logo
EAGLE: An Efficient Global Attention Lesion Segmentation
Model for Hepatic Echinococcosis

- PRCV 2025 -

Jiayan Chen1, Kai Li2, Yulu Zhao1, Jianqiang Huang1,*, Zhan Wang1,**

1Qinghai University      2Tsinghua University


  arXiv     Code  

Abstract

Hepatic echinococcosis (HE) is a widespread parasitic disease in underdeveloped pastoral areas with limited medical resources. While CNN-based and Transformer-based models have been widely applied to medical image segmentation, CNNs lack global context modeling due to local receptive fields, and Transformers, though capable of capturing long-range dependencies, are computationally expensive. Recently, state space models (SSMs), such as Mamba, have gained attention for their ability to model long sequences with linear complexity. In this paper, we propose EAGLE, a U-shaped network that integrates CNNs, Transformers, and SSMs. We introduce the Convolutional Vision State Space Block (CVSSB) to fuse local and global features and employ the Haar Wavelet Transformation Block (HWTB) for lossless downsampling. EAGLE leverages the synergy between the proposed Progressive Visual State Space (PVSS) Encoder and Hybrid Visual State Space (HVSS) Decoder to achieve efficient and accurate segmentation of HE lesions. Due to the lack of publicly available HE datasets, we collected CT slices from 260 patients at a local hospital. Experimental results show that EAGLE achieves state-of-the-art performance with a Dice Similarity Coefficient (DSC) of 89.76%, surpassing MSVM-UNet by 1.61%.

Experiments and Results

Overview of the HE dataset

We collected abdominal CT scans from 260 patients who were clinically diagnosed with hepatic echinococcosis (HE), including 130 cases of cystic echinococcosis (CE) and 130 cases of alveolar echinococcosis (AE) . Here are some representative samples.

Results on HE Lesion Segmentation

Comparison of different methods on the HE dataset. The optimal values are highlighted in bold, while the second-best values are underlined. Specifically, we compared to the CNN-based UNet, EAGLE achieves a 4.54% improvement in DSC . When compared with the Transformer-based SwinUNet, EAGLE yields a 2.21% gain in DSC. Additionally, EAGLE outperforms the SSMs-based MSVM-UNet by 1.61% in DSC. Furthermore, EAGLE also achieves the highest scores in both Precision and Recall metrics.

Ablation study on different CVSSB stacking configurations. Configurations marked with CUDA Out of Memory could not be evaluated due to GPU memory limitations. We focused on enhancing stage 3 while keeping the overall parameter count within a manageable range. The final configuration, L = [2, 2, 4, 2], proved to be the optimal setting.

Ablation study of different modules. The optimal values are highlighted in bold. We use a Vanilla UNet as the baseline. Specifically, when all three modules are applied together, the DSC, Precision, and Recall are improved by 3.1%, 1.68%, and 2.24%, respectively.

BibTeX

The following demos illustrate a comparison between EAGLE and other existing segmentation models on our dataset. In the visualization, true positives (TP) are highlighted in green , false positives (FP) in blue, and false negatives (FN) in hotpink.

Cystic Echinococcosis (CE) Lesion Segmentation Sample
Alveolar Echinococcosis (AE) Lesion Segmentation Sample

BibTeX

                        
@article{chen2025eagle,
        title={EAGLE: An Efficient Global Attention Lesion Segmentation Model for Hepatic Echinococcosis},
        author={Chen, Jiayan and Li, Kai and Zhao, Yulu and Huang, Jianqiang and Wang, Zhan},
        journal={arXiv preprint arXiv:2506.20333},
        year={2025}
}