MemSAM-2.5D: overcoming volumetric discontinuity and boundary ambiguity for 3D liver tumor segmentation.

INTRODUCTION

Accurate segmentation of liver tumors from 3D computed tomography (CT) volumes is essential for the clinical management of hepatocellular carcinoma (HCC), but remains challenging because of extreme lesion-scale variation, volumetric discontinuity across slices, and ambiguous tumor boundaries.

METHODS

We propose MemSAM-2.5D, a unified 2.5D segmentation framework built upon the MedSAM foundation model. The framework integrates a Hybrid Mamba-Adapter (HMA) for intra-slice multi-scale representation, a Z-axis State Flow (ZSF) module for continuous inter-slice dependency modeling, and a Confidence-Gated Prototype Memory (CGPM) module for uncertainty-aware boundary refinement.

RESULTS

Extensive evaluations on MSD08, HCC-TACE-Seg, and WAW-TACE demonstrate that MemSAM-2.5D consistently outperforms representative CNN-based, Transformer-based, Mamba-based, and MedSAM-based baselines. The improvements are reflected not only in overlap-based metrics, but also in boundary-sensitive, lesion-level, and continuity-related measures.

DISCUSSION: These results suggest that coordinated modeling of multi-scale lesion variability, z-axis continuity, and boundary ambiguity provides an effective and transferable solution for clinically relevant HCC segmentation in CT volumes.