فاطمه آقاگلي

Despite significant advances in the field of tumor detection in medical images, achieving an accurate and reliable detection is still an important and challenging task for the success of treatment, saving the patient's life and reducing costs in the shortest possible time. In this thesis, for the first time, an innovative method is proposed for the accurate and automatic detection of tumors and their localization in medical images using local dimension reduction after reducing the size of the original image. For this purpose, an algorithm is proposed called the extension of modified mixture of factor analyzers model based on co-clustering and dynamic co-clustering for gray medical images. First, the proposed algorithm optimizes the number of row and column clusters simultaneously by estimating model parameters based on tumor size. Then, using the obtained optimal values, the location of the rows and columns is changed by clustering them simultaneously until a block containing the tumor is detected. In this process, the image is segmented into a certain number of blocks, one of which contains the tumor. The number of blocks is the Cartesian product of the optimal number of row clusters and the optimal number of column clusters. Finally, the detected block is binarized using minimum error thresholding and localized on the input image by the proposed algorithm, while the rest of the image is considered as background. The performance of the proposed algorithm is eva‎luated on mammography images from MIAS and DDSM datasets as well as brain MRI images from BraTS2018, BraTS2019 and BraTS2020 datasets. The results show the high accuracy of the algorithm in the fast and very accurate detection of tumors with different sizes and locations in high-dimensional and complex images. Also, in this thesis, the preprocessing of raw mammography images and contrast enhancement of medical images is discussed using a novel algorithm. Since the ground truth for the pectoral muscle, additional margins and labels are not available in mammography images, to eva‎luate the performance of the algorithm in the preprocessing step, the data augmentation method is used to simulate the ground truth for pectoral muscle, and novel criteria are defined to eva‎luate the performance of the algorithm in removing additional margins and labels. The results show that the proposed algorithm is able to completely and accurately remove the unnecessary areas and extract the desired area and enhance the image quality without causing adverse effects. Also, this algorithm shows well its stability against noise in the preprocessing step.

تحليل عاملي آميخته اصلاح‌شده , هم‌خوشه‌بندي , هم‌خوشه‌بندي پويا , تشخيص تومور , ‌پيش‌پردازش و بهبود كنتراست , شبيه‌سازي عضله سينه , روش افزايش داده

Modified mixture of factor analyzers , co-clustering , dynamic co-clustering , tumor detection , preprocessing and contrast enhancement , pectoral muscle simulation , data augmentation technique

Fatemeh Aghagoli

Dr. Rahman Farnoosh