Slice-aware MedSAM Adaptation with cross-slice consistency for 2.5D multi-organ CT segmentation
TL;DR
This work proposes a 2.5D multi-organ CT segmentation method based on slice-aware MedSAM adaptation and cross-slice consistency regularization, which achieves a Mean Dice of 0.8598 on the Synapse multi-organ CT dataset, outperforming U-Net and TransUNet.
Abstract
Abdominal CT multi-organ segmentation is important for computer-aided diagnosis and quantitative analysis, but it remains challenging due to large variations in organ size, shape, and boundary clarity. Existing 2D methods lack inter-slice contextual modeling, while 3D methods require high computational and memory costs. Most 2.5D methods stack adjacent slices and lack explicit constraints on inter-slice structural continuity. To address these issues, we propose a 2.5D multi-organ CT segmentation method based on slice-aware MedSAM adaptation and cross-slice consistency regularization. The proposed method consists of Slice-Attention input fusion, a LoRA-adapted MedSAM ViT-B encoder, a weak-gated multi-path DPAM decoder, and Tversky-Consistency joint optimization. Experimental results show that the proposed method achieves a Mean Dice of 0.8598 on the Synapse multi-organ CT dataset, outperforming U-Net and TransUNet.