Semi-automated infarct segmentation from follow-up noncontrast CT scans in patients with acute ischemic stroke

Transcription

1 Semi-automated infarct segmentation from follow-up noncontrast CT scans in patients with acute ischemic stroke Hulin Kuang, Bijoy K. Menon, and Wu Qiu a) Department of Clinical Neurosciences, University of Calgary, Calgary, Alberta, T2N 2T9 Canada (Received 13 December 2018; revised 30 May 2019; accepted for publication 28 June 2019; published 6 August 2019) Purpose: Cerebral infarct volume observed in follow-up noncontrast computed tomography (NCCT) scans of acute ischemic stroke (AIS) patients is as an important radiologic outcome measure of the effectiveness of endovascular therapy (EVT). In this paper, our aim is to propose a semiautomated segmentation approach that can accurately measure ischemic infarct volume from NCCT images of AIS patients. Methods: A novel cascaded random forest (RF) learning is first employed to classify each voxel into normal or ischemic voxel, leading to an infarct probability map. Four types of features: intensity, statistical information in local region, symmetric difference compared to the contralateral side, and spatial probability of infarct occurrence generated by the STAPLE method, are extracted. These features are input into RF to train a first-stage classifier. The coarse segmentation results generated by the first-stage classifier are then used to train a fine second-stage classifier with fivefold cross validation. The RF estimated infarct probability map obtained in the second-stage testing as well as user input high-level knowledge are then incorporated into a convex optimization function to obtain final segmentation. One hundred AIS patients were collected in this study, of which 70 patient images were used for evaluation while the remaining 30 patient images were used for RF training. Results: Quantitative results show that the proposed approach is capable of yielding a dice coefficient (DC) of 79.42%, significantly outperforming some state-of-the-art automated segmentation methods, such as the RF-based methods and convolutional neural network (CNN)-based segmentation method, U-net. The infarct volume obtained by the proposed method is strongly correlated with the manually segmented volume. In addition, interobserver variability analysis initialized by two observers suggests low user dependency. Conclusions: Our proposed semiautomated segmentation method can accurately segment infarct from NCCT of AIS patients American Association of Physicists in Medicine [ /mp.13703] acute ischemic stroke, convex optimization, infarct segmentation, noncontrast CT, ran- Key words: dom forest 1. INTRODUCTION Acute ischemic stroke (AIS) is the leading cause of severe disability in adults and the third leading cause of death worldwide. Endovascular therapy (EVT) has become the standard of care in patients with disabling anterior circulation ischemic stroke due to proximal intracranial thrombi. 1 Cerebral infarct volume observed on follow-up noncontrast computed tomography (NCCT) scans is an important radiologic outcome measure of the effectiveness of EVT. 2 Recent studies have been shown that clinical outcome after EVT could be correlated with a reduction in post-treatment infarct volume. 2 It is therefore possible that post-treatment infarct volume estimation would become a robust surrogate outcome measure in future clinical trials in AIS patients. Measurement of posttreatment infarct volume, however, relies on manual segmentation in practice, which is a tedious, time-consuming, and observer dependent method. An auto or semiautomated infarct segmentation approach with less human input would be beneficial for quantitatively measuring infarct volume from clinically used follow-up NCCT images. Segmenting ischemic infarct from NCCT images suffers from low signal to noise ratio (SNR), anisotropy of images, anatomical asymmetry, and interference from chronic infarcts, leukoaraiosis, and partial volume effect in the area around the cerebrospinal fluid (CSF). While many efforts have been made on segmenting lesions from MR images, 3 there are not many well-established methods for segmenting ischemic lesions from CT scans. Most existing methods segment ischemic lesions in an automatic way using machine learning. Joint features including mean, standard deviation, histogram, and gray level co-occurrence matrix, were used to automatically segment ischemic stroke region on CT scans. 4 Ischemic stroke lesions were segmented based on texture features and classification. 5,6 Nowinski et al. detected, localized, and quantified the stroke infarct by analyzing hemisphere 4037 Med. Phys. 46 (9), September /2019/46(9)/4037/ American Association of Physicists in Medicine 4037

