Chapter 13 - Optimization-based multilevel threshold image segmentation for identifying ischemic stroke lesion in brain MR images

Abstract

Brain stroke, a noncommunicable disease, is the third prime cause of death and disability worldwide. According to World Health Organization, 17 million people worldwide suffer from stroke. Every year 6.2 million die and 5 million remain permanently disabled. In India, every year the stroke population varies from 116 to 163 per 100,000 persons. Around 87% of all strokes are ischemic strokes (IS). Stroke is also termed brain ischemia or cerebral ischemia. The interior structural brain regions are investigated to identify the presence of IS. Magnetic resonance (MR) imaging is useful to diagnose IS and to determine the treatment strategy in the acute phase. In this work, diffusion-weighted (DW) MR images have been acquired from the Ischemic Stroke Lesion Segmentation Challenge 2015 dataset. Though MR imaging has good image quality, the low contrast between tissues could make it very difficult to analyze the abnormalities in the image; this also affects the accuracy of clinical diagnosis. The quality of the image is improved by using preprocessing techniques such as filtering. An anisotropic diffusion filter (ADF) provides better visualization, which aids diagnostics. The edges of the MR images are preserved in spite of noise being removed by ADF. Performance of the preprocessing technique is analyzed by using statistical-based histogram features. An ischemic stroke lesion (ISL) in an MR image is identified by a multilevel thresholding-based harmony search optimization algorithm and an electromagnetism-like optimization (EMO) algorithm for Otsu and Kapur object functions. The segmented outcomes acquired from these two algorithms are compared with expert's ground truth images and their performance is analyzed using a statistical measure of similarity indices. It is observed that the multilevel thresholding EMO-based Kapur method performs better in identifying ISL in the DW modality of brain MR images.