Visualization of MRI Datasets for Anatomical Brain Segmentation by Pixel-level Analysis

Abstract

In this paper, segmentation of the brain MRI images are carried out by performing pixel-level analysis with deep learning approaches. Image segmentation is often used for medical image analysis and is an early diagnosis tool for clinical applications. Brain images acquired by MRI machines need to be segmented into anatomical structures for detecting morphological changes. One can then visualize the anatomical structures and take quantitative measurements if required. This allows surgeons to plan out surgeries and the method of approach when removing tumors or other lesions. As large MRI datasets become available, deep learning can be used to carry out pixel-level analysis for determining its class in the brain anatomical structure, which can enhance visual intelligence. This paper aims to use the latest deep learning techniques like dropout, batch normalization and data preprocessing to accurately segment images based the MICCAI 2012 dataset, which contains 35 manually segmented MRI into 134 anatomical brain structures. The segmented result into 134 classes/regions are then visualized in colored images. The model uses a multi-patch data extraction algorithm which helps to classify the center voxel of the extracted patches. The result obtained reaches a mean dice coefficient of 76.2%, which has outperformed the best previous result of a mean dice coefficient of 72.5%.