Four-Class Brain Tumor Classification Using ConvNeXt with Grad-CAM-Based Explainable

Abstract

Brain tumor is one of the most critical diseases with high mortality rates requiring early and accurate diagnosis. This study proposes a deep learning-based brain tumor classification system using the ConvNeXt-Base architecture to classify four categories: glioma, meningioma, pituitary tumor, and no tumor. To further advance this field, this study addresses two open challenges in prior literature, the need for larger and more diverse datasets to ensure model robustness, and the demand for model transparency (resolving the black-box problem) to facilitate clinical adoption.. To address these, this study combines MRI images from three public sources, Figshare (Cheng et al., 2016), Br35H (Hamada, 2020), and Mendeley (Hira et al., 2025), totaling 11,474 unique images after perceptual hash-based deduplication. The model was trained using progressive unfreezing with AdamW optimizer, CosineAnnealingLR scheduler, and class-weighted cross-entropy loss. Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated as an Explainable AI (XAI) approach to visualize the regions influencing model decisions. The proposed ConvNeXt-Base model achieved an accuracy of 99.25%, macro precision of 99.23%, macro recall of 99.22%, macro F1-score of 99.22%, and macro AUC-ROC of 0.9995 on the test set. Grad-CAM visualizations confirm that the model focuses on clinically relevant tumor regions, particularly the sella turcica for pituitary tumors and distinct mass boundaries for meningioma, thereby improving interpretability and clinical trustworthiness of the AI system.