BTS-ADCNN: brain tumor segmentation based on rapid anisotropic diffusion function combined with convolutional neural network using MR images

• Published in: The Journal of supercomputing Vol. 80; no. 9; pp. 13272 - 13294
• Main Authors: Mbarki, Zouhair, Ben Slama, Amine, Amri, Yessine, Trabelsi, Hedi, Seddik, Hassene
• Format: Journal Article
• Language: English
• Published: New York Springer US 2024; Springer Nature B.V
• Subjects: Algorithms; Artificial neural networks; Brain; Brain cancer; Compilers; Computer Science; Filtration; Interpreters; Magnetic resonance imaging; Medical imaging; Metric space; Neural networks; Processor Architectures; Programming Languages; Tumors; Brain tumor; MRI datasets; U-Net architecture; Anisotropic diffusion
• ISSN: 0920-8542; 1573-0484
• DOI: 10.1007/s11227-024-05985-2

Brain cancer is a fatal and debilitating condition that has a profoundly negative impact on patients' lives. Therefore, early diagnosis of brain tumors enhances the effectiveness of treatment and raises patient survival rates. However, it is a challenging task and an unmet need to identify brain tumors in their early stages. In the presented work, a rapid and efficient algorithm for tumor segmentation that supports doctors in the practice of screening brain tumors is proposed. The proposed method is divided into two phases: Firstly, a preprocessing operation is performed using an anisotropic diffusion filtering function for noise removal with details and edge conservation. The second phase is a segmentation operation of brain tumors based on deep convolutional neural network. Simulation results on reel data approve the efficiency of the proposed method. In fact, the combined filtering and segmentation methods have improved the segmentation results of Dice similarity coefficient (Dice = 89.65 ± 0.81%), Hausdorff distance (95%) (HD95) = 7.53, and Intersection over Union value (IOU = 90.12 ± 0.76%) using a set of 520 MR images divided into: 364 images for the training and 156 images for the validation. Compared to different recent segmentation methods, the proposed technique offers an advanced performance by detecting Glioblastoma tumor regions. The obtained results are very interesting and prove the efficiency of the proposed algorithm compared to other recent works in the literature.