Optimizing Brain Tumor Classification with GridSearchCV Using Multimodal Image Fusion: A Hyperparameter Tuning Approach

Brain tumor classification to reliably detect brain cancers or diseases accurately is an important goal of early detection and treatment planning of a brain tumor. However, the classification accuracy of such models has been previously challenged because they tend to display irregularities in the appearance of various imaging conditions. This paper proposes a deep learning architecture for image fusion based on the combination of MRI and PET image acquisition, which provides better structural and metabolic representations for brain tumors. One of the contributions of this work is a systematic hyperparameter tuning approach by GridSearchCV to optimize learning rate, batch size, optimizer selection, and network depth of Convolutional Neural Networks (CNNs) trained on fused images. The evaluation results on the BraTS and TCIA datasets show that the proposed architecture provides an effective performance improvement over commonly used untuned models and conventional analog imaging approaches. Based on classification evaluation the model optimized obtains 93.5% higher classification accuracy, 92.8% precision, 94.0% recall, 93.2% F1-score, and 95.2% ROC-AUC, lower misclassification rates, and predictive reliability. Our study suggests that the proposed method is feasible in terms of clinical application.