An Optimized Capsule Network Architecture for Brain Tumor Classification with Comparative Insights from Transfer Learning Models

Accurate classification of brain tumors from MRI scans is vital for early diagnosis and effective treatment planning. This study proposes an optimized deep learning approach based on Capsule Networks (CapsNet) for robust brain tumor classification. The dataset used includes 7,023 MRI images categorized into four types: glioma, meningioma, pituitary tumor, and no tumor. The proposed CapsNet model emphasizes the preservation of spatial relationships and pose information, addressing limitations found in traditional CNNbased models. For comparative analysis, three baseline models—custom CNN, VGG16, and DenseNet121 were also trained and evaluated under identical experimental conditions. All models were assessed using accuracy and confusion matrix metrics. The results demonstrate that the CapsNet model achieves superior classification accuracy (97%) compared to CNN (89%), VGG16 (94%), and DenseNet121 (92%). These findings highlight the potential of CapsNet in advancing medical image classification, particularly in scenarios requiring spatial hierarchy preservation. This work supports the application of CapsNet in computer-aided diagnosis systems for brain tumor detection.