EDGE AI FRAMEWORK FOR BRAIN TUMOR CLASSIFICATION USING GOOGLENET FEATURES AND MACHINE LEARNING

Abstract

Brain tumor classification is a pivotal process in medical imaging. The precision and promptitude of a diagnosis will dictate treatment and affect patient prognosis. Here, I propose a two-phase framework for the automation of Brain Tumor Detection which integrates Feature Extraction and Classification using Machine Learning Algorithms. The aim is to extend the concept of Edge AI and run the system on embedded devices for real-time clinical applications. In the first phase of the project, I designed a feature extraction system using the GoogLeNet (InceptionV3) model and evaluated a range of classifiers: SVM, MLP, XGBoost, LightGBM, Random Forest, AdaBoost, K-NN, and Softmax on the CE-MRI dataset. Results showcase the system performance, as well as GoogLeNet+ SVM being the best performing model overall with precision, recall, F1-score, and ROC-AUC measures. As such, this model is the best predictor of the classifiers designed. Phase 2 involved deploying the optimized models on NVIDIA Jetson Orin Nano, the embodiment of edge-AI. Advanced quantization (FP16/INT8) and pruning helped to reduce complexity while sustaining accuracy. A simplified GUI was designed to show input MRI slices, the predicted tumor class (glioma, meningioma, or pituitary), confidence scores, and device metrics including latency, throughput, power consumption, and temperature. Live demonstrations proved the framework’s self-sufficiency, eliminating the need for remote (cloud) resources, allowing rapid, bedside feedback with increased privacy, and reduced latency. Even though many classifiers showed strong diagnostic accuracy in Phase 1, in Phase 2 only a few classifiers, specifically GoogLeNet + SVM and GoogLeNet + MLP, combined the accuracy with the low latency, high efficiency, and thermal stability necessary for edge deployment. The system’s practicality was further demonstrated by the GUI, which in real time, successfully classified even the pituitary tumor cases. To summarize, the two-part evaluation shows that GoogLeNet + SVM achieves the best compromise between diagnostic precision and edge efficiency, thereby making it the most realistic option for Edge AI healthcare applications in the real-world. Proposed future directions for this framework include integration of multimodal imaging and federated learning along with TensorRT model serving for deployment, to build a more complete and clinician-centered robust decision support system that AI will assist at the edge. This will include additional tumor types.