Bridging modalities: a deep learning framework for brain tumor classification via CT-MRI integration and model fusion

Artificial intelligence (AI) and machine learning (ML) have shown remarkable promise in advancing medical image analysis, yet their potential in neurology and psychiatry remains underexplored. This work explores the use of deep learning approaches for automated brain tumor classification, leveraging multimodal neuroimaging data comprising computed tomography (CT) and magnetic resonance imaging (MRI) scans. Two model families were evaluated: a custom CNN trained from scratch and a transfer-learning approach based on ResNet-18. Models were trained and validated separately on CT and MRI datasets, and further extended to a combined dataset through multimodal fusion. Experimental results demonstrate that the CNN achieved accuracies of 97 and 99% on CT and MRI datasets, respectively, outperforming ResNet18, which yielded 95 and 97% under the same settings. On the combined dataset, CNN maintained superior performance (98%) compared to ResNet18 (94%), highlighting the adaptability of CNNs to domain-specific features in medical imaging. These findings suggest that lightweight CNNs can be highly effective for neuroimaging-based tumor detection, particularly when multimodal data are leveraged. Beyond clinical utility in early diagnosis, the authors underscore the importance of exploring modality-specific characteristics and model adaptability in designing AI-driven diagnostic systems for neurological disorders.