|
|
 |
| |
|
|
|
A Novel Brain Tumor Segmentation Approach Based on Combining Fine-Tuned Deep Learning and Augmentation Techniques |
|
|
|
PP: 357-370 |
|
|
doi:10.18576/amis/200205
|
|
|
|
Author(s) |
|
|
|
Hanan H. Amin,
E. A. Zanaty,
Walaa M. Abd-Elhafiez,
|
|
|
|
Abstract |
|
|
| Recently, brain tumor diagnosis has increasingly relied on medical image analysis, providing essential inputs for diagnosis, prognosis, and treatment planning. Deep learning models are an important method used to predict patient progression, impacting the selection of effective medical prescriptions. Deep learning models have achieved impressive results in automatic classification, but diverse datasets face challenges in achieving high generality and robustness. In this paper, we propose a new approach that combines fine-tuning of hyperparameters with advanced computational scripting techniques to increase segmentation accuracy. In this case, we adopted a modified FCNN architecture to systematically adjust learning speeds, batch size, dropout, and optimizer parameters. We incorporated data scripting strategies such as rotation, flipping, density variation, and elastic deformation to increase dataset diversity. This combination helps improve the performance of deep learning models when applied to a wide variety of datasets and study samples. The proposed method was evaluated on the Brats 2020 dataset and compared to existing methods, including Patchnet, Deeplabav3, and Baseline FCNN. The results indicate that our augmented model improves state-of-the-art methods in terms of Dice coefficients, accuracy, and memory. The results demonstrate the effectiveness of combining computer-centric and model-centric adaptation strategies for improving brain tumor segmentation. |
|
|
|
|
|
|
|
