MemSAM-2.5D: overcoming volumetric discontinuity and boundary ambiguity for 3D liver tumor segmentation
By
Yinyin Hou
Ningning Chen
Tingting Huo
Weijia Wang
July 8, 2026
Objective: To propose a unified 2.5D segmentation framework, MemSAM-2.5D, for accurate liver tumor segmentation in 3D CT volumes.
Approach: Hybrid Mamba-Adapter (HMA): Integrates intra-slice multi-scale representation to capture contextual correlations.Z-axis State Flow (ZSF) module: Models continuous inter-slice dependencies using a one-dimensional state-space model.Confidence-Gated Prototype Memory (CGPM): Refines boundaries by evaluating predictive uncertainty and excluding ambiguous regions.Key Findings: MemSAM-2.5D outperforms CNN-based, Transformer-based, Mamba-based, and MedSAM-based baselines on the MSD08, HCC-TACE-Seg, and WAW-TACE datasets. Improvements are noted in overlap-based metrics, boundary-sensitive measures, and continuity-related assessments. Interpretation:
Limitations: The study does not address potential overfitting risks associated with the proposed model. Further validation on diverse datasets, including those with varying tumor characteristics, may be required to confirm generalizability. Conclusion: MemSAM-2.5D provides an effective solution for clinically relevant liver tumor segmentation.
