Zhang Qinghua, Hou Yulei, He Changchun, Zhai Zhengyu, Deng Yunjiao
Department of Neurosurgery, Huazhong University of Science and Technology Union Shenzhen Hospital, Shenzhen 518052, China.
The 6th Affiliated Hospital of Shenzhen University Health Science Center, Shenzhen 518060, China.
Curr Med Imaging. 2025;21:e15734056370555. doi: 10.2174/0115734056370555250426140155.
Semantic segmentation algorithms are essential for identifying and segmenting human organs and lesions in medical images. However, as U-Net variants enhance segmentation accuracy, they often increase in parameter count, demanding more sophisticated and costly hardware for training.
This study aims to introduce a lightweight U-Net that optimizes the trade-off between network parameters and segmentation accuracy, while fully leveraging the encoder's feature extraction capabilities.
We propose a lightweight full-encoder U-shaped network, termed LFE-UNet, which employs full-encoder skip connections, encompassing all encoder layers. This model is designed with a reduced number of basic channels-specifically, 8 instead of the typical 64 or 32-to achieve a more efficient architecture.
The LFE-UNet, when integrated with ResNet34, achieved a Dice score of 0.97385 on the ISBI LiTS 2017 liver dataset. For the BraTS 2018 brain tumor dataset, it obtained 0.87510, 0.93759, 0.87301, and 0.81469 on average, WT, TC, and ET, respectively. The paper also discusses the impact of varying basic channel numbers n and encoder layer counts N on the network's parameter efficiency, as well as the model's robustness to different levels of Gaussian noise in images and salt and pepper noise in labels. Additionally, the influence of different loss functions is explored.
The LFE-UNet proves that high segmentation accuracy can be attained with a markedly lower parameters, fully utilizing the full-scale encoder's feature extraction. It also highlights the significance of loss function selection and the effects of noise on segmentation accuracy.
语义分割算法对于识别和分割医学图像中的人体器官及病变至关重要。然而,随着U-Net变体提高分割精度,其参数数量往往会增加,需要更复杂且昂贵的硬件进行训练。
本研究旨在引入一种轻量级U-Net,在优化网络参数与分割精度之间权衡的同时,充分利用编码器的特征提取能力。
我们提出了一种轻量级全编码器U形网络,称为LFE-UNet,它采用全编码器跳跃连接,涵盖所有编码器层。该模型设计为减少基本通道数量——具体为8个,而非典型的64或32个——以实现更高效的架构。
LFE-UNet与ResNet34集成时,在ISBI LiTS 2017肝脏数据集上的Dice分数达到0.97385。对于BraTS 2018脑肿瘤数据集,在WT、TC和ET上的平均分数分别为0.87510、0.93759、0.87301和0.81469。本文还讨论了不同基本通道数n和编码器层数N对网络参数效率的影响,以及模型对图像中不同程度高斯噪声和标签中椒盐噪声的鲁棒性。此外,还探讨了不同损失函数的影响。
LFE-UNet证明了在显著减少参数的情况下仍可实现高分割精度,充分利用了全尺度编码器的特征提取能力。它还突出了损失函数选择的重要性以及噪声对分割精度的影响。