Abstract
Automated segmentation of brain lesions in magnetic resonance images (MRI) is a difficult procedure due to the variability and complexity of the location, size, shape, and texture of these lesions. In this study, four algorithms for brain lesion detection and segmentation using MRI are proposed. In the first algorithm, an automatic algorithm for brain stroke lesion detection and segmentation using single-spectral MRI is proposed, which is called histogram-based gravitational optimization algorithm (HGOA). HGOA is a novel intensity-based segmentation technique that applies enhanced gravitational optimization algorithm on histogram analysis results to segment the brain lesion. The ischemic stroke lesions are segmented with 91.5% accuracy and tumor lesions are segmented with 88% accuracy. Since histogram analysis limits the extracted information to the number of pixels in specific gray levels and does not include any region-based information, the accuracy of a histogram-based method is limited. In the second algorithm, in order to increase the accuracy of brain tumor segmentation, a texture-based automated approach is presented. The experimental results on T1-weighted, T2-weighted, and fluid-attenuated inversion recovery (FLAIR) images on both simulated and real brain MRI data prove the efficacy of our technique in successfully segmentation of brain tumor tissues with high accuracy (95.9 ± 0.4% for database of simulated MR images, and 93.2 ± 0.3% for database of real MR images). In order to reduce the computational complexity and expedite the segmentation algorithm, and also to improve the system performance, some modifications are applied in the algorithm presented in previous algorithm. In the third algorithm, a fully automatic tumor system, which is combination of texture-based and contour-based algorithms is presented. Skippy greedy snake algorithm is capable of segmenting the tumor area; however, the algorithm’s accuracy and performance depends significantly on its initial points. Here, we modify the previous algorithm to automatically find proper initial points, which not only obviates the requirement of manual interference, but also increase the accuracy and speed of optimization convergence. Comparing with previous method, this method achieves higher accuracy in tumor segmentation (96.8 ± 0.3% for database of simulated MR images, and 93.8 ± 0.1% for database of real MR images) and lower computational complexity. The intensity similarities between brain lesions and some normal tissues result in confusion within segmentation algorithms, especially in the database of real MR images. In order to improve the system performance for this database, a multi-spectral approach based on feature-level fusion is presented in forth algorithm. Even though using multi-spectral MRI has several drawbacks and limitations, since it makes use of complementary information, it increases the accuracy of the system. Here, a feature-level fusion technique based on canonical correlation analysis (CCA) is proposed. It is worth mentioning that for the first time CCA is applied for combining MRI sequences in order to segment tumors. Even though data fusion increases computational complexity of the segmentation algorithm, it results in a higher accuracy (95.8 ± 0.2% for database of real MR images).