Clinically oriented automatic 2D liver tumor segmentation: LCMambaNet with a state-space model and liver...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Liver cancer presents significant global health challenges with rising incidence and mortality.
• Accurate tumor segmentation is difficult due to complex imaging characteristics and limitations of current deep learning models.
• Existing 3D deep learning architectures require substantial computational resources, hindering clinical application.

Purpose of the Study:

• To introduce the Liver Cancer Mamba Network (LCMambaNet), an efficient 2D segmentation framework for liver tumors.
• To address limitations of existing models in balancing computational efficiency and segmentation accuracy.
• To improve the detection and characterization of liver lesions, particularly small ones.

Main Methods:

• Developed LCMambaNet, a 2D segmentation framework utilizing selective state-space models.
• Implemented a scan-patch mechanism for extracting salient texture and density features.
• Incorporated the Liver Cancer Attention Module (LCAM) to differentiate tumor characteristics from normal tissue.
• Evaluated performance on LITS (CT) and CirrMR160+ (MRI) datasets.

Main Results:

• LCMambaNet achieved high segmentation accuracy, with Dice scores of 92.94% on LITS and 92.08% on CirrMR160+.
• Demonstrated statistically significant superior performance on small lesions (< 2 cm) compared to baseline models.
• Ablation studies confirmed the effectiveness of individual model components.

Conclusions:

• LCMambaNet provides an efficient and clinically viable solution for 2D liver tumor segmentation.
• The model's design overcomes computational and accuracy limitations of current deep learning approaches.
• Its strong performance on small lesions supports early diagnosis and precise treatment planning in liver cancer management.