XAI-BT-EdgeNet: Explainable Edge-Aware Deep Learning with Squeeze-and-Excitation for Brain Tumor Detection and Prediction

Deependra Rastogi^1*Prashant Johri²Massimo Donelli^3,4Tarun Agarwal⁵Shrikant Tiwari²Pushpa Singh¹

• ¹School of Computer Science and Engineering, IILM University, Greater Noida, India
• ²School of Computer Science and Engineering, Galgotias University, Greater Noida, India
• ³Department of Civil, Environmental, Mechanical Engineering University of Trento, Trento, Italy
• ⁴Radiomics Laboratory, Department of Economy and Management, University of Trento, Trento, Italy
• ⁵Department of Computer Science and Engineering, JIIT, Noida, India

Accurate and early detection of brain tumors is critical for effective treatment and improved patient outcomes, yet manual radiological analysis remains time-consuming, subjective, and error-prone. To address these challenges and improve clinical trust in AI systems, this study presents XAI-BT-EdgeNet, an explainable, edge-aware deep learning framework integrated with Squeeze-and-Excitation (SE) modules for brain tumor detection using MRI scans. The architecture employs a dual-branch design that fuses high-level semantic features from InceptionV3 with low-level edge representations via an Edge Feature Block, while SE modules adaptively recalibrate feature importance to enhance diagnostic accuracy. To ensure transparency, the model incorporates four XAI techniques—LIME, Grad-CAM, Grad-CAM++, and Vanilla Saliency, which provide interpretable visual justifications for predictions. The framework was trained and evaluated on the Brain Tumor Dataset by Preet Viradiya, comprising 4,589 labeled MRI images divided into Brain Tumor (2,513) and Healthy (2,076) classes. The model achieved outstanding results, with 99.58% training accuracy, 99.71% validation accuracy, and 100.00% testing accuracy, alongside minimal loss values of 0.0103, 0.0051, and 0.0026, respectively. This work includes, the development of a dual-branch CNN architecture that combines semantic and edge features for enhanced classification, the integration of SE modules to highlight clinically significant regions, and the application of multi-method XAI to offer transparent, interpretable outputs for clinical applicability. Overall, XAI-BT-EdgeNet delivers a high-performing, interpretable solution that bridges the gap between deep learning and trustworthy clinical decision-making in brain tumor diagnosis.