Introduction: Segmentation of brain tumors is crucial in proper diagnosis and the plan of treatment. Annotation is time-consuming and inaccurate using manual methods, and is incentivizing deep learning automated methods. Despite the good results of CNN and U-Net models, their performance is highly hyperparameter-sensitive to accuracy. In this patent-oriented work, a Modified Gorilla Troops Optimization (MGTO) algorithm was combined to optimize these parameters automatically and enhance the performance of segmentation on BRATS MRI scans, as well as guaranteeing greater generalization across tumor subregions. Materials and Methods: The BRATS dataset, which includes T1, T1ce, T2, and FLAIR MRI scans, was preprocessed with normalization, slice removal, and resizing. The CNN/U-Net architectures were trained to perform multi-class segmentation. The hyperparameters optimized using MGTO were the learning rate, batch size, and dropout. Dice Score fitness was applied to each candidate configuration, and the most successful one was chosen to be fully trained. The segmentation- optimized models were further evaluated for multiclass tumor classification to demonstrate the downstream diagnostic benefits of MGTO-guided feature learning. Results: The model CNN/U-Net based on MGTO was superior to the baseline architectures on all tumor regions. Dice Score, IoU, and sensitivity were significantly improved, in particular, to enhance the tumor regions that are normally hard to segment. The model was optimized and depicted less overfitting and quicker convergence. The qualitative findings also revealed more distinct tumor boundaries, and the quantitative measures ensured consistent robustness in the training, validation, and test subsets of BRATS MRI data. Discussion: To enhance the development of segmentation, MGTO successfully selected excellent hyperparameters to ensure the CNN/U-Net models could learn outstanding spatial and contextual features. The exploration-exploitation mechanism facilitated optimization more than manual search or grid search. Similar improvements in multi-dimensional tumor areas indicated the flexibility of MGTO to non-homogeneous MRI images. Results indicated that deep learning models that are informed by optimization can be used to enhance clinical reliability in automated tumor analysis. Conclusion: This patent-oriented work has shown that MGTO, utilized along with CNN/U-Net systems, can improve the brain tumor segmentation on BRATS data by a significant margin. The optimized model was more accurate and had better boundary detection and generalization to tumor subregions. The adaptive hyperparameter tuning of MGTO minimized overfitting and training stabilization. Comprehensively, the suggested framework offers an excellent basis in clinical practice in neuro-oncology, which allows for accelerated and more accurate diagnostic processes.
Mishra et al. (Mon,) studied this question.