Precision Brain Tumour Diagnosis with Convolutional Recurrent Attention Networks

Abstract

Histological heterogeneity and complex imaging patterns make brain tumour diagnosis difficult. In order to make timely clinical decisions and plan personalised treatment, MRI multi-class brain tumour classification must be accurate. Expert evaluation in conventional image interpretation increases diagnostic burden and inter-observer variability. While current AI-based tumour classification methods have improved, learning discriminative spatial-contextual representations and giving clinically relevant model interpretability remain challenges. This article offers Convolutional Recurrent Attention Network (CRANet) for multi-class brain tumour classification using T1-weighted MRI data to overcome these issues. Convolutional Neural Networks (CNNs) extract hierarchical spatial features, Recurrent Neural Networks (RNNs) model contextual dependencies among deep feature representations, and a Channel-Spatial Attention Module emphasises diagnostically informative features during classification in the proposed architecture. Cross-validation evaluates generalisation performance, whereas noise reduction, intensity normalisation, and data augmentation improve data quality and model resilience. Experiments show that CRANet outperforms state-of-the-art deep learning methods in conventional diagnostic measures. Attention activation maps improve model transparency and clinician interpretation by visualising picture areas contributing to classification decisions, without localising tumours.