LiT-WSAG: high-precision 3D liver tumor segmentation via 2D training and 3D reconstruction.

Liver cancer remains a major cause of cancer mortality, and precise CT-based liver tumor segmentation is critical for early diagnosis and personalized treatment. In practice, fully 3D‑supervised training is limited by prohibitive annotation costs and memory and compute demands. Achieving 3D segmentation from purely 2D supervision is attractive yet challenging due to through-plane context loss, inter‑slice inconsistency, weak boundaries, and low contrast and noise. We propose LiT‑WSAG, a 2D‑trained framework that reconstructs voxel‑wise 3D masks. Built upon SAM‑Med2D, it incorporates a wavelet‑guided dual‑scale channel adapter (WDSC‑Adapter) to decouple low- and high-frequency content and enhance multi-scale representations, a lightweight L2.5D neighborhood fusion to recover through‑plane context at near‑2D cost, and adversarial training with a ramped GAN weight ('mean-loss' scheduling) to enforce shape and boundary consistency. Extensive experiments demonstrate significant gains over the baseline across multiple datasets. On the LiTS dataset, LiT‑WSAG achieves slice‑wise and volumetric DSC of 89.74% and 69.27% (+ 8.85 and + 33.64 percentage points over SAM-Med2D); on the WAW‑TACE dataset, the corresponding scores are 96.08% and 82.03%. Furthermore, on the Synapse multi-organ dataset, it attains a state-of-the-art average DSC of 88.23%. These results indicate that LiT‑WSAG delivers robust 3D liver tumor segmentation from 2D supervision, balancing boundary accuracy, noise robustness, and computational efficiency.