2 4038 Kuang et al.: Semi-automated infarct segmentation 4038 attenuation value distributions using percentile difference ratios. 7 Tyan et al. proposed an ischemic stroke detection system including contrast enhancement, the brain tissue image area extraction, and an unsupervised region growing algorithm, and tested this method on 90 CT scan slices from 26 ischemic stroke patients. 8 Gillebert et al. proposed an automated delineation of stroke lesions via the accurate normalization of CT images from stroke patients into template space and the subsequent voxel wise comparison with a group of control CT images. 9 However, there was reduced sensitivity in regions close to the ventricles and the brain contours. These methods were reported to be able to generate decent accuracy. Most of them, however, were evaluated with limited data, and heavy manual refinement after automatic segmentation was typically required to remove the detected false positives caused by the aforementioned interference. In order to reduce the detected false positives, high-level human knowledge was introduced to guide the segmentation, leading to some semiautomated approaches for infarct segmentation. Boers et al. proposed a semiautomated intensitybased region-growing algorithm to segment cerebral infarct for volume measurement in follow-up NCCT Scans of AIS patients. 10 In this technique, manual placement of seed points and midline was required. This method was evaluated on 34 CT scans and obtained a dice coefficient (DC) of 74%. Another semiautomated lesion demarcation approach, 11 the clusterize algorithm with initialization by experts, was evaluated on multiple image modalities of acute stroke patients. It achieved a DC of about 84% with 13 CT scans. A joint segmentation approach based on random forest (RF) followed by multi-region contour evolution was proposed to segment ischemic infarct and hemorrhage simultaneously. 12 It only dealt with the cases where ischemic infarct and hemorrhage were confounded, and was evaluated on a limited data of 16 NCCT images. The relative small sample sizes of these studies limited their clinical applicability. In this study, a novel semiautomated ischemic infarct segmentation approach is proposed for follow-up NCCT scans of AIS patients, which integrates machine learning learned semantic information based on well-designed features and interactive segmentation concept into the same segmentation framework. The proposed approach consists of three major parts: expert initialization, RF classification, and convex optimization-based segmentation. Our contributions are in three aspects: (a) A two-stage RF training/testing strategy is employed to estimate infarct probability, which is following a coarse-to-fine training/testing mechanism. Although cascaded training has been proven to be an effective strategy for some computer vision problems, 13 it is unknown whether this strategy can be adapted on this particular medical imaging problem. (b) Expert knowledge and the RF estimated infarct probability are incorporated into an energy optimization function. The final segmentation is obtained by globally minimizing the defined energy function using convex optimization technique. (c) The proposed segmentation approach is extensively evaluated to demonstrate its efficacy, compared to manual segmentation and the state-of-the-art automatic segmentation methods. 2. INFARCT SEGMENTATION APPROACH 2.A. Approach overview The proposed semiautomated segmentation pipeline is demonstrated in Fig. 1, which includes a few steps: preprocessing, user initialization, RF learning, and convex optimization-based segmentation. 2.B. Preprocessing All the original NCCT images I(x), x 2 Ω (Ω denotes image space) are first skull-stripped, 14 and are then transformed to the Montreal Neurological Institute (MNI) space (the image size is mm 3 ) using affine registration. The images in MNI space are then smoothed by a nonlocal transform-domain filter method, and the smoothed images IðyÞ, y 2 X MNI (X MNI denotes MNI space) are accordingly used as input for the subsequent feature extraction and RF learning. During the testing stage, the RF estimated infarct probability map in the MNI space is transformed back to the image space Ω by inverting affine transformation matrix. The final convex optimization-based segmentation is performed to partition the image space Ω into two disjoint regions: X f for infarct region and X b for background, where X f þ X b ¼ X; X f \ X b ¼ : (1) 2.C. User initialization The proposed segmentation approach requires users to prelabel some voxels ^x, ^x 2 X. ^x denotes the labeled voxels, for infarct region and background, on a few axial slices of the input three-dimensional (3D) NCCT images, as illustrated in Fig. 2. The number of initialization slices depends on the size of the infarct of the patient. Three to five views were typically used for most of patient images. The intensities of the labeled voxels are used to estimate two prior intensity probability density functions (PDFs) k s;t ðxþ for the infarct and background, where s and t denote infarct and background respectively. The user initializations encode local context information, including infarct location and intensity profile. This information is of importance for segmenting the patient images with small infarcts and low contrast between infarct and background, which are quite challenging for most of automatic infarct segmentation methods. The obtained PDFs from prelabeled voxels ^x are used to assign each voxel a data cost to facilitate the subsequent optimization procedure. It should be noted that the userselected voxels ^x also serve as hard constraints supervising the optimization scheme.

