CPU-Efficient U-Net-Transformer with Quantized Soft Labels

Leong Chit Jeff Kwan, Albert Chung

Keywords: Medical image segmentation, Abdomen, Computed tomography, Semi-supervised learning, U-Net, Transformer

TL;DR: U-Net-Transformer hybrid model trained with quantized soft pseudo labels on the FLARE 2025 challenge Task 2 can run on CPU with great performance.

Abstract: This paper introduces a U-Net-Transformer hybrid model designed for the MICCAI FLARE 2025 Abdominal CT Organ Segmentation on Laptop Challenge. The model achieves both efficiency and performance through the integration of depthwise separable convolutions and transformer layers within the bottleneck. Our method incorporates quantized soft labels for improved boundary accuracy, aggressive multi-category data augmentation for enhanced robustness, class-weighted loss function and class-specific post-processing for precise segmentation of organs. On the online validation set, the pseudo-labeling model achieves mean Dice 0.9110 and NSD 0.9575, while the CPU-inference model achieves 0.8912 and 0.9518, respectively. Average inference time over 50 public validation cases is 19.3s per volume. These findings highlight the potential for practical deployment of the model in clinical environments with limited computational resources.

Submission Number: 13