Intracranial tumors are life-threatening and severe health hazards attributed to the unregulated and excessive growth of brain cells; thus, a diagnosis needs to be made at an early and accurate stage to achieve successful treatment. As much as medical practitioners tend to involve Brain MRI in diagnostic purposes, some cases of oversight or misdiagnosis may be caused by human interpretation. Machine learning-based detection can be effectively used to reduce this issue. In medical imaging experiments, especially in essential medical fields, such as brain tumor classification, the method of feature selection is also vital in enhancing the effectiveness and performance of models. The suggested study applies a pre-trained transfer learning model (MobileNet) to extract deep features, namely, textures, edges, lines, and shapes of objects. However, such high-dimensional features are effective; they can involve redundant and irrelevant factors that negatively impact on the model performance. The best tool in establishing the most pertinent subset of features is a powerful metaheuristic optimization plan. This is useful to reduce dimensionality efficiently as well as preserve the ability to discriminate. The proposed research offers a new hybrid system that brings to bear the TL with a nature-inspired Particle Swarm Optimization (PSO) algorithm used to select the best features based on the Brain MRI data. The proposed PSO-based feature selection algorithm proves good results when used in a Support Vector Classifier with a score of the R2 as 93.7211%, a precision reached 95.2828%, an accuracy of 95.0995%, a recall as 95.0995%, and an F1-score as 95.0996%, which shows the effectiveness of the proposed framework for classifying the brain tumor.
