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 composed of a Progressive Visual State Space (PVSS) encoder and a Hybrid Visual State Space (HVSS) decoder that work collaboratively to achieve efficient and accurate segmentation of hepatic echinococcosis (HE) lesions. The proposed Convolutional Vision State Space Block (CVSSB) module is designed to fuse local and global features, while the Haar Wavelet Transformation Block (HWTB) module compresses spatial information into the channel dimension to enable lossless downsampling. 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%.
Introduction
Hepatic Echinococcosis (HE) is a parasitic disease that poses a significant health challenge in pastoral regions. Accurate diagnosis and treatment planning rely heavily on the precise segmentation of lesions in CT scans. However, this task is complicated by the diverse morphology of lesions, which often exhibit irregular shapes, varying sizes, and indistinct boundaries that blend into surrounding healthy tissue.
Current deep learning approaches for medical image segmentation primarily utilize Convolutional Neural Networks (CNNs) or Transformers. CNNs are effective at extracting local features but struggle with global context due to their limited receptive fields. Conversely, Transformers excel at capturing long-range dependencies but suffer from quadratic computational complexity relative to image resolution, making them difficult to deploy in resource-constrained environments. To address these limitations, we introduce EAGLE, an efficient segmentation model leveraging State Space Models (SSMs) to achieve global awareness with linear complexity.
Methodology
The EAGLE architecture is a U-shaped network designed to balance global context modeling with local feature preservation. It consists of a Progressive Visual State Space (PVSS) encoder and a Hybrid Visual State Space (HVSS) decoder.
The PVSS encoder utilizes a series of Convolutional Vision State Space Blocks (CVSSB) and Haar Wavelet Transformation Blocks (HWTB). The CVSSB is the core computational unit, integrating a Mamba-based 2D Selective Scan (SS2D) block to capture global dependencies and a Depthwise-Aware FeedForward Network (DA-FFN) to enhance local feature extraction.
A key innovation in EAGLE is the use of the HWTB for downsampling. Traditional pooling layers often discard critical spatial information, which can be detrimental when segmenting lesions with subtle boundaries. The HWTB employs Haar wavelet transforms to compress spatial data into the channel dimension, facilitating "lossless" downsampling that preserves fine-grained edge information.
The HVSS decoder mirrors the encoder's structure but incorporates a hybrid approach. In deeper layers, it utilizes SSMs combined with Convolutional Block Attention Modules (CBAM) to suppress noise and focus on the most relevant lesion features.
Experimental Setup
As no large-scale public datasets for HE exist, we developed the HE-CT dataset, comprising CT scans from 260 patients (130 cystic and 130 alveolar echinococcosis cases). CT Hounsfield Unit (HU) values were windowed to a range of [-40, 110] to improve lesion contrast. The dataset was split into training (80%), validation (10%), and testing (10%) sets.
Results and Discussion
EAGLE was evaluated against several baseline models, including UNet, Swin-UNet, and the SSM-based MSVM-UNet. EAGLE achieved a state-of-the-art Dice Similarity Coefficient (DSC) of 89.76%, a Precision of 88.95%, and a Recall of 89.47%. This represents a 1.61% improvement over MSVM-UNet and a 4.54% improvement over the standard CNN-based UNet.
Qualitative analysis indicates that EAGLE is particularly effective at delineating invasive alveolar lesions and avoiding false positives in complex anatomical regions, such as near the spine. The integration of SSMs allows the model to maintain high performance while being more computationally efficient than Transformer-based alternatives.
Conclusion
EAGLE provides a robust and efficient solution for Hepatic Echinococcosis lesion segmentation. By combining the global modeling capabilities of State Space Models with lossless wavelet-based downsampling, the model achieves superior accuracy without the prohibitive computational costs of Transformers. Future work will focus on expanding the dataset and testing the model's generalizability across diverse clinical settings.
Summary updated