Stroke is a condition where the supply of blood from the heart to the brain through cerebral vascular arteries is interfered. This interference arises as a result of cholesterol and fat deposition inside the blood vessels of arteries [1]. A high percentile of 66% death with only 34% recovery rate, as per world health organization reports suggest the significance of early diagnosis of stroke [2]. The two classes of stroke found among human are ischemic and haemorrhagic stroke. Here, ischemic stroke falls under the major category, while haemorrhagic falls under severe category [2].

Grading, assessment of spread and severity, planning and procedures of the treatment can be sped up through automation in the diagnosis of stroke. Haemorrhagic stroke is a condition where rupture of a blood vessel occurs, haemorrhagic stages are usually followed by ischemic stroke. Hence, diagnosis of stroke at the ischemic stage is essential for preventing death and disability [3]. The two medical imaging methods that identify stroke lesions and predict the area of stroke spread are computed tomography and magnetic resonance imaging. [4] [5]

Stroke infarcted location and volume occupancy in medical images enables experts to categorise the proportion of physical disability among stroke patients [6]. MRI supports stroke diagnosis in an upper hand over CT due to softer tissue differentiation available in the Diffusion Weighted modality of MRI images [7]. Clustering is an image segmentation technique where an image is divided into different regions based on features like intensity and other measurable characteristics. The clustering algorithm K-means and its variants employ sharp boundaries delineating each cluster, while FCM and its variants allows a pixel to be part of more than one group or cluster based on a membership value. Both these methods are intuitive to the boundary leakage problem, where a disease pixel could be missed or a normal pixel could be misclassified. These problems are addressed by active contour or level set methods-based image segmentation.

Clustering algorithm provides non-precise or rough object boundaries, these rough boundaries can be initialized as initial contour for level set segmentation and further refined through level set-based evolution convergence method for fine-tuning boundary detection, making the segmentation more accurate. Medical images often face the problem of no clear separation between object boundaries, when such images are processed through clustering techniques, it might result in multiple fragments, while the level set method holds the possibility of combining these different fragments into a coherent object. Thus, there is a demanding need for hybrid approaches that combine both clustering and level set-based segmentation.

In the proposed work, a hybrid approach that makes use of clustering and level set method, along with machine learning classifier and optimization techniques are utilized. The pre-processed diffusion weighted MRI image gets divided in to multiple clusters by FCM algorithm. The best cluster that matches the stroke lesion are identified through Error Corrected code output Multi Classifier Support Vector Machine Classifier using GLCM features. Level set evolution begins with the edge of the best cluster as the initial level set. The progression of level set evolution is carried out through competition of pixel for ROI region through binary arithmetic optimization algorithm based arbitrator. This article is organized with introduction followed by related works, materials and methods, results, discussion, conclusion and references.

