Optimization Techniques for the Multilevel Thresholding of the Medical Images

Introduction

Image segmentation is the process of extracting meaningful objects from the image. Computed tomography and MRI are the common imaging modalities that help physicians to non-invasively study the brain structures (Gondal and Khan, 2013). MRI is preferred modality of choice due to its higher soft tissue contrast and the availability of multispectral images. Upon careful analyzing these multiple images, a radiologist provides a decision about the lesion extent and the therapy treatment. Various tissue constituents in an MR image can be either manually marked by the radiologists (Georgiadis et al., 2007, 2008; Zacharaki et al., 2009) or they can make use of the semi-automatic or fully automatic computer-assisted approaches. Manual segmentation of MR image is both tedious and time-consuming due its 3D nature. Also, the outlined tissue constituent suffer from the effect of interobserver variability (Dou et al., 2007; Corso et al., 2008; Hemanth et al., 2013). Though several segmentation methods have been presented in the past year's, thresholding is undoubtedly one of the most attractive technique due to its simplicity.

For the analysis of the medical images, many thresholding techniques have been developed in recent years which are either based on Kapur’s entropy maximization (Kapur, Sahoo and Wong, 1985; Manikandan et al., 2014) or the cross entropy minimization (Kaur, Saini and Gupta, 2016). The entropy criteria proposed by Kapur’s and Otsu’s based between class variance approach has been widely used to find the optimal thresholds in the case of images with bimodal histogram but for images with multimodal histogram, these approaches suffer from the drawback of an exponential surge in the computational time (Manikandan et al., 2014). To overcome this problem heuristic based algorithms have been applied for dividing the image into several regions by obtaining the optimal set of threshold values.

Yin (Yin, 2007) presented a method that used particle swarm optimization (PSO) for finding of the optimal thresholds. The criteria employed for the multilevel thresholding was minimum cross entropy, and the efficacy of the method was validated by the measure of the computation time that was faster in than the exhaustive search methods. Nakib et al. (Nakib et al., 2007) devised two-dimensional survival exponential entropy criteria based on PSO for MR image segmentation. The results were compared with those obtained using 2D Shannon entropy as the objective function in terms of the misclassification error.

Djerou et al. (Djerou et al., 2009) proposed binary PSO for determining the optimum number and the corresponding values of the various thresholds. The Kapur’s entropy and the Otsu’s criteria were employed as objective function, uniformity measure, and computation time were used as a performance metric. The results illustrate that the Otsu's method was faster than the Kapur’s method.

De et al. (De et al., 2010) performed the segmentation of diseased MRI using PSO based Shannon entropy criteria. The segmented region was post-processed by applying masking operation. The method was based on the conception that diseased area would have different intensity level in contrast to the normal regions. Firstly, the normalized image histogram was obtained and then the entropy maximization was used to find the range of gray values for the diseased region using PSO. After that, region growing in the form of masking was also done to attain the final segmented image.