A minimal-net CNN model for an IoT-based brain tumor detection and monitoring system.

In this study, we propose M-Net, a lightweight Convolutional Neural Network (L-CNN) designed for real-time brain tumor classification and detection. Large, state-of-the-art (SOTA) CNN models often face deployment challenges on edge devices due to their high computational resource requirements. To address this, M-Net utilizes a sequential architecture with 4 × 4 kernels and progressive filter expansion, balancing computational efficiency and classification performance. The model was trained and tested on three diverse MRI datasets (3-class, 4-class, and 15-class) to evaluate its robustness and generalizability across varying brain tumor modalities. In k-fold cross-validation (k = 5) and on an unseen test set, M-Net achieved an average accuracy of 99%, outperforming SOTA CNNs (96%). The study also integrates explainable AI techniques-LIME, SHAP, and Grad-CAM-to visualize regions of MRI images contributing to the model's decision-making process. Additionally, a web and mobile application for smart brain tumor management (SBTM) was developed using M-Net, demonstrating real-time processing at 1 FPS, minimal thermal consumption, and low energy usage. The novel contributions of M-Net lie in its efficient architecture and integration with XAI, making it a suitable candidate for clinical deployment in brain tumor diagnosis. With only 495,972 parameters, M-Net offers a practical, deployable solution for real-time tumor classification on edge devices.