Brain CT and MRI images provide visual support for the diagnosis of stroke by medical experts [1]. The molecular movement of blood along structures of tissues portrays a wide range of diffusion in grey matter, white matter, Cerebro Spinal Fluid (CSF) and other disease lesions [1]. Stroke lesion has the nature of appearing as hypo intense in chronic cases and hyper intense in acute cases due to absorption of water molecules in the core region of stroke. The progression of stroke among patients and its diagnosis can thus be evaluated through Diffusion Weighted MRI image [1]. Diagnosis time could be reduced through the development of methods that provide automatic extraction of the stroke lesion. The automation algorithm and its performance depend on the way it support this task without going through mishaps. Clustering is a category of image segmentation algorithms that serves the purpose of grouping pixels of an image on the basis of similarity of intensity, texture and other features. The two important classes of clustering algorithm that serves the purpose of image segmentation are K means and Fuzzy C Means algorithms. K-means algorithm are crisp in partitioning, i.e. a pixel is not allowed to participate in more than one cluster or grouping, while fuzzy c-means allows a pixel to belong to more than one category on the basis of a fuzzy membership value. These algorithms suffers from random cluster centre initialization, missing of local spatial information, boundary leakage problems, and traps at local optimum. In histogram-based filter-enhanced FCM [8] the authors have modified the working by grouping pixels of similar intensity into super pixel before clustering operation, they have adopted morphological operation along with filtering to include local spatial information, yet this algorithm slows down in segmentation result due to random cluster centres initialization. In arithmetic optimization-based K Means algorithm [9] the authors adopted arithmetic optimization algorithm for finding best cluster centres before K Mean’s cluster grouping, while in exploration enhanced dynamic arithmetic optimization, the [10] is enhanced through exploratory operator by replacing multiplication, division with power and logarithmic operators and Canberra distance metric. Despite these improvements, segmentation algorithm suffers from boundary leakage problem where there are discrete discontinuities between normal and stroke lesions. Deformable model-based image segmentation method [11] has the ability to address boundary leakage problem of clustering algorithm. One such method is level set segmentation which accumulate information from image to drive for optimized segmentation. Level set operates either through level set evolution followed by edge detection or level set evolution followed by region competition. A clear and Distinct object boundary is the essential prerequisite for edge-based level set methods. The stroke lesion in an MRI image is ambiguous due to several leakages caused by the dynamic interface. Region-based Level set method does not require clear, distinct boundaries, but the images are to be of high contrast. Ciofolo et. al [12] utilized a competitive level set with fuzzy control where a fuzzy decision system works on a priori atlas data and intensity profile to make a decision. Rebouças et.al [13] adopted thresholds at the edge and centre for zero level set implementation. Ho, S et al. [14] used a region competitive level set method where local statistical force is modulated to bring in stable solution. Abdol-Reza Mansouri et al [15] adopted a multi-region competition based level set based on motion disparity removal. All these surveyed methods suffer from volume overlap error, boundary leakage problems, random initial contour which are addressed through the proposed work. In the proposed method, Fuzzy C Means is adopted to segment initial cluster and further post processing is carried out via best cluster identification from Error Corrected Code Output Multi Class SVM. It is followed by pixel identification, which fits in the disease region through binary optimization-based pixel fitness evaluator operating over competitive level set evolution.

The input diffusion weighted image MRI is fed to the clustering unit, where the initial clustering operations are carried out through Fuzzy C Means. FCM allows a pixel to be part of more than one cluster group based on the membership or cost function. The membership value of a particular pixel to each cluster group is determined on the basis of the cost function given in Eq. (1).

Where

-Centroid of cluster K

- fuzziness control parameter

- Intensity color or texture at each pixel location

Fuzzy membership of pixel at location x and y in cluster k, is given by Eq. (2)

Centroid of cluster k, is obtained from Eq. (3)

In this work the value of is chosen as two for maintaining Euclidean distance norm. The number of clusters to be segmented is taken as four resembling grey matter, white matter cerebro Spinal Fluid (CSF) and stroke lesion lesions.

unique arrangement of intensity or Gray level that contributes for differently shaped object classes in a single image are obtained through GLCM texture features. Eq. 6 to 11 are used to calculate contrast, energy, homogeneity and correlation (using mean and variance) for individual clusters from FCM clustering unit. The average values of above features obtained for Grey Matter, White Matter, CSF and stroke lesions are shown in Table 1.

The input output of clustering unit are given in figure `1. Implementation and results of proposed ECCO-MC-SVM-BAO-RCLS for a sample image is given in Fig. 2. Figure 2(a) represents the original DWI image. Figure 2(b) represents the accumulated outputs of clustering unit. Figure 2(c) represents the best cluster selected by ECCO-MC-SVM. The cumulative accumulation of remaining all cluster except the best cluster is given in Fig. 2(d). 2(c) and 2(d) are outputs from competition region formulator. Initial level extracted from canny edge detector operating on best cluster output of ECCOMC SVM is given in Fig. 2e. Final segmentation result through region competition on the evolving level set through BAO on fitness function is given in Fig. 2(f).

