EXPLAINABILITY OF BRAIN TUMOR CLASSIFICATION MODEL BASED ON INCEPTIONV3 USING XAI TOOLS

Detection and diagnosis of brain tumors swiftly and accurately using noninvasive approaches are essential for the prognosis of the patients. Magnetic resonance imaging (MRI) scans combined with artificial intelligence (AI)-based algorithms provide invaluable insights regarding patient health and disease information, leading to a breakthrough of AI in healthcare. Although the model is mathematically explained, lack of trust and explainability of black-box algorithms pose a major challenge for deployment. Determining trust in the AI system helps users to deploy the model in a clinical setup and to take actions based on the predictions in real time. In this paper, a high-performing deep learning (DL) model using transfer learning based on InceptionV3 is applied to classify images of brains into different types of tumors: meningioma, glioma, and pituitary. To address the trust concerns, this study uses open-source explainable artificial intelligence (XAI) techniques to provide after-the-fact (post hoc) explanations using libraries such as local interpretable model-agnostic explanations (LIME), shapley additive explanations (SHAP), and gradient-weighted class activation mapping (Grad-CAM). The InceptionV3-based model achieved an accuracy of 93% and an F1 score of 0.93. The XAI techniques achieved explainability of nearly 60% using SHAP compared with LIME and Grad-CAM explaining less than 50% of the cases using the relevant tumor features as expected by medical practitioners. The experimental study concludes that in the ML-based classification process, there is minimal impact of the features related to the tumor and heavy influence of non-relevant features.