Deep Convolutional Routing Network with Probabilistic Hierarchical Learning for Multi-Class Brain Tumor Detection in MRI Scans

Abstract

Automated detection of brain tumors from magnetic resonance imaging (MRI) data remains a critical challenge in clinical oncology, due to the high morphological diversity of neoplastic tissue, inter-patient anatomical variability, and subtle intensity transitions at tumor margins. This paper introduces the Deep Convolutional Routing Network with Probabilistic Hierarchical Learning (DCRN-DBN), a novel two-stage architecture integrating dynamic routing capsule modules with a stacked generative Deep Belief Network for robust multi-class tumor identification. The DCRN stage replaces conventional max-pooling with agreement-based routing, preserving part-whole spatial relationships across encoding levels. Extracted capsule vectors are fused through a trainable five-level scale-attention mechanism, producing orientation-invariant, scale-comprehensive feature tensors that are processed by the DBN through contrastive divergence training for discrimination among glioma, meningioma, pituitary adenoma, and healthy tissue. Evaluated on a benchmark neuroimaging corpus, DCRN-DBN achieves 99.94% classification accuracy, 99.89% sensitivity, 99.91% specificity, and AUC values exceeding 0.9987 across all classes, outperforming ResNet-50, VGG-16, YOLO-based, and conventional CNN baselines. This research contributes to Sustainable Development Goal 3 (SDG 3) by advancing reliable non-invasive tools for early brain tumor diagnosis.