HMC-transducer: hierarchical mamba-CNN transducer for robust liver tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate liver tumor segmentation in CT scans is crucial but challenging due to tumor variability.
• Current deep learning models (CNNs, Transformers) have limitations in capturing both local and global features efficiently for 3D data.

Purpose of the Study:

• To introduce a novel hybrid deep learning architecture, the hierarchical mamba-CNN transducer (HMC-transducer), for enhanced liver tumor segmentation.
• To overcome the limitations of existing models by integrating CNNs with Mamba's efficient long-range dependency modeling.

Main Methods:

• Developed a novel HMC-transducer architecture featuring direction-aware 3D Mamba (DA3D-Mamba) blocks for volumetric data processing.
• Incorporated a Mamba-CNN Transducer block with gated fusion for adaptive combination of local and global features.
• Evaluated the model on public benchmarks: LiTS17, MSD-liver, and KiTS21.

Main Results:

• The HMC-transducer achieved state-of-the-art segmentation accuracy on multiple liver tumor datasets.
• Demonstrated superior generalization capabilities compared to existing CNN- and transformer-based methods.
• Showcased improved computational efficiency.

Conclusions:

• The HMC-transducer offers a significant advancement in automated liver tumor segmentation from CT scans.
• This hybrid approach provides a practical and efficient solution for clinical applications.
• The model's effectiveness highlights the potential of integrating Mamba with CNNs for complex medical image analysis.