DC-CNN-SE: A DEEP CONVOLUTIONAL CAPSULE NETWORK WITH SQUEEZE-AND-EXCITATION AND BAYESIAN LEARNING FOR MULTI-MODAL MRI BRAIN TUMOR CLASSIFICATION

Abstract

The classification and segmentation of brain tumors in multi-modal MRI scans are highly complex tasks due to variations in shape, size, and appearance across different imaging modalities. This paper presents a novel Deep Convolutional Capsule Neural Network with Squeeze-and-Excitation (DC-CNN-SE) blocks, incorporating a Bayesian uncertainty module and prototype-guided representation learning for classifying brain tumors in multi-modal MRI. By combining T1, T2, T1C, and FLAIR images, our model uses squeeze-and-excitation blocks to recalibrate feature maps, emphasizing modality-specific tumor features. Additionally, a Bayesian uncertainty quantification mechanism provides reliable prediction confidence estimates, addressing clinical concerns about misclassification. Moreover, a prototype-guided representation learning approach is included to improve the separability of features in the latent space, boosting generalization across patients. The architecture, enhanced with deep convolutional layers, Leaky ReLU activations, batch normalization, and dropout, achieves excellent performance on the BraTS2020 dataset, with accuracy (99.00%), precision (98.95%), recall (98.91%), F1-score (98.90%), MCC (98.65%), and CSI (97.84%) metrics are used for segmentation and classification of MR Images. This framework marks a significant step forward in clinical decision-making for brain tumor diagnosis. Both qualitative and quantitative results demonstrate that our proposed model is effective and surpasses current state-of-the-art (SOTA) methods.