Modified Watershed Transform for Automated Brain Segmentation from Magnetic Resonance Images

Abstract

The segmentation of human brain from Magnetic Resonance Image (MRI) is one of the most important parts of clinical diagnostic. Brains' anatomical structures can be visualized and measured through image segmentation. Especially, while clinical analysis of magnetic resonance images, accurate segmentation is a crucial task for precise subsequent analysis. Watershed transform is a widely used segmentation method in medical image analysis filed. Regarding MRI images, they always contain noise caused by different operating equipment and environmental situation. However, the performance of the watershed transform depends on converges of numerous local minima on the image. Wrong regional minima on the image cause a high rate of over-segmentation of the watershed transform method. To address this problem, in this paper we propose a modified watershed transform method to prevent over-segmentation using k-means clustering method. Our modified watershed transform utilizes the k-means clustering method for region classification to remove wrong regional minima on image and provides a guideline for watershed transform to prevent the over-segmentation problem. Experimental results on brain MRI images evaluations (Dice coefficient: 95.32%) demonstrate that the proposed method can substantially prevent the over-segmentation problem of conventional watershed transform method.