A CNN-based method with capuchin search algorithm-based weighted constrained optimization for brain tumor classification

Early detection of brain tumors is crucial for improving patient survival rates and treatment options. Accurate classification and stratification of brain tumors are also critical for developing individualized treatment plans. Despite the increasing use of Magnetic Resonance Imaging (MRI) for brain evaluation and advances in AI-based detection techniques, building an accurate and efficient model to identify and classify brain tumors from MRI images remains a challenge. To tackle this issue, this study develops a deep Convolutional Neural Network (CNN)-based structure to automatically classify brain tumors into four prevalent groups: meningiomas, pituitary tumors, non-tumor tumors, and gliomas. To this end, the proposed deep CNN model utilizes a segmentation model and a preprocessing approach combined with the Capuchin Search Algorithm (CSA) to improve image contrast. These methods demand super-computing power and real-time performance, as well as parallel or distributed processing to further enhance their effectiveness. The proposed CNN model-based structure is used in the classifier to improve the diagnostic procedure for tumor classification. Using four broadly available reference datasets of varying complexity, along with tumor regions exhibiting varying degrees of variability, we trained the segmentation model and assessed the classification model. This enables us to perform a side-by-side comparison of the effects of the segmentation process on tumor classification. The efficiency level of the presented classification method was evaluated using many related metrics. On all four adopted datasets, the developed deep learning-based classification model performs better than many pre-trained models. The results showed that the proposed classification model achieved a maximum classification accuracy of 97.64% on dataset 1 with preprocessing and 96.27% without. The highest classification accuracy was 99.58%, reported on dataset 4. Thus, this proposed framework can be used with high efficiency in clinical settings to automatically identify and segment brain tumors from MRI images.