3 4039 Kuang et al.: Semi-automated infarct segmentation 4039 FIG. 1. Framework of the proposed semiautomated infarct segmentation approach. FIG. 2. Initialization scheme. Red strokes are labeled as foreground voxels while green strokes are labelled as background voxels. [Color figure can be viewed at wileyonlinelibrary.com] 2.D. Random forest learning 2.D.1. Feature extraction We design four types of features, 9 features in total as follows: 1. Intensity information: Smoothed voxel intensity IðyÞ, y 2 X MNI, after the smoothing of raw NCCT images is used as a feature, denoted as IntFeat. In order to improve the differentiation of gray and white matter, the input smoothing image IðyÞ is enhanced by histogram equalization. The enhanced voxel intensity HðIðyÞÞ is used as the second feature, denoted as HistEqIntFeat. 2. Statistical information in the local region: For each voxel, the statistical context information in its local region (5 9 5 in this study) is taken into account, such as the mean (MeanIntFeat) and standard deviation (StdIntFeat) of image intensity of its neighboring voxels, defined as IðyÞ ¼ 1 NðyÞ R k2nðyþi k ðyþ and rðyþ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 NðyÞ R k2nðyþfi k ðyþ IðyÞg 2, where k = 1,..., N(y) is the number of neighboring voxels of y, and N(y) denotes all neighboring voxels. 3. Symmetric difference compared to the contralateral side: Since ischemic infarct has lower intensity values compared to the normal tissue in the contralateral side, symmetric abnormalities between the ischemic and contralateral side are typically used when radiologists manually contour infarct. This prior information perceived by human eyes should be coded into the algorithm, resulting in 2 features: IntDifFeat, dðyþ ¼IðyÞ Iðy Þ, and HistEqIntDifFeat, HðdðyÞÞ ¼ HðIðyÞ Iðy ÞÞ, where y denotes the contralateral and symmetrical voxel of the voxel y. We assume the brain is approximately symmetrical. Once an image is transformed into MNI space using affine registration, it is spatially aligned with the atlas in MNI space, which is symmetrical. The middle line of the transformed image is accordingly considered as symmetrical axis. However, the brains of AIS patients are typically not completely symmetric due to stroke pathology. In order to account for this quasi-symmetry, the differences of statistical information in local region are introduced, which tends to be more robust, defined as MeanIntDifFeat, dðyþ ¼IðyÞ Iðy Þ, and StdInt- DifFeat, rðdðyþþ ¼ rðiðyþ Iðy ÞÞ. 4. Spatial probability of infarct occurrence: An additional feature l(y) associated with infarct occurrence location is introduced in this work, which is obtained by statistically analyzing the infarct occurrence probability of each location using all the training images. Specifically, all the training images as well as their corresponding infarct labels are nonlinearly registered into MNI space using deformable registration in NiftyReg

4 4040 Kuang et al.: Semi-automated infarct segmentation 4040 FIG. 3. An examples of the extracted 9 features in an axial slice of one patient. Left to right, first row: NCCT slice with infarct manual segmentation, IntFeat, HistEqIntFeat, MeanIntFeat, and StdIntFeat; second row: IntDifFeat, HistEqIntDifFeat, MeanIntDifFeat, StdIntDifFeat, and LocProbFeat. toolbox. The STAPLE algorithm 15 is then used to fuse all transformed infarct labels to generate an infarct location probability map (LocProbFeat). An example of the generated feature maps of one patient image is shown in Fig. 3, visually demonstrating the effectiveness of all the used features for infarct segmentation. 2.D.2. Two-stage training and testing Cascaded training was employed in this application using the aforementioned features by cascading two RF classifiers, which adopts a coarse-to-fine mechanism. The strategy of holdout validation randomly subsampled was used in our experiments, as the leave-one-out training and testing is computationally expensive. During the training at the first stage, positive samples belonging to infarcts are randomly selected, which are matched with the same number of negative samples belonging to normal tissues. The features of both the selected positive and negative voxels from the randomly selected 30 training patients are input to RF to train the first-stage classifier. The objective of the first-stage classifier is to include as many candidate infarct voxels as possible. An optimal segmentation is not necessary at this stage. Conversely, a more accurate segmentation at the first stage may leave out some suspicious infarct voxels, which might