A Quality of Service Analysis of FPGA-Accelerated Conv2D Architectures for Brain Tumor Multi-Classification

In medical imaging, accurate brain tumor classification in medical imaging requires real-time processing and efficient computation, making hardware acceleration essential. Field Programmable Gate Arrays (FPGAs) offer parallelism and reconfigurability, making them well-suited for such tasks. In this study, we propose a hardware-accelerated Convolutional Neural Network (CNN) for brain cancer classification, implemented on the PYNQ-Z2 FPGA. Our approach optimizes the first Conv2D layer using different numerical representations: 8-bit fixed-point (INT8), 16-bit fixed-point (FP16), and 32-bit fixed-point (FP32), while the remaining layers run on an ARM Cortex-A9 processor. Experimental results demonstrate that FPGA acceleration significantly outperforms the CPU (Central Processing Unit) based approach. The obtained results emphasize the critical importance of selecting the appropriate numerical representation for hardware acceleration in medical imaging. On the PYNQ-Z2 FPGA, the INT8 achieves a 16.8% reduction in latency and 22.2% power savings compared to FP32, making it ideal for real-time and energy-constrained applications. FP16 offers a strong balance, delivering only a 0.1% drop in accuracy compared to FP32 (94.1% vs. 94.2%) while improving latency by 5% and reducing power consumption by 11.1%. Compared to prior works, the proposed FPGA-based CNN model achieves the highest classification accuracy (94.2%) with a throughput of up to 1.562 FPS, outperforming GPU-based and traditional CPU methods in both accuracy and hardware efficiency. These findings demonstrate the effectiveness of FPGA-based AI acceleration for real-time, power-efficient, and high-performance brain tumor classification, showcasing its practical potential in next-generation medical imaging systems.