Hierarchical Mamba-CNN Transducer for Enhanced Liver Tumor Segmentation in CT Imaging

Overview

The hierarchical mamba-CNN transducer (HMC-transducer) introduces a novel hybrid architecture combining CNNs with the Mamba state space model to improve liver tumor segmentation from CT scans. It achieves state-of-the-art accuracy, superior generalization, and computational efficiency on multiple public datasets.

Background

Accurate liver tumor segmentation from CT imaging is essential for diagnosis and treatment planning but is challenged by tumor heterogeneity and unclear boundaries. Traditional CNNs capture local features well but lack long-range spatial modeling, while transformers provide global context at high computational cost. The HMC-transducer addresses these limitations by integrating CNNs with a direction-aware 3D Mamba block, enabling efficient volumetric data processing and adaptive fusion of local and global features.

Data Highlights

Key Findings

• The HMC-transducer combines CNNs with the linear-complexity Mamba state space model for effective local and global feature integration.
• The direction-aware 3D Mamba block preserves spatial topology along all three axes, enhancing volumetric data processing.
• A gated fusion mechanism adaptively weighs local and global features at each network hierarchy level.
• Extensive evaluation on LiTS17, MSD-liver, and KiTS21 datasets demonstrates superior segmentation accuracy over existing CNN and transformer models.
• The model achieves better generalization and computational efficiency, addressing the quadratic cost limitations of transformers in 3D imaging.

Clinical Implications

The HMC-transducer offers a practical and generalizable solution for liver tumor segmentation in clinical CT imaging, potentially improving diagnostic accuracy and treatment planning. Its computational efficiency facilitates deployment in real-world clinical workflows, enabling faster and more reliable tumor delineation.

Conclusion

The hierarchical mamba-CNN transducer represents a significant advancement in liver tumor segmentation by effectively integrating local and global features with computational efficiency. This approach sets a new benchmark for accuracy and generalization in volumetric medical image analysis.

References

1. Gu & Dao 2023 -- Mamba: linear-time sequence modeling with selective state spaces
2. Ronneberger et al. 2015 -- U-net: convolutional networks for biomedical image segmentation
3. Chen et al. 2021 -- Transunet: transformers make strong encoders for medical image segmentation
4. Hatamizadeh et al. 2022 -- Unetr: transformers for 3d medical image segmentation
5. LiTS17 Dataset -- Liver Tumor Segmentation Challenge
6. MSD-liver Dataset -- Medical Segmentation Decathlon
7. KiTS21 Dataset -- Kidney Tumor Segmentation Challenge