be accurately classified by the secondstage classifier. During the training at the second stage, the positive samples are selected from the voxels obtained from the first-stage training, which are confirmed by manual segmentation. The same procedure is utilized to select negative samples for the second-stage training. In particular, we conduct fivefold cross validation to select the optimal training parameters for the second-stage training. An example of the infarct probability maps exported from the first-stage and second-stage training is shown in Fig. 4. Visual inspection shows that the final infarct probability map decreases false positive voxels substantially compared to the probability map generated from the first stage, which may justify the efficacy of the cascaded training mechanism. 2.E. Convex optimization-based segmentation A convex optimization-based segmentation approach is subsequently applied to optimize the infarct probability map p(x) generated from RF, prior intensity PDFs k s;t ðxþ, as well as user initialization information, to obtain final segmentation result. If u (x) 2 {0,1} is defined as the indicator function of the segmented infarct region associated with two cost functions C s ðxþ and C t ðxþ for infarct and background voxels, respectively, the final segmentation u(x) can be obtained by minimizing the following variational formulation associated with the data costs: Z min hu; C siþh1 u; C t iþ gðxþjrujdx (2) uðxþ2f0;1g subject to the labelling constraint (2), where g(x) is an edge indicator defined by image gradient, the second weighted length term in (2) is encoded as the weighted total-variation function regularizing the segmentation region, and the data cost functions C s ðxþ and C t ðxþ comprise of three components and are defined as: X C s :¼ x 1 1 log k s ðxþ x 2 log pðxþþx 3 h gdistðx; ^xþ where x 2 S; 0; otherwise. (3) C t :¼ x 1 1 log k t ðxþ x 2 logð1 pðxþþ þ x 3 h gdistðx; ^xþ where x 2 X=S; 0; otherwise. (4)

5 4041 Kuang et al.: Semi-automated infarct segmentation 4041 FIG. 4. om] An example of the infarct probability maps generated after the first stage and second stage training. [Color figure can be viewed at wileyonlinelibrary.c where S denotes the segmented region at the current iteration, gdistðx; ^xþ is geodesic distance function, which measures the geodesic distances from the voxel x during the current iteration to the boundary of user prelabeled infarct region ^x, using the shortest path along the image intensities, the parameter h implicitly adapted is time step size controlling the convergence rate of optimization. 16,17 This distance cost enforces the segmented infarct region close to the user initializations. In (3) and (4), the weighting parameters x 1 ; x 2 ; x 3 [ 0 and x 1 þ x 2 þ x 3 ¼ 1, weight the contributions from the local intensity profile k s;t ðxþ, RF learned information p(x), and geodesic distance for each voxel, respectively. If the binary constraint u(x) 2 {0,1} in (2) is substituted by its convex relaxation of u(x) 2 [0,1], the minimization of the energy function (2) can be globally solved by a multiplier-based continuous max-flow/min-cut algorithm. In particular, the user-prelabeled voxels in the foreground and background were used both as initialization and hard region constraints for the minimization procedure. 3. EXPERIMENTS AND RESULTS 3.A. Image acquisitions One hundred AIS patients were included in this study. 1 Ischemic infarcts were manually segmented on parallel axial views slice by slice using the paintbrush mode in the software of ITK-SNAP by a trained clinician and verified by an expert neuroradiologist. Thirty out of 100 patients were randomly selected for training while the remained 70 patients were used for testing. The size of NCCT images was (18 37) with a voxel spacing from 0:37 0:37 3mm 3 to 0:49 0:49 5mm 3. The infarct volume size for 100 patients varied greatly, ranging from 3:7mm 3 to 393:3mm 3. 3.B. Evaluation Methods Our segmentation method was evaluated by comparing the algorithm segmentations with manual segmentations in terms of the spatial overlap of infarct, using DC, the mean absolute surface distance (MAD), and maximum absolute surface distance (MAXD) with 95% percentile. 16,17 The proposed semiautomated segmentation approach was also compared to the method of thresholding the RF output probability map directly using 0.5 as a cutting point (denoted as RF), and two postprocessing methods after RF learning using convex optimization (RF+CO) 16,17 and Markov random field (RF+MRF). 