CDL-UNet:Curriculum-Driven Lightweight 3D U-Net for Abdominal Organ Segmentation

Zhehao Wang, Xian Lin, Junjie Shi, Caozhi Shang, Li Yu, Zengqiang Yan

Keywords: Medical image segmentation，Curriculum learning，Computational efficiency

Abstract: Deep learning has significantly advanced medical image segmentation, particularly for large-scale datasets. However, deploying these models on resource-limited devices like laptops presents two new difficulties: (1) Data-level challenges due to the large volume of data, limited availability of high-quality labels, and varying quality of pseudo-labels; and (2) Computational constraints, as CPU-only inference requires achieving both speed and performance under limited resources. To tackle these issues, we propose a novel curriculum-driven lightweight 3D U-Net (CDL-UNet) approach, which integrates a curriculum learning strategy, a label-based difficulty discriminator, and an adaptive sliding window inference method. Our curriculum learning strategy progressively trains the model with increasingly complex samples to enhance learning efficiency and accuracy. The label-based difficulty discriminator refines pseudo-labels and categorizes samples by difficulty, optimizing the training process. Finally, the adaptive sliding window inference ensures fast and accurate segmentation even with CPU-only hardware. Our method achieved an average score of 88.28% and 93.80% for organ DSC and NSD on the online validation set, with an average inference time of 38 seconds, demonstrating its effectiveness for high-quality segmentation on resource-constrained devices.

Submission Number: 11