Performance analysis of various segmentation techniques for detection of brain abnormality

In the field of neuropsychiatric disorders, it is known that brain segmentation is important for both detection and diagnosis. Processing of MRI/PET images for brain pathology detection is a challenging task for radiologists. To address this issue the performance of various segmentation techniques has been analyzed with Independent component analysis (ICA) as a pre-processing step for removal of salt and pepper and speckle noise at 5dB noise level. There are so many segmentation methodology used for medical diagnosis, which comes under the classification strategy like supervised and unsupervised segmentation, region and edge based segmentation. The methodologies considered in this paper will cover the main classification strategy of segmentation for brain abnormality. Mean Shift (MS), Fuzzy C Means (FCM), Hough Transform (HT), Normalized Graph Cut (NGC), Thresholding by Histogram (ThH) and Support Vector Machine (SVM) are taken here for performance analysis. Experimental results suggest that, for PET images, pathology detection is found good while using ICA as denoising method for removing salt and pepper and speckle noise at 5dB and SVM as segmentation technique. Whereas for MRI images the performance of both ThH and SVM goes hand in hand as a segmentation methodology with ICA as noise removal method. The evaluation measure used here are Jaccard and Dice Coefficient, Peak Signal to Noise Ratio (PSNR), Global Consistency Error (GCE), Under Segmentation (UnS), Over segmentation (OvS) and Incorrect Segmentation (InC), Selectivity, Specificity, Accuracy, Positive predictive value (PPV) and Negative predictive value (NPV) . From the obtained results it is seen that SVM gives better result for detection of mild cognitive impairment in PET scan images and both SVM and ThH is performing good for brain tumor detection in MRI images, without affecting the image quality.