<p>Stroke is one of the largest contributors to death and disability across the world. There is increasing demand for automation methods that detect and characterize different classes of stroke. Unsupervised region of interest extraction through clustering algorithms like K-means, Fuzzy C-means and their variants exhibits degradation when these algorithms go through image modalities with intensity inhomogeneity, noise and traps at local maximum. Active contour-based level set method is widely used in medical imaging as it accounts for fine-tuning the segmentation results from clustering approaches. These level set image segmentation methods fail in their evolution-convergence process when they are supplied with images having weak boundaries and a random initial contour over a fixed region of interest. These issues are addressed by our proposed Level set framework, which uses Error Correcting Code Output Multi Class Support Vector Machine for identifying stroke lesion group from the output of the FCM algorithm followed by region competition using binary arithmetic optimization-based pixel fitness detector for evolution and convergence of the level set. Encouraging Segmentation results are obtained in the proposed method with average values of 98.87% accuracy, 76.8% sensitivity, 83.1% Dice and 91.9% precision.</p>