| چکیده انگلیسی مقاله |
Introduction: Brain tumors are caused by abnormal cell proliferation and can lead to neurological disorders by affecting the structure and function of the brain. Therefore, their accurate and timely diagnosis plays a significant role in reducing clinical risks for patients. Magnetic resonance imaging, as a non-invasive and highly accurate method, is widely used in identifying tumor areas in the diagnosis and treatment planning process.
Materials & Methods: This study is a qualitative systematic review based on the PRISMA guideline, which was conducted by comprehensively searching the Web of Science, Google Scholar, Springer, Scopus, IEEE Xplore, and Elsevier scientific databases. Studies related to convolutional neural networks (CNN) for brain tumor segmentation in Magnetic Resonance Imaging (MRI) images, focusing on advanced architectures including U-Net, nnU-Net, V-Net, DeepMedic, and DeepLabV3+, were selected based on specific inclusion and exclusion criteria and subjected to qualitative and comparative analysis.
Results: The results of the reviewed articles showed that the DeepLabV3+ model had the highest accuracy with an average Dice score of 0.917 and the other models U-Net, nnU-Net, V-Net, and DeepMedic had average Dice scores of 0.827, 0.793, 0.819, and 0.752, respectively. All of these models performed better than manual or traditional methods in different data conditions such as 2D, 3D, and unbalanced data. However, the reported performance for each model is affected by several factors, including the quality and volume of training data, data augmentation strategies, the loss function used, and post-processing steps.
Conclusion: This study shows that novel image analysis methods significantly improve the accuracy of diagnostic assistance systems in brain imaging by automatically extracting diagnostic features. However, the performance of the models is the result of a combination of the main architecture and complementary techniques, and their evaluation should be performed within the overall framework of the analysis pipeline (from preprocessing to post-processing). Developing models with high generalizability in diverse data conditions is the main path of progress in this field. Given the time-consuming and complex nature of manual interpretation of MRI images, deep learning-based systems can help reduce human errors and facilitate clinical decision-making. Consequently, optimization and development of these models is an important step towards improving the diagnosis and management of patients with brain tumors. |
| نویسندگان مقاله |
هاجر دانش | Hajar Danesh
Dept of Electrical and Biomedical Engineering, Faculty of Engineering and Technology, Shahid Ashrafi Esfahani University, Isfahan, Iran
گروه برق و مهندسی پزشکی، دانشکده فنی و مهندسی، دانشگاه شهید اشرفی اصفهانی، اصفهان، ایران
زهرا قاسمی | Zahra Ghasemi
Dept of Electrical and Biomedical Engineering, Faculty of Engineering and Technology, Shahid Ashrafi Esfahani University, Isfahan, Iran
گروه برق و مهندسی پزشکی، دانشکده فنی و مهندسی، دانشگاه شهید اشرفی اصفهانی، اصفهان، ایران
|