Using image processing and deep learning for automatic brain tumor detection in MRI images
Using image processing and deep learning for automatic brain tumor detection in MRI images
Mahsa Ghasemi Samani1 Narges Karimzadeh Dehkordi2
1) Department of Biomedical Engineering, Shahrekord Branch, Islamic Azad University, Shahrekord, Iran
2) Department of Biomedical Engineering, Shahrekord Branch, Islamic Azad University, Shahrekord, Iran
IOI Code
:
AADK-AAHGFF
Publication :
3rd International Congress on Science, engineering & New Technologies(secongress.com/3rd)
Abstract :
Automatic brain tumor detection from MRI images using image processing and deep learning is one of the prominent challenges in medical sciences. This research focuses on analyzing and comparing various deep learning models for brain tumor detection, with an emphasis on the impact of preprocessing, image size, and model architectures. Results obtained from the BRATS 2020 and BraTS 2021 datasets demonstrated that DenseNet-121 and ResNet-50 models outperformed simpler models like VGGNet-16 in terms of accuracy. In particular, the DenseNet-121 model achieved the best performance with an accuracy of 94.3% in identifying different types of brain tumors (glioma, astrocytoma, and meningioma), while the ResNet-50 model, with an accuracy of 91.2%, ranked second. Preprocessing results showed that noise reduction with Gaussian filters, signal intensity normalization, and data augmentation had a significant impact on improving model accuracy. Specifically, data augmentation improved model accuracy by 8% and helped prevent overfitting. T2-weighted MRI images had the highest accuracy of 92.1%, while T1-weighted MRI images showed an accuracy of only 86.5%. Compared to traditional methods such as manual thresholding, deep learning models demonstrated significantly better performance. For example, the DenseNet-121 model achieved an accuracy of 94.3% in detecting brain tumors, compared to 75.2% accuracy with traditional methods. This difference highlights the ability of deep learning models to learn complex and nonlinear features from MRI images. Additionally, model interpretability analysis using techniques like Grad-CAM and LIME showed that these models could more transparently identify tumor-related regions and provide more understandable decisions for doctors. This feature helps clinicians utilize the model s results more confidently in clinical processes. Ultimately, the results of this study suggest that the use of complex deep learning models, combined with precise preprocessing and optimal data selection, can serve as reliable tools for automatic brain tumor detection.
Keywords : image processing deep learning brain tumor detection mri images densenet 121 resnet 50 image preprocessing glioma astrocytoma