Towards Eight-Bit Quantization for 3D U-Net Medical Image Segmentation via ROI-Based Calibration and Background-Aware Shift

3D U-Net medical image segmentation is essential for precise diagnosis and treatment planning but comes with massive bandwidth requirements and considerable computational costs. Eight-bit fixed-point quantization is crucial for efficient 3D U-Net inference on modern deep learning accelerators. However, quantizing a 3D U-Net model is non-trivial and frequently leads to unfavorable accuracy degradation. We empirically observed that large range imbalances between tumor/non-tumor voxels and foreground/background areas are the main contributing factors. Based on these findings, we propose an eight-bit quantization method with a simple yet effective ROI-based calibration and a background shift that captures outliers in tumor areas to avoid accuracy degradation. Moreover, our method incorporates instance normalization folding, which eliminates computation-intensive operations during inference. Experimental results on the KITS19 dataset show that the proposed method with integer-only arithmetic achieves a negligible accuracy drop, i.e., 0.4%p in the dice score, compared to the FP32 model.