Deployment of an Optimised Brain Tumour Segmentation Model on the ZCU104 FPGA Using Xilinx Vitis AI

Source abstract

Brain tumour segmentation from magnetic resonance imaging requires accurate delineation of sub-regions under clinical constraints on computational resources. Deploying deep segmentation networks on edge hardware demands model optimisation that preserves accuracy while reducing complexity. This paper presents the deployment of a pruned RAAGR2-Net brain tumour segmentation model on the Xilinx ZCU104 FPGA using the Vitis AI 3.0 toolchain. The model was optimised through sensitivity-guided channel pruning, which suppresses channels identified as redundant while maintaining segmentation stability. The pruned model was trained and evaluated on the BraTS 2019 dataset. At a pruning ratio of 0.1, the pruned model achieved Dice scores of 0.8578, 0.8131, and 0.8635 for tumour core, enhancing tumour, and whole tumour respectively, compared to 0.8169, 0.7896, and 0.8424 for the base model. The deployment further serves as a verification step, confirming that suppressed channels do not contribute to inference on hardware before structural removal is applied. Following INT8 post-training quantisation and compilation for the DPUCZDX8G accelerator, inference on the ZCU104 produced an average latency of 69.16 ms per sample and a throughput of 14.46 frames per second. These results demonstrate that sensitivity-guided channel pruning preserves segmentation accuracy and is compatible with FPGA deployment without requiring network restructuring.