarct from DWI and infarct prediction from CTP respectively.

Table 2 represents mean predicted infarct volumes per CTP slice (slices have a thickness of 0.5 mm), with 95% confidence intervals, for each of the tested CTP parameters. Additionally, mean infarct volume from DWI is included in the table along with the mean infarct difference per slice calculated between DWI and the CTP parameter. Positive infarct differences indicate and underestimation of infarct by the CNN while negative infarct differences indicate an overestimation of infarct by the CNN. Figure 8 indicates an instance of failed infarct segmentation by the CNN with the input CTP parameter in the figure being CBF.

Table 3 shows the amount of spatial overlap between the reconstructed, CNN predicted, infarct volume from each CTP parameter with the final infarct volume from DWI. Ninety-five percent confidence intervals are additionally included for each parameter in the table and were determined using Monte Carlo cross-validation. Figure 9 represents the reconstructed infarct volumes for each CTP parameter along with the reconstructed final infarct volume from DWI. In the figure, blue regions represent DWI infarct, green regions represent CTP infarct predicted by the CNN, and maroon regions indicate overlap of infarct from CTP and DWI.

Table 4 indicates the mean reconstructed infarct volume in each patient, with 95% confidence intervals, for each CTP parameter and for DWI. In addition, Table 4 indicates the mean infarct difference between each of the predicted infarct from each CTP parameter with DWI infarct.

Table 5 represents the degree of spatial overlap between the volume reconstruction of infarct using each CTP parameter, following watershed volume correction, with final infarct from DWI. Ninety-five percent confidence intervals are included with the table. Figure 10 indicates the watershed volume correction for each of the CTP parameters. In the images, blue represents DWI infarct, green indicates CTP predicted infarct, and marron is an overlap of the two infarcts.

Table 6 shows the mean infarct volumes predicted for each CTP parameter using our CNN following watershed volume correction. Additionally, Table 6 shows the average difference between the indicated DWI final infarct volume with each CNN CTP predicted infarct volume following the watershed correction. Ninety-five percent confidence intervals are included for each metric in the table.

4. DISCUSSION

Analysis of the spatial overlap metrics in Table 1 indicate CBF is the most accurate CTP parameter in predicting infarct for two-dimensional slices. This is indicative based on the highest Dice coefficient, sensitivity, and PPV all corresponding to the CBF parameter. Figure 6 additionally indicates this, as the majority of the predicted CBF infarct lesion overlaps with the final infarct lesion from DWI. TTP and MTT meanwhile are the next two more accurate predictors of infarct based on the spatial overlap metrics in Table 1. The CBV parameter meanwhile indicates much lower spatial overlap metrics which can be explained by Figure 7. This figure shows a common trend with the CBV parameter predicting infarct volumes in both hemispheres of the brain when only one hemisphere actually contains infarct. Within Table 1, it should be noted that PPVs increase for each parameter compared to their corresponding Dice coefficients since infarct growth will occur between initial perfusion imaging and follow-up DWI imaging, or until a successful reperfusion procedure occurs. Due to this infarct growth between baseline and follow-up imaging, it is impossible to obtain a perfect Dice coefficient [7]. Therefore, PPV may be a better indicator of how accurate segmentation can be using CTP parameters since PPV indicates the percentage of the predicted CTP infarct lesion that is encompassed within the final DWI infarct lesion [11, 25]. This indicates the PPV metric is not reliant on the infarct growth that occurs between initial and final imaging.

Further analysis of two-dimensional predicted CTP infarct with DWI infarct indicates that CBF and CBV are better predictors of infarct compared with the other parameters based on volume measurements. Within Table 2, it can be seen that only the CBF and CBV CTP parameters underestimate the amount of infarct present in each slice, on average. This underestimation of infarct is preferred in clinical practice as it allows for in increased chance in enrollment for endovascular reperfusion procedures. From a clinical perspective, many clinicians would prefer to attempt the reperfusion procedure and have the patient regain their lost neurological function than eliminate them from consideration due to infarct overestimation [11]. Although this underestimation of infarct by CBF and CBV is preferred, there are certain instances where underestimation through a failed segmentation of infarct can occur and be harmful. Figure 8 indicates an instance of this failed segmentation which could lead to a patient receiving no treatment if they are deemed to have healthy cerebral tissue. This failed segmentation is likely due to fewer instances of decreased posterior circulation causing infarct compared to the majority of instances where infarct occurs in the middle cerebral artery territory. Based on the combined spatial and volumetric analysis, it can be deemed that CBF is the most accurate CTP parameter in determining infarct from two-dimensional slices.

