INDEREX-NET: AN ATTENTION-GUIDED MULTI-STREAM CNN FRAMEWORK WITH ADAPTIVE FEATURE FUSION FOR BRAIN TUMOR CLASSIFICATION

Abstract

Accurate classification of brain tumors from magnetic resonance imaging (MRI) data remains a challenging problem for computer-aided diagnosis systems due to inter-class similarity, intra-class variability, and heterogeneous tumor morphology. In this study, an attention-guided multi-stream deep learning architecture, termed IndereX-Net, is proposed for brain tumor classification from axial MRI images. The proposed model integrates three heterogeneous ImageNet-pretrained convolutional neural network backbones, namely ResNet50, DenseNet121, and InceptionV3, arranged in parallel streams to extract complementary multi-scale feature representations. Each stream is enhanced with a Convolutional Block Attention Module (CBAM), while Generalized Mean (GeM) pooling is employed for adaptive feature aggregation. In addition, a gated feature fusion mechanism is incorporated to dynamically regulate inter-stream interactions. The model is trained using a two-phase transfer learning strategy and evaluated on a publicly available brain tumor MRI dataset. Experimental results demonstrate that IndereX-Net outperforms individual backbone models, achieving an accuracy of 98.22% and a ROC-AUC of 0.9986, indicating that attention-guided multi-stream architectures provide an effective and reliable solution for computer-aided brain tumor diagnosis.