19 The RF+MRF method consists of a data term and a regularization term. The data term encodes the cost (penalty) associated with assigning the voxel x as negative logarithm of the likelihood: log(p(x)). The regularization encoding pair-wise interaction potential of neighboring voxels penalizes the total surface area of the segments in the output. Furthermore, the latest convolutional neural network (CNN)-based segmentation approach, such as U-net, 20 was also used as a benchmark to evaluate the performance of the proposed semiautomated infarct segmentation approach. Around 1700 NCCT two-dimensional (2D) slices with the corresponding manual labels were used to train a U-net model. Note that the input of U-net is the NCCT intensity image in MNI space. The U-net is configured as follows. The number of layers is 5, the number of features is 64 and the size of max pooling kernel is A momentumbased stochastic gradient decent with an exponentially decaying learning rate with an initial value of 0.2 is used to optimize a dice loss function. The number of epochs is 1000 and each epoch has 500 iterations and use a training mini-batch size of 3. Online data augmentation was performed, which applied random translation, scaling, intensity variation, rotation, and left-right flipping to each image in a mini-batch before feeding them to the U-net network. The U-net segmentation results were obtained by thresholding the U-net output probability map by a value of 0.5. Additionally, the U-net estimated infarct probability was integrated into the same semiautomated covnex optimization framework, leading to another semiautomated segmentation approach, Semi_Unet. The same initializations were applied on Semi_Unet for fair comparison.

6 4042 Kuang et al.: Semi-automated infarct segmentation 4042 FIG. 5. Two examples of segmentation results using the proposed semiautomated infarct segmentation method compared to some state-of-the-art methods. The first column shows two NCCT sections of two patient images with manual segmentations (green contour) superimposed. The algorithm segmentations are colored in red, shown in other columns. DC values obtained by each method for these two cases are shown in the parentheses following the method symbol. [Color figure can be viewed at wileyonlinelibrary.com] 3.C. Results Figure 5 shows two examples of infarct segmentation results from two NCCT images using the proposed algorithm as well as other five methods, such as the RF, RF+CO, RF+MRF, Unet, and Semi_Unet methods. The proposed method yielded the most accurate segmentation for the case with a large infarct (shown in the first row in Fig. 5) while the RF, RF+CO, and RF+MRF methods under-estimated segmentation and the U-net and Semi_Unet methods over-estimated segmentation, compared to manual segmentations. For the case with a small infarct (shown in the second row in Fig. 5), the proposed method provided the best match with manual segmentation while other five methods showed under-estimation of segmentation. Further quantitative evaluations using the metrics of DC, MAD, and MAXD are shown in box plots in Fig. 6. The proposed method obtained the highest DC value while keeping the MAD and MAXD the lowest compared to the RF, RF+CO, RF+MRF, U-net and Semi_Unet methods. Values of DC, MAD, and MAXD of all methods used in this analysis are shown in Table I. The proposed method was capable of generating a DC of %, a MAD of , a MAXD of , respectively, outperforming the other five methods greatly. One-way ANOVA tests show that there is statistically significant difference among the five methods regarding DC (P < 0.001). A multiple comparison test using the Dunn Sidak method shows that our proposed method is statistically significantly better than the RF (P < 0.001), RF+CO (P < 0.01), RF+MRF (P < 0.001), U-net (P < 0.001), and Semi_Unet (P < 0.001) methods regarding DC. There is a statistical significance observed between the proposed method and other five methods in terms of MAD (all P < 0.01). This suggests that the proposed segmentation method generated overall the smallest distance error against manual segmentation compared to other methods. For MAXD, no significant difference was observed between the proposed method and the RF method (P = 0.167), although multiple pair comparisons show that there is statistical significance between the proposed method and the RF+CO (P = 0.029), RF+MRF (P = 0.021), U-net (P < 0.001), and Semi_Unet (P < 0.001). The MAXD FIG. 6. Quantitative results using 70 AIS patients in terms of DC, MAD, and MAXD. [Color figure can be viewed at wileyonlinelibrary.com]

