Spatio-Temporal Decoupling Network for Multiphasic Liver Lesion Analysis

Overview

This study introduces STD-Net, a novel spatio-temporal decoupling network that separately models spatial features and temporal dynamics in multiphasic liver imaging. The approach outperforms state-of-the-art methods in segmentation accuracy and lesion characterization, particularly improving detection of small or low-contrast hepatocellular carcinoma lesions.

Background

Hepatocellular carcinoma (HCC) is a leading cause of cancer mortality, where early and accurate diagnosis via imaging is critical. Multiphasic CT and MRI scans provide dynamic contrast-enhancement information across arterial, portal venous, and delayed phases, essential for identifying characteristic HCC patterns such as arterial phase hyperenhancement and venous washout. Existing deep learning methods often treat these phases as simple multi-channel inputs, failing to capture the temporal evolution and suffering from entangled spatial and temporal features, which limits diagnostic performance.

Data Highlights

Key Findings

• STD-Net explicitly separates spatial feature extraction from temporal dynamics modeling, mirroring clinical diagnostic reasoning.
• A shared-weight 3D encoder robustly captures anatomical features from each phase independently.
• A transformer-based temporal module effectively models sequential contrast patterns such as arterial hyperenhancement and venous washout.
• STD-Net achieves higher Dice scores, lower HD95 distances, and better classification accuracy than state-of-the-art baselines across multiple datasets.
• The approach provides more stable and generalizable performance, especially for small or low-contrast liver lesions.
• Spatio-temporal decoupling reduces feature entanglement caused by motion artifacts and phase misalignment.

Clinical Implications

The proposed STD-Net framework enhances the accuracy and reliability of liver lesion segmentation and characterization in multiphasic imaging, facilitating improved diagnostic confidence. By explicitly modeling temporal contrast dynamics, it better captures hallmark HCC features, potentially aiding earlier detection and treatment planning. Its robustness to motion artifacts and generalizability suggest applicability in routine clinical workflows and other dynamic imaging contexts.

Conclusion

STD-Net demonstrates that spatio-temporal decoupling is a powerful paradigm for dynamic liver lesion analysis, yielding superior segmentation and characterization performance. This approach aligns closely with clinical reasoning and offers a promising blueprint for future dynamic medical imaging applications.

References

1. HCC Imaging and Diagnosis Context
2. Deep Learning for Liver Lesion Segmentation and Characterization
3. Vision Transformers in Medical Imaging