In this study, 98 ischemic stroke patients (31 women and 66 men, with age range 24 - 78 years) within 6 hours from the symptom onset were used. They were collected from Tiantan Hospital and Tianjin Huanhua Hospital since 2005. The protocol of the research has been approved by the Institutional Review Board of both hospitals. All patients gave written consent and provided permission for scientific and educational purpose. The baseline DW imaging was carried out with two 3.0 Tesla scanners (Trio-Tim, Erlagen, Germany) with a spin-echo, multislice, and single shot echo-planar of b = 0 (B0 image) and 1000 or 1500 s/mm² (DWI) and the corresponding ADC map. The imaging covered the whole brain, with 19 - 24 axial slices of a 5 mm slice thickness and 1 - 1.5 mm gap, most matrixes being 128 × 128 with few being 156 × 156 or 384 × 384 to have in-plane resolution of 0.60 mm to 1.80 mm.

Our study consist of the following components: extracting relevant features, employing feature selection algorithms to choose more robust features, and applying SVM to segment ischemic infarcts. The algorithm was implemented with Matlab version 2013a running on Microsoft windows with Core i5 3.2 GHz processor and 8GB memory. The schematic diagram is shown in Figure 1.

All images were normalized to the range of [0, 255]. Asymmetry map (ASM) was calculated to reflect the characteristics of infarcts in DWI and ADC: being hyper-intense on DWI, hypo-intense on ADC, and being generally asymmetrical with respect to the midsagittal plane (MSP) extracted from B0 [6].

where (x₀, y₀) and (x, y) are symmetrical with respect to the midsagittal line (which is the intersection between the MSP and the axial slice) on the axial slice z, N_s(x₀, y₀) is the neighborhood of (x₀, y₀) and the window size of window is 5 × 5. Axial slices of B0, DWI, ADC and ASM of a patient are shown in Figure 2. It is clear that the infarct region on the right hemisphere has been enhanced to be more easily discernable than both the DWI and ADC.

To alleviate the interference of normal brain tissues and speed up the calculation, candidate infarcts in the form of regions of interest (ROIs) are extracted from thresholding ADC maps subjected to DWI [5]. Specifically, denote the most frequent ADC value for all voxels within the brain as ADC_ref, any voxels with ADC not greater than 0.75*ADC_ref are to formulate connected components. Any connected components with average DWI values not smaller than the intensity average plus the intensity standard deviation of brain voxels on DWI at the corresponding axial slice are kept as part of region R₁. Voxels with ADC value within (0.75*ADC_ref, 0.85*ADC_ref] are checked to formulate connected components and added to R₁ if the components has at least one neighboring voxels in R₁. R₁ is then increased by morphological dilation with a structuring element of radius R_d (being 2 voxels). Suppose the z coordinates of R₁ are in the range of z₀ to z_n with z_n >= z₀, then R₁ region on the axial slice z₀ and z_n are respectively pasted to axial slice z₀ − 1 and z_n + 1 to attain the eventual ROIs. Figure 3 shows an axial slice of DWI, ADC and the corresponding ROIs (in color), where regions in red are initial ischemic infarcts and those in yellow are added voxels through the described morphological dilation.

For training, the voxels in infarct regions are the positive samples and drawn by a specialist. While the negative samples are those voxels close to the infarct regions obtained through iterative dilation of the infarct regions such that the number of negative samples is just equal to or greater than that of the positive samples. The features for each voxel are: intensity value, mean value, standard deviation, the weighted mean intensity value with Gaussian kernel (mean being 0 and standard deviation being 1), minimum value, maximum value, median value, skewness, kurtosis. All of these features are calculated from a local window of size 3 × 3 except the intensity value. The feature vector is formulated form the 9 quantities of B0, DWI, ADC and ASM to be 36 dimensional.

Segmentation of ischemic infarcts can be considered as a classification of two classes: infarcts and non-infarcts. Two-fold cross-validation is applied in this research. The classifier is SVM while features are optimized through the feature selection procedure detailed below.

SVM was first proposed by Vapnik [7] as a supervised learning method to transform training samples into a multidimensional hyperspace and then computes decision boundary in hyperplane by solving the convex optimization problem. The margin between the positive and negative samples from the separating hyperplane is to be maximal. To obtain the maximal margin, we employ LIBSVM tool box to obtain optimal parameters [8], using radial basis kernel functions (RBFs). Two important parameters of the RBF, penalty factor C and the parameter γ, are respectively set to 32 and 0.5 using grid search method to obtain a good performance for training samples.

Redundancy and irrelevant features may lead to non-effective classification model. It is desirable to rank the importance of the features and select the distinguishing features. Five feature selection methods are employed and compared in our experiments.

Fisher score (FS) [9] identifies features with small within-class distance and large between-class distance. Threshold value is used to differentiate effective features.

Gini index (Gini) [10] quantifies the ability of feature to differentiate the classes. The features are relevant with smaller Gini Index. The top k features with the smallest Gini index are selected.

Mutual-information (MI) [11] describes the relationship between the features and categories. It calculates information between the feature and the category with a larger MI for larger information.

Minimum redundancy maximum relevance (mRMR) [12] is a filtering method based on space search. It searches for the feature subsets by using information gap and information entropy to obtain the feature subsets.

ReliefF algorithm [13] computes feature weight to rank features with linear time according to the number of features extracted from datasets.

To assess the performance of the segmentation, three popular evaluation metrics are used, the Dice coefficient (Dice), sensitivity and specificity as calculated below: Dice = 2TP/(FP + FN), sensitivity = TP/(TP + FN), and specificity = TN/(TN + FP), where TP is for the number of correctly segmented infarct voxels, TN is for the number of correctly segmented non-infarct voxels, FP is for the number of wrongly segmented infarct voxels, and FN is for the number of wrongly segmented non-infarct voxels.

To see the relative importance of different images in the classifier, we designed classifiers with each singular image and classifiers with one image being excluded. For instance, we designed a classifier from ASM image to obtain an average Dice of 0.749 ± 0.122. The results are shown in Figure 4.

We compared the suitability of a number of feature selection methods. Each selector is applied to the training data and returns a ranking of the feature vector. Different sizes of features are tested on each selector. The mean Dice over all testing images on SVM are shown in Figure 5.

The highest Dice is achieved from feature selection using on MI algorithm with 15 features on SVM. The mean Dice, sensitivity and specificity are 0.763 ± 0.121, 76.08% ± 14.8%, 99.95% ± 0.098% respectively. Two slices with good and bad segmentation are shown in Figure 6.

The proposed method has been compared with divergence measure based method (DM method) [3] and FCM based method [4]. For a fair and relevant comparison, we have tried our best to implement them to attain the best performance on the 98 datasets within 6 hours from symptom onset. The Dice, sensitivity and specificity for [3] ([4]) are respectively 0.215 ± 0.213 (0.597 ± 0.204), 0.565 ± 0.346 (0.585 ± 0.221) and 0.983 ± 0.024 (0.999 ± 0.001).