7 4043 Kuang et al.: Semi-automated infarct segmentation 4043 TABLE I. Quantitative evaluations of our proposed method and other four methods using 70 AIS patient images. Method DC (%) MAD (mm) MAXD (mm) Ours * * RF RF + CO RF + MRF U-net Semi_Unet Values are demonstrated as mean standard deviation. *denotes statistically significant. results indicate that the proposed method still included some false positives away from the ground truth as other methods. An infarct volume analysis was performed to further quantify the difference regarding the infarct volume size measured by the proposed semiautomated approach and manual segmentation. It is found from the scatter plot in Fig. 7(a) that our algorithm segmented infarct volume is linearly related with the manually segmented infarct volume. Besides, the Pearson correlation coefficient between the two infarct volume sizes is 0.968, (P < 0.001) with 95% confident intervals of These evidences demonstrate excellent agreement regarding the infarct volume between the proposed method and manual segmentation. Since the proposed approach requires user input as initialization, which introduces observer variability, it is necessary to investigate the impact of user initialization on segmentation accuracy. To this end, an additional experiment was performed, in which, a second observer (observer 2) was asked to sample voxels from those 70 patient images again to initialize the proposed algorithm, and interobserver variability was accordingly assessed between the two observer initializations in terms of DSC, MAD, MAXD, and volume measurement. The assessment results in Table II demonstrate that two segmentation results initialized by two observers are highly consistent in terms of DC, MAD, and MAXD. Infarct volume measurements obtained by the two initializations demonstrate a strong correlation regarding a high Pearson correlation coefficient of with 95% confident intervals of (P < 0.001) in Table II and Figs. 7(b) and 7(c). A t-test demonstrates no statistically significant difference between these two measurements (P = 0.449). The low interobserver variability suggests that the proposed infarct segmentation method is insensitive to user initialization. 4. DISCUSSION AND CONCLUSION In this study, we propose a semiautomated segmentation approach to segmenting ischemic infarct from follow-up NCCT images of AIS patients. The proposed approach makes use of two cascaded RF classifiers followed by a convex optimization-based postprocessing technique while using user initialization as priors. Quantitative evaluations against manual segmentations show that the proposed approach is accurate in terms of metrics of DC, MAD, and MAXD, using 70 AIS patient NCCT images. Further infarct volume analysis shows that the infarct volume measured by the proposed segmentation approach agrees very well with manual measurement, suggesting its potential in being used in clinical studies. We deployed a two-stage cascaded training mechanism using a few well-designed hand-crafted features to tackle this challenging segmentation problem. We have shown the advantage of two-stage training mechanism over the traditional one-stage training in Fig. 4, which shows that the infarct probability map output by the two-stage training mechanism contains substantially smaller false positive rate than on-stage training while maintaining high precision. Compared to deep learning-based methods, such as the U- net, the RF methods using our proposed hand-crafted features yielded better accuracy regarding DC, MAD, and MAXD in our experiments, suggesting that the shallow learning technique is not inferior to deep learning technique in this study, provided that the used features are sufficient. In addition, user interaction and the RF learnt features are integrated into a global optimization-based contour evolution framework, demonstrating significant improvement over other postprocessing techniques, such as MRF. We additionally explored the feasibility of inputting the hand-crafted features used in FIG. 7. Infarct volume measurement analysis. Two algorithm segmented infarct volumes initialized by two different observers (named Vol_1 and Vol_2, respectively) are shown in (a) and (b) against manually segmented infarct volume (named GT_Vol). The relationship between two algorithm segmented infarct volumes is plotted in (c). [Color figure can be viewed at wileyonlinelibrary.com]

