Enhanced Brain Tumor Detection with Hybrid Approaches and Explainable AI

Abstract

Accurate brain tumor classification poses significant challenges due to cellular diversity, complicating reliable radiological diagnoses. Recent advancements in MRI have enhanced CADS for tumor detection. However, DL models struggle to extract significant characteristics from medical images, differing substantially from natural Images. To overcome this limitation, it uses multiple stages feature extraction and a XAI technique to assess the effectiveness of hybrid DL and ensemble ML in brain tumor detection. This work employs a two-pronged strategy to improve brain tumor categorization. Initially, ResNet50 extracts features, which are then optimized using GSMVO and MI. Classification is performed using either a MLP or an ensemble of Random Forest and XGBoost models. Analysis of SIAR dataset MRI images confirms the hybrid DL model’s 95.5% accuracy, outperforming the ensemble model’s 93.5%. Furthermore, XAI techniques - Grad-CAM, LIME, SmoothGrad with Guided Backpropagation enhance interpretability, fostering clinical trust and transparency. The proposed hybrid approach emerges as a reliable diagnostic tool for brain tumors.