Enhanced Mamba Filtering Networks for Precise HCC Lesion Segmentation in CT

Overview

Prompt-Mamba-AF, a novel deep learning framework, significantly improves hepatocellular carcinoma (HCC) lesion segmentation in abdominal CT scans by integrating anatomy-aware prompts and Mamba-based state-space modeling. Validated on multiple benchmark datasets, it outperforms existing CNN and Transformer models in Dice similarity, boundary accuracy, and small nodule sensitivity while maintaining computational efficiency.

Background

Accurate segmentation of hepatocellular carcinoma lesions in abdominal CT imaging is critical for early diagnosis and effective surgical planning. Challenges include morphological heterogeneity of tumors, low contrast especially in small lesions, and variability across imaging scanners. Traditional convolutional neural networks and Transformer-based models have limitations in capturing long-range volumetric dependencies and maintaining topological consistency. Advances in state-space modeling and anatomy-guided feature extraction offer potential to overcome these obstacles.

Data Highlights

Key Findings

• Prompt-Mamba-AF integrates anatomy-aware prompts to focus feature extraction within liver regions, enhancing lesion delineation.
• Mamba-based state-space modeling captures long-range volumetric dependencies with linear computational complexity.
• Structure-aware filtering enforces topological consistency along lesion boundaries, improving segmentation accuracy.
• The model achieves superior Dice similarity and boundary accuracy across LiTS, 3DIRCADb, and CHAOS datasets compared to current CNN and Transformer architectures.
• Notably improved sensitivity for detecting small HCC nodules, addressing a critical clinical challenge.
• Compact model size (27.6 million parameters) facilitates efficient deployment in multi-center clinical workflows.

Clinical Implications

The enhanced segmentation accuracy and improved sensitivity for small lesions offered by Prompt-Mamba-AF can support earlier and more precise diagnosis of hepatocellular carcinoma, potentially improving patient outcomes. Its computational efficiency and generalizability across diverse imaging domains make it suitable for integration into routine clinical practice and multi-center studies. This approach may streamline surgical planning and treatment monitoring by providing reliable lesion delineations.

Conclusion

Prompt-Mamba-AF represents a significant advancement in automated HCC lesion segmentation, combining anatomy-aware guidance and efficient long-range modeling to outperform existing methods. Its robust performance and compact design position it as a promising tool for clinical adoption in abdominal CT imaging workflows.

References

1. Singal et al. 2023 -- Global trends in hepatocellular carcinoma epidemiology: implications for screening, prevention and therapy
2. Bilic et al. 2023 -- The liver tumor segmentation benchmark (LiTS)
3. Gu & Dao 2024 -- Mamba: Linear-time sequence modeling with selective state spaces
4. Yao et al. 2024 -- From CNN to Transformer: a review of medical image segmentation models
5. Zhou et al. 2019 -- Unet++: Redesigning skip connections to exploit multiscale features in image segmentation