EXPLAINABLE AI-BASED MULTI-SCALE FEATURE FUSION FOR ACCURATE TUMOUR DETECTION AND CLASSIFICATION IN RADIOLOGICAL IMAGES

Abstract

Detection and classification of brain tumors are essential in medical imaging, as early and accurate diagnosis aids in treatment planning and patient outcomes. Magnetic Resonance Imaging (MRI) is a common imaging technique used in medicine, especially for brain imaging. Magnetic Resonance Imaging (MRI) is a medical imaging technique used for brain scanning, which is time-consuming and traditionally highly dependent on expert interpretation. While many AI models have been developed for the automatic detection of tumors, recent achievements in Artificial Intelligence (AI) and deep learning have shown great potential; however, many available models lack interpretability and fail to leverage the multi-level features of images. For this reason, this study presents an Explainable AI (XAI) framework for accurate brain Tumor detection and classification using MRI images. The study employed the Brain Tumor MRI Dataset, which contains 7,200 MRI images belonging to four classes (Glioma, Meningioma, Pituitary tumor, and No tumor). To enhance the model's performance, the images underwent preprocessing techniques such as resizing, normalization, and data augmentation. A basic CNN model and a CNN with High-Quality Gradient Class Activation Maps (CNN-XAI) were developed, both using batch normalization, dropout, Early Stopping, and ReduceLROnPlateau. Four metrics, accuracy, precision, recall, and F1-score, along with ROC-AUC analysis, a confusion matrix, and an interpretability assessment, were used to evaluate performance. The experimental results showed that the CNN baseline model achieved the highest classification accuracy of 87.44%, with Precision, Recall, and F1-score values of 88.69%, 87.44%, and 87.13%, respectively. The proposed CNN-XAI framework achieved 82% accuracy and improved transparency and interpretability by localizing the tumor with Grad-CAM. Both models demonstrated good discriminative ability, with near-perfect performance for the pituitary and non-tumor classes in ROC-AUC analysis. Glioma tumors continued to be the most challenging category owing to their imaging heterogeneity. These results suggest that CNN-based deep learning models contribute to automated brain tumor classification, and that explainable AI techniques provide a stronger understanding of and trust in the clinical application of these models. The proposed framework demonstrates the potential of combining deep learning and explainability for reliable, interpretable AI-assisted diagnosis in radiological imaging systems.