Clinical Report: MemSAM-2.5D for 3D Segmentation of Liver Tumors
Overview
MemSAM-2.5D is a novel segmentation framework designed for liver tumor delineation in 3D CT scans. It addresses challenges related to volumetric discontinuities and boundary uncertainties.
Background
Accurate segmentation of liver tumors is critical for the management of hepatocellular carcinoma (HCC). Traditional methods face challenges due to variability in tumor size, discontinuities in volumetric imaging, and ambiguous tumor boundaries.
Data Highlights
MemSAM-2.5D was evaluated on multiple datasets, including MSD08, HCC-TACE-Seg, and WAW-TACE.
Key Findings
MemSAM-2.5D integrates a Hybrid Mamba-Adapter (HMA) for enhanced intra-slice feature representation.
The Z-axis State Flow (ZSF) module models long-range dependencies between slices.
The Confidence-Gated Prototype Memory (CGPM) module refines boundaries by evaluating predictive uncertainty.
Evaluations show improvements in overlap-based, boundary-sensitive, and continuity-related metrics.
MemSAM-2.5D provides a solution for HCC segmentation.
Clinical Implications
The MemSAM-2.5D framework offers a promising approach to improve the accuracy of liver tumor segmentation in clinical settings. Its ability to address volumetric and boundary challenges may enhance diagnostic precision and treatment planning for HCC.
Conclusion
MemSAM-2.5D represents an advancement in the segmentation of liver tumors from 3D CT volumes.