8 4044 Kuang et al.: Semi-automated infarct segmentation 4044 TABLE II. Inter-observer variability assessment using two initializations from two observers in terms of DC, MAD, MAXD, and volume measurement. DC (%) MAD (mm) MAXD (mm) Correlation Observer [ ], P < Observer [ ], P < The values of DC, MAD, and MAXD are shown in mean standard deviation. FIG. 8. Two examples of low DC values. (a) Interference from partial volume effect of CSF (b) Infarct accompanied by hemorrhage. Green contour: manual segmentation, Red regions: algorithm segmentations. DC values are listed in the parentheses. [Color figure can be viewed at wileyonlinelibrary.c om] this study as different channels into the U-net. It was found that this method generated a promising DC of %, which is significantly higher than % of the U-net, but still lower than % of the proposed well-trained RF. This might be due to the limited training data for the U-net-based methods. When the probability maps generated by this method were integrated into the proposed semiautomation framework using convex optimization, it achieved a DC of %, lower than % of the proposed method, but higher than the DC values obtained by other methods. In spite of using online data augment, the U-netbased method still cannot be trained well to account for large acquisition variability of medical imaging data in our experiments. However, it is worthy of exploring further on this direction in this application as well as other medical imaging applications with limited data and ground-truth. Although some automated infarct segmentation methods have been reported, 7,9 most of them were evaluated with limited data, and heavy manual editing after automatic segmentation was required. Compared to the semiautomated methods 10,11 in the literature, our obtained DC of 79.42% was higher than 74%, 10 but lower than 85%. 11 However, the insufficient validation with only 13 patient images 11 was less convincing as DC values might be biased by volume size. Furthermore, the latest CNN-based method, U-net, 20 was applied on the same dataset as well. However, it generated the poorest accuracy in this application with a DC of 40.18%. When combined with the proposed semiautomated postprocessing method, the DC of the Semi_Unet was significantly improved to 59.72%, but still lower than the proposed method. Comparing the intensity and appearance between infarct and the normal tissue on the contralateral side is an intuitive way to pick up ischemic infarct from NCCT images for radiologists. The features used in this study are therefore designed to simulate the way radiologists interpret NCCT images. In addition, the feature of spatial probability of infarct occurrence is designed to estimate how likely a voxel will infarct in a specific anatomic region in brain. In the last subfigure in Fig. 3, it can be seen that the middle cerebral artery region (MCA) has higher probability of being infarcted than other brain regions, whereas other regions, such as anterior and posterior cerebral artery regions (ACA and PCA), have lower probability. This correlates to this fact that the MCA region is the most common site for the occurrence of ischemic stroke. We have performed an additional experiment by training a new model using all the designed features excluding this infarct occurrence feature. The dice obtained by this model is 75.24%, significantly lower (P < 0.01) than 79.42% reported by the proposed method in this study, which suggests the effectiveness of this feature. In order to assess the sensitivity and stability of the lesion occurrence probability feature, we calculated the proportion of the infarcted ASPECTS regions 5 in both 30 training images and all 100 images. It was found that the proportions of ASPECTS regions with infarction in 30 training images are similar to the ones in all 100 images. Fisher s exact tests did not show significant differences in each ASPECTS region between these two datasets, which suggests that the lesion probability map generated from 30 training images randomly selected is able to approximately represent the spatial distribution of the infarct in this application. In addition, we generated another

