Error-Tolerant Medical Image Segmentation Using Resource-Efficient Approximate Booth Multipliers

CNN-based medical image segmentation models, such as U-Net, achieve high accuracy but demand significant computation and power, limiting real-time deployment in resource-constrained clinical settings. This letter presents AxC-UNet, a hardware-aware segmentation accelerator employing two Radix-4 approximate Booth encoders (AxBEM1 and AxBEM2) and three low-power approximate 4:2 compressors to optimize multiply-and-accumulate (MAC) operation. FPGA synthesis on a Xilinx ZCU104 shows up to 16.1% LUT reduction and 10% dynamic power savings compared to exact Booth multipliers, while maintaining segmentation accuracy within <6% Dice variation—comparable to natural inter-observer disagreement in BraTS2020—thereby ensuring clinical reliability. This co-design approach demonstrates a practical trade-off between clinical performance and hardware efficiency, enabling energy-efficient real-time medical image segmentation on edge platforms.