HCAC: Trustworthy Hybrid Attentive Framework Tackling Brain Tumor Identification With XAI

More than all other types of cancer, brain tumors may affect a patient’s physical, mental, and emotional health in ways that are permanent and life-changing. A timely and accurate diagnosis is essential to improve therapy and outcomes. Despite the fact that medical imaging techniques like magnetic resonance imaging (MRI) give valuable insights, accurately identifying brain tumors within these images remains a challenge. In this work, a hybrid framework using AI-based techniques is proposed for achieving enhanced multiclass MRI brain tumour identification and classification. Specifically, Deep Learning (DL) methodologies have drawn more focus due to their excellent computational efficiency and precision in categorizing various diseases. In particular, the proposed work involves preprocessing MRI images using an Improved Rolling Guidance Filter (IRoLF) to enhance image quality, and ROIs are segmented and processed using the Modified Dilated U-Net (MDU-Net) model. Efficient features are retrieved using the Swin patch Convolutional Transformer (SwinCT). The Oppositional Osprey Algorithm (OOA) is used to select optimal features and reduce complexity. Then, classification is accomplished by means of the proposed hybrid Dwarf Attention Capsule Routing Autoencoder network model, named HCAC. Moreover, reliability is also an essential need when dealing with medical imaging. Thus, in the proposed study, Explainable AI (XAI) methods like Gradient-based Class Activation Mappings (Grad-CAM), Local Interpretable Model-Agnostic Explanations (LIME), and Shapley Additive Explanations (SHAP) are utilised to understand the model’s reasoning. The proposed model has been validated considering three well-known publicly available MRI datasets, and eleven methods from the state-of-the-art are also considered for comparison. The outcomes indicate that in brain tumour classification situations, the proposed HCAC performs considerably better than existing methods. Specifically, the Figshare, Brain Tumor MRI, and GBM MRI datasets were considered for testing our proposal, achieving high accuracy levels of 99.9%, 99.04%, and 99.25%, respectively, for each of the three tested medical scenarios.