چكيده به لاتين
Brain tumor detection has been one of the most critical and competitive issues for researchers in the medical field over the years. Numerous methods have been developed to distinguish between normal and abnormal tissues in MR images. Unsupervised methods, particularly clustering, have been widely employed in these studies. However, Co-Clustering, which refers to the simultaneous clustering of rows and columns of a matrix, has rarely been used in these issues due to certain limitations. This has prevented the exploitation of Co-Clustering’s advantages, such as high execution speed and superior ability to detect similar patterns in matrix data. After applying a Co-Clustering algorithm on matrix data, a new matrix with block-shaped co-clusters is generated. The block-shaped Co-Clusters make these methods ineffective in segmenting tumors, which come in various shapes. Moreover, after execution, Co-Clustering algorithms alter the positions of pixels in the original matrix based on their inclusion in the Co-Clusters, which further weakens the performance of these methods in accurately localizing tumors, an essential aspect of detection. The aim of this thesis is to address these limitations and modify Co-Clustering methods to enhance their applicability in tumor detection. Two algorithms are proposed for this purpose.
The first algorithm, called Iterative Co-Clustering and K-Means, uses a Latent block model for Co-Clustering and applies it iteratively. By integrating the K-Means clustering method, the algorithm performs tumor segmentation. The results of this method, along with comparative analysis on the BraTS 2019 dataset using various evaluation metrics, are presented. For accuracy and Dice similarity coefficient, this method achieves 99.28% and 84.87%, respectively, demonstrating its strong performance compared to other methods, particularly in detecting small and challenging tumors.
The second method, called Iterative Spectral Co-Clustering and Fuzzy C-Means, introduces a novel perspective to this iterative structure, called pseudo-deep structure. The iterative or pseudo-deep structure of the algorithm enhances its accuracy and performance with each iteration or layer, especially for complex images with small tumors. This algorithm also includes a method for relocating the displaced pixels and identifying their original positions. Spectral Co-Clustering is used in each layer, while the Fuzzy C-Means method is employed for tumor segmentation. The results of the algorithm’s performance on the BraTS 2020 and 2021 datasets are presented. For BraTS 2020, accuracy and Dice similarity coefficient are 99.12% and 81.42%, respectively, while for BraTS 2021, these values are 99.21% and 82.03%, respectively. These results, along with other presented evaluation metrics, demonstrate the proposed method's efficiency and high speed in tumor segmentation and localization.
