Exploring Radiologists' Reluctance Towards Machine Learning Models and Explainable AI in Brain Tumor Detection

Machine Learning (ML) models have made significant advancements in medical sciences, yet their application in clinical practice is hindered by concerns over transparency and accountability. Explainable Artificial Intelligence (XAI) emerges as a solution to these challenges, offering interpretable insights into ML model predictions for clinicians. This research investigates the extent to which existing XAI techniques support decision-making in brain tumor detection for radiologists. A highly accurate ML model based on the VGG16 architecture was developed for detecting brain tumors from MRI images, achieving training and testing accuracy of 99.46% and 96.28%, respectively. A series of experiments were conducted integrating various XAI techniques. The study reveals, despite high model accuracy, false positives can mislead clinical decisions, especially when datasets contain multiple MRI sequences with conflicting diagnostic information. The findings, validated by expert radiologists, underscore the importance of using 3D imaging or at least consecutive 2D slices from the same sequence to improve diagnostic accuracy. The paper presents the experimental details highlighting the critical gap between 2D MRI slice-based diagnosis and the need for comprehensive 3D analysis. The research concludes with recommendations for improved dataset design and the integration of XAI to enhance the reliability and accountability of AI -driven healthcare diagnostic systems.