FPGA-based brain tumor MRI classification using multi-scale discrete hahn moments and a lightweight CNN

Brain tumor classification from magnetic resonance imaging (MRI) is an important task in computer-aided diagnosis, requiring high accuracy to provide reliable clinical decision support and assist radiologists in their diagnostic workflow as a complementary tool, without aiming to replace expert clinical judgment. In this paper, we propose a hardware-oriented method for four-class brain tumor MRI classification, based on a multi-scale representation using discrete Hahn moments and a lightweight convolutional neural network. Instead of processing raw images directly, the proposed approach transforms each image into a compact multi-channel tensor obtained by extracting Hahn coefficients at three complementary spatial levels: global, regional, and local. This representation captures discriminative information at different scales while reducing input redundancy and enabling parallel execution on FPGA. The Hahn polynomial matrices are pre-computed offline, while multi-scale moment extraction and lightweight network inference are performed online on a Xilinx Zynq UltraScale+ ZCU106 platform. Furthermore, the student network is trained offline via knowledge distillation to improve accuracy under complexity constraints. Experiments conducted on the Kaggle Brain Tumor MRI dataset, comprising glioma, meningioma, no tumor and pituitary classes, show that the proposed method achieves an accuracy of 89.21% while significantly reducing model complexity compared to raw image-based approaches.