NeuroScanXNet: An Explainable and Evidence-Grounded Framework for Brain Tumor Classification

Abstract

Deep learning has been widely used in medical imaging, significantly improving the accuracy of brain tumor classification. However, many existing models focus primarily on prediction accuracy without explaining how decisions are made, making their deployment in real clinical settings challenging. Convolutional Neural Networks (CNNs), though effective, are often treated as black-box models, which makes it difficult to trust their outputs. This study proposes a framework — NeuroScanXNet — that addresses both classification accuracy and result interpretability. A CNNbased model is used to classify MRI images into various tumour categories. Grad-CAM is applied to highlight the important regions influencing the model's predictions. In addition, a quantitative consistency analysis using masked cosine similarity is introduced to evaluate whether the model focuses on similar regions across different inputs. A Mini-RAG module based on TF-IDF retrieves relevant medical knowledge to generate a structured diagnostic report. The proposed system achieves a test accuracy of 94.66% while improving transparency and usability for real-world clinical decision support

Keywords