Following reconstruction of the tested infarct volumes for each patient, it was again observed that the CBF parameter was the most accurate in segmenting infarct tissue based on the spatial overlap metrics in Table 3 and the volumetric comparison metrics in Table 4. For all perfusion parameters, there was an increase in the Dice coefficient, sensitivity, and PPV spatial overlap metrics compared to when slices were assessed. The increase in these metrics for reconstructed volumes likely indicates most failed segmentations were in slices that contained small amounts of infarct. Since these slices with smaller amount of infarct no longer carry the same weight as those with larger amounts of infarct, the segmentation metrics will intuitively increase. Many of these failed segmentations with smaller infarct tend to occur near the top of the skull where there is naturally less infarct that can be present and also where the calculation of perfusion parameters tends to be more erroneous due to less distal blood flow compared to other regions. Volumetric assessment of infarct differences between CTP and DWI again indicate CBF and CBV as the only parameters to underestimate infarct making them more clinically practical. Although CBV does underestimate the amount of infarct present, an underestimation of this magnitude could lead to the potential risk of hemorrhage in patients. Additionally, based on Figure 9, it can be seen that the CBV parameter is actually predicting erroneous lesions of infarct in the contralateral hemisphere compared to where the actual infarct is. TTP, MTT, and CBF additionally do this, but to a much smaller degree. For the particular reason of erroneous lesions being predicted, a watershed method was implemented to remove these incorrectly segmented infarct regions.

From the images indicated within Figure 10, it can be seen that the watershed method eliminates the erroneously segmented infarct, especially those that was incorrectly predicted within the contralateral hemisphere. Dice coefficient spatial overlap metrics remained relatively constant when utilizing the watershed method with a slight increase for the TTP parameter. PPV however increased for all CTP parameters after using the watershed method following our CNN. This indicates that there is a greater percentage of the predicted CTP lesion within the DWI lesion after using this method. The tradeoff with increasing the PPV however is the sensitivity slightly decreased for most CTP parameters. This is likely due to a slight decrease in the number of false negative infarct voxel which potentially occur when there is an error in segmentation in the middle of a large infarct lesion. Specificity, accuracy, and NPV metrics meanwhile all remained near 1 for all methods since the number of infarct negative voxels greatly outnumbers the amount of infarct positive voxels. This then leads to a much easier classification of these voxels for the network.

Volume measurements using the watershed method however led to larger underestimations of infarct using the CBF and CBV maps while the MTT and delay time maps have less of an overestimation of infarct. Although the CBF volumes appear to be less accurate compared to the ground truth for this method, the spatial overlap metrics indicate otherwise. This is due to the erroneously segmented CTP infarct, that is not in the DWI lesion, inflating the volumetric results for the CTP predicted infarct. This inflation of infarct from errors in segmentation is clearly seen within Figure 9, especially for the CBV parameter.

Limitations of this study include the 48-hour delay between initial perfusion imaging and follow-up DWI. This delay in follow-up imaging allowed for the growth of infarct to occur making it impossible to achieve perfect segmentation accuracy using CTP parameters from initial imaging. The imperfect registration technique used to align the CTP and DWI volumes is another limitation of the study. A dice coefficient of 0.78±0.02 was calculated when determining ventricle overlap between the CTP and DWI volumes. This indicates there may not have been a perfectly accurate placement of infarct between the ground truth labels and CTP slices when training the model. Additionally, the method used to segment infarct from the DWIs could have negatively influenced the results if erroneous regions of infarct were segmented as ground truth labels. Furthermore, the use of only one CTP map at a time to train the network and predict infarct location is a limitation as multiple maps are sometimes taken into consideration for infarct segmentation with CTP software [9]. In addition, although the watershed method helped improve segmentation accuracy based on spatial overlap metrics, it eliminates the ability to detect bilateral infarctions. This is due to the inclusion of only the largest infarct lesion in the segmented volume. Although identification of bilateral infarcts is a limitation within some CTP software and is rather uncommon, it should still be taken into consideration [11]. A final limitation of the study is all cases came from one comprehensive stroke center and no validation was conducted on patients outside of this institution.

5. CONCLUSION

This study assessed the effectiveness of a CNN in segmenting cerebral infarct tissue utilizing CBF, CBV, TTP, MTT, and delay time CTP maps. It was determined that CBF is the most accurate CTP parameter in segmenting infarct tissue based on spatial overlap metrics with TTP showing the most accurate assessment of predicted infarct volume. Therefore, CNNs can be utilized in combination with various CTP parameter maps to spatially locate and volumetrically assess cerebral infarct tissue in acute ischemic stroke patients. This automated segmentation of infarct using a CNN has the potential to eliminate non-universal contralateral hemisphere comparison thresholds which are currently utilized within CTP software to segment infarct tissue.

Supplementary Material