ECCO-MC-SVM classifier is supplied with features from gray level Cooccurrence Matrix as listed earlier from equations 6 to 11. ECCO-MC-SVM is trained at different learning rates, and the evaluations are repeated for different number of iterations to obtain the optimal learning rate. This hyperparameter tuning through different learning rates gives maximum accuracy and minimum loss. The optimal learning is obtained at a learning rate of 0.5, where average loss is 0.19 and average accuracy is 0.96. These results are accurate for the ECCO-MC-SVM network that is supplied with the current data set. Additionally, there are possibilities for the optimal learning rate to differ based on variations in the input data set. The binary arithmetic optimization-based pixel fitness detector is supplied with input from a competitive region evaluator, which formulates two regions: the best cluster and the accumulated value of all clusters except the best cluster. The BAO-based pixel fitness detector is thus the BAO algorithm modelled as a binary classifier. It evaluates the fitness function of region competition through the balloon force of level set evolution, modifying the fitness value for each pixel over the competitive region i.e. if the region is Interior to ROI (shrinking) or if the region is exterior to ROI (expanding). Thus, binary arithmetic optimization plays a major role in the decision of the shrinking or expansion of active contour.

Table 1 provides confusion matrix components of the proposed segmentation algorithm, and Table 2 represents segmentation scores for the subsample dataset of 28 images from the ischemic stroke lesion segmentation challenge 2015 dataset, Diffusion Weighted MRI image. The segmentation metrics are obtained by comparing the ground truth against each segmentation result. These segmentation metrics are obtained using the following formulas in equations (31) to (34)

Where TP = ground truth’s positive matches with segmentation results’ positive

TN = ground truth’s negative matches segmentation result’s negative

FP = positive in segmentation result but negative in ground truth

FN = Negative in segmentation result but positive in ground truth

Comparison of state of art methods with proposed method for clustering approaches on stroke lesion segmentation is given in Table 3. This comparison shows the proposed method outperforms the available methods in parameters of recall and dice while the precision is maintained in methods as FCM allows a pixel to belong to more than one category through its fuzziness.

The clustering-based approaches in stroke lesson segmentation suffers from random cluster centroid initializations, traps through local optimum, loss of local or neighborhood information, while the level set approaches which enhance results from clustering approaches that suffers from boundary leakage problems. These sufferings are addressed through the proposed Error Corrected Code Output Multi Class SVM and Binary Arithmetic Optimization based Region Competitive Level Set Frame work. In this approach FCM clustering is applied over 2D diffusion weighted MRI image for initial grouping (clusters). These operation through clustering provides segmentation result, which in turn is enhanced by finding the disease cluster that matches with the ground truth using ECCO-MC-SVM. Canny Edge detection is applied over the best cluster to determine the initial level set function. In the subsequent step the clusters obtained through FCM algorithm is separated into two regions where region one corresponds to the best cluster that is in alignment with the ground truth and the union sum of all clusters except the best cluster forms region two. This binary competitive region is formed through competitive region formulator. The binary arithmetic optimization-based pixel fitness detector evaluate the fitness function computed by the competitive region formulator through the probability of occurrence at each grey level. Through this fitness evaluation by BAO-based pixel fitness evaluator a pixel taken for region competition is thus fitted either into region one or region two. This in turn activates the balloon force of the initial level set which either shrinks or expand the initial level set. This process is repeated till reaching the convergence or until reaching the minimal error rate. The final segmentations is obtained through evolution and convergence of the initial level set taken care by region competitive level set framework. The final segmentation results are correlated against the ground truth to obtain the segmentation metrics of accuracy, Sensitivity or Recall, Dice and Precision. The results of the segmentation through proposed approach are best when compared to the other state-of-the-art methods with 83% DICE score, 77% recall value and 92% precision in average. In future new techniques that addresses the complexities of adapting classifiers for detecting the best cluster and evaluating the pixel for region competition can be sought through. This could produce a faster segmentation process over 2D images when compared with 3D deep network-based segmentation and classification.