Research map / Detection and segmentation / Open-set and few-shot detection

UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation

NeurIPS 2026 · Detection and segmentation › Open-set and few-shot detection · approach: Interactive and referring segmentation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UniPro, a unified slice-based medical segmentation model that handles semantic, in-context, interactive, and reference-conditioned propagation segmentation across 2D images and 3D volumes, with bidirectional and 3D verification losses for propagation.

Problem

Medical image segmentation is fragmented across segmentation paradigms (semantic, in-context, interactive) and data dimensionality (native 2D vs 3D volumes), and 2D edits do not propagate coherently through volumes.

Approach

Treat volumetric propagation as reference-conditioned in-context segmentation where references are previously segmented neighboring slices, so one shared slice-based backbone with dense-plus-sparse query conditioning and MoE decoder covers all modes; add backward seed-reconstruction and a 3D verification loss to curb drift.

Key results

• Single-click propagation: UniPro (prop) reaches 54.89% 3D Dice on unseen BraTS vs 40.74% for MedSAM2, and 84.40% on BTCV.
• Semantic setting: UniPro (prop) reaches 89.49% on ACDC, surpassing fully supervised nnU-Net (3D) at 88.59%.
• Interactive correction on BraTS: propagation cuts edited slices from 62.4 to 12.1 versus slice-wise editing.

Surprising findings

• Stage-2 propagation training degrades native 2D accuracy (ISIC drops from 87.54% to 82.40% without 2D interleaving), a trade-off only partly recovered by interleaving.
• UniPro (prop) outperforms the dedicated 3D interactive model nnInteractive on most 3D datasets in the single-click setting.

Limitations

• Stage-2 propagation adaptation reduces native 2D accuracy relative to Stage-1 only (ISIC still 1.43 points below).
• Trails fully supervised nnU-Net (3D) on most datasets in the semantic setting; no explicit spacing inputs.
• Single-click initialization may miss disconnected or fragmented components; longer inference time from sequential propagation.

Experimental setup

Datasets: Abd-1K CHAOS M&Ms LiTS KiTS Breast BKAI Chest ISIC BUS BTCV ACDC

Benchmarks: Abd-1K CHAOS M&Ms LiTS KiTS BTCV ACDC BraTS Breast BKAI Chest ISIC

Metrics: Dice

Baselines: nnU-Net (2D) UniSeg Hermes UniverSeg SegMIC Iris nnU-Net (3D) nnInteractive SAM2 PAM MedSAM2 iSegFormer

Builds on: Mask2Former SAM nnU-Net

Closest papers by meaning

• VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.89)
• Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation (NeurIPS 2024 · similarity 0.88)
• Clustering Propagation for Universal Medical Image Segmentation (CVPR 2024 · similarity 0.88)
• SegVol: Universal and Interactive Volumetric Medical Image Segmentation (NeurIPS 2024 · similarity 0.87)
• MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling (CVPR 2024 · similarity 0.87)
• Show and Segment: Universal Medical Image Segmentation via In-Context Learning (CVPR 2025 · similarity 0.86)