Event-Driven Neuromorphic AI for Brain Tumor Classification: A CNN–SNN Framework with Hardware Validation

Abstract

Classifying brain tumors using magnetic resonance imaging (MRI) is a crucial medical image processing problem; however, deployment on platforms with limited resources is limited by the computational expense of traditional deep learning models. Convolutional neural network (CNN) feature extraction, spike encoding, spiking neural network (SNN) inference, fixed-point quantization, and hardware validation are all integrated in this paper’s Event-Driven Fixed-Point Neuromorphic Inference Engine (ED-FPNIE). Compact features are extracted by a lightweight CNN and transformed into temporal spike sequences for SNN-based classification. While inference employs event-driven conditional accumulation in place of traditional multiplier operations, quantized synaptic weights are mapped to ROM-based memory. A four-class brain MRI dataset consisting of the Glioma, Meningioma, No Tumor, and Pituitary categories was used to assess the framework. With class-wise accuracies of 69.00%, 46.25%, 96.00%, and 82.25%, respectively, RTL validation on 1600 test images yielded an accuracy of 73.38%. Using the Nangate45 library, gate-level synthesis produced a cell area of 7866.68 μ"m" ^2 and an estimated power of 1.38 mW. The results demonstrate the feasibility of a hardware-oriented event-driven neuromorphic inference framework for resource-constrained medical AI applications.