9 4045 Kuang et al.: Semi-automated infarct segmentation 4045 lesion probability map using 80 patient images, and used this map to retrain a new RF model to segment the remained 20 patient images. It is found that the Dice value increased slightly from 79.42% to 79.63% for these 20 patients regardless of extra computational time. We acknowledge that this study has limitations. First, false positives caused by partial volume effect of CSF is still challenging to remove, especially for the cases with small infarcts close to ventricles [Fig. 8(a)]. In these cases, more user interactions are required. Second, the proposed segmentation approach cannot deal with the cases with ischemic infarct accompanied by hemorrhage [Fig. 8(b)]. However, it is possible to extend the current approach by using two different classifiers for ischemic infarct and hemorrhage sequentially or by using a multi-label classifier. Third, the proposed method is able to segment both focal and multi-focal unilateral infarcts, but not bilateral infarcts. Finally, validation on a larger population would be desired to account for variability in image acquisition with different levels of noise, artifacts, motion, and scanning parameters. To conclude, a semiautomated infarct segmentation approach is proposed for measuring ischemic infarct volume from post-treatment NCCT scans of AIS patients. The quantitative evaluations using 70 AIS NCCT images show that the proposed technique is accurate and robust, suggesting its potential in being used to measure radiologic outcome in AIS patients. ACKNOWLEDGMENTS This work was supported by the Canadian Institutes of Health Research and Alberta Innovate Health Solution. CONFLICT OF INTEREST The authors have no conflicts to disclose. a) Author to whom correspondence should be addressed. Electronic mail: qiu.wu.ch@gmail.com REFERENCES 1. Goyal M, Menon BK, van Zwam WH, et al. Endovascular thrombectomy after large-vessel ischaemic stroke: a meta-analysis of individual patient data from five randomised trials. Lancet. 2016;387: Boers AM, Jansen IGH, Beenen LFM, et al. Association of follow-up infarct volume with functional outcome in acute ischemic stroke: a pooled analysis of seven randomized trials. J Neurointerv Surg. 2018;10: Maier O, Menze BH, von der Gablentz J, et al. ISLES 2015-A public evaluation benchmark for ischemic stroke lesion segmentation from multispectral MRI. Med Image Anal. 2017;35: Usinskas, A, Dobrovolskis RA, Tomandl BF. Ischemic stroke segmentation on CT images using joint features. Informatica. 2004;15: Kuang H, Najm M, Chakraborty D, et al. Automated ASPECTS on noncontrast CT scans in patients with acute ischemic stroke using machine learning. Am J Neuroradiol. 2019;40: Qiu W, Kuang H, Nair J, et al. radiomics-based intracranial thrombus features on CT and CTA predict recanalization with intravenous alteplase in patients with acute ischemic stroke. Am J Neuroradiol. 2019;40: Nowinski WL, Gupta V, Qian G, et al. Automatic detection, localization, and volume estimation of ischemic infarcts in noncontrast computed tomographic scans: method and preliminary results. Invest Radiol. 2013;48: Tyan Y-S, Wu M-C, Chin C-L, Kuo Y-L, Lee M-S, Chang H-Y. Ischemic stroke detection system with a computer-aided diagnostic ability using an unsupervised feature perception enhancement method. J Biomed Imaging 2014;2014: Gillebert CR, Humphreys GW, Mantini D. Automated delineation of stroke lesions using brain CT images. Neuroimage Clin 2014;4: Boers AM, Marquering HA, Jochem JJ, et al. Automated cerebral infarct volume measurement in follow-up noncontrast CT scans of patients with acute ischemic stroke. Am J Neuroradiol 2013;34: de Haan B, Clas P, Juenger H, Wilke M, Karnath H.-O. Fast semiautomated lesion demarcation in stroke. Neuroimage Clin 2015;9: Kuang H, Najm M, Menon BK, Qiu W. Joint segmentation of intracerebral hemorrhage and infarct from noncontrast CT images of post-treatment acute ischemic stroke patients. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages , Springer, Sun, J, Rehg JM, Bobick A. Automatic cascade training with perturbation bias. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 2, pages II II, IEEE, Najm M, Kuang H, Federico A, et al. Automated brain extraction from head CT and CTA images using convex optimization with shape propagation. Comput Methods Programs Biomed 2019;176: Warfield SK, Zou KH, Wells WM. Simultaneous truth and performance level estimation (STAPLE): an algorithm for the validation of image segmentation. IEEE Trans Med Imaging 2004;23: Qiu W, Yuan J, Ukwatta E, Sun Y, Rajchl M, Fenster A. Prostate Segmentation: an efficient convex optimization approach with axial symmetry using 3D TRUS and MR images. IEEE Trans Med Imaging 2014;33: Qiu W, Yuan J, Ukwatta E, Sun Y, Rajchl M, Fenster A. Dual optimization based prostate zonal segmentation in 3D MR images. Med Image Anal 2014;16: Qiu W, Chen Y, Kishimoto J, et al. Longitudinal analysis of pre-term neonatal cerebral ventricles from 3D ultrasound images using spatialtemporal deformable registration. IEEE Trans Med Imaging 2017;36: Boykov Y, Veksler O, Zabih R. Fast approximate energy minimization via graph cuts. IEEE J Biomed Health Inform 2001;23: Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages , Springer, 2015.