Multiscale MoE: A Mixture of Experts Framework with Attention-Driven Multi-Scale Learning for Brain Tumor Classification

The impact of brain tumors as a global health concern is due to their aggressive behavior, high mortality, and complexities in their diagnosis.While MRI remains the gold standard for identifying, monitoring, and detecting brain tumors, automated classification methods encounter many complications with respect to the diverse morphologies of tumors, similarities in their imaging features, and the potential variability in imaging conditions.CNNs can capture spatial hierarchies, but cannot generalize effectively and ViTs rely on the context to characterize the image modalities which means that, whilst they address some deficiencies of CNNs, they require extensive data and computational resources.To remedy some of the issues that each approach presents, we present multiscale MoE that leverages CNNs and attention-oriented modules.The proposed architecture uses multi-scale feature extraction, channel-spatial attention, and dynamic expert routing, which adequately collects tumor-specific features efficiently.We applied two different publicly available datasets, namely the Bangladesh Brain Cancer MRI and Figshare Brain Tumor dataset.For the Bangladesh dataset, the proposed model achieved overall accuracy of 96.92% and for FigShare dataset, the highest results achieved 96.42% accuracy.In contrast to state-of-the-art models, multiscale MoE achieved the highest testing accuracy 96.14%, and the lowest Brier score 0.0603.The proposed model has shown to have balanced classification results across the tumor classes and reduced the number of false predictions whilst maintaining efficient computational performance and thus has the potential to provide a valuable resource for clinical practice with respect to real-time applications.