Medical Image Segmentation on LIDC-IDRI (5-fold cross-val)
69.1IoUOurs (Full Model w/ SAFT)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Ours (Full Model w/ SAFT)Tuning Strategy (PEFT/SAFT/Full Tuning)=SAFT, Evaluation Protocol=Fine-Tuning Strategy2026.04 | 69.1 | 81.5 | |
| + Semantic-Topological Graph Reasoning (STGR)Reasoning Module (STGR)=Enabled, Evaluation Protocol=Ablation Strategy2026.04 | 67.5 | 80.1 | |
| Full Model (No PEFT, Full Tuning)Tuning Strategy (PEFT/SAFT/Full Tuning)=Full Tuning, Evaluation Protocol=Fine-Tuning Strategy2026.04 | 66.8 | 79.5 | |
| LISAModel Category=Foundation & Referring Segmentation Models2026.04 | 62.8 | 76.2 | |
| + Text-to-Vision Intent Distillation (TVID)Distillation Strategy (TVID)=Enabled, Evaluation Protocol=Ablation Strategy2026.04 | 62.1 | 75.3 | |
| SegNetModel Category=Standard Medical Segmentation Models2026.04 | 61.2 | 74.3 | |
| SEEMModel Category=Foundation & Referring Segmentation Models2026.04 | 59.8 | 73 | |
| BaselinePrompting Strategy=LLaMA-3-V + MedSAM, Evaluation Protocol=Ablation Strategy2026.04 | 59.5 | 72.5 | |
| TransUNetModel Category=Standard Medical Segmentation Models2026.04 | 58.4 | 72.1 | |
| MedSAMPrompting Strategy=Text-Prompt, Model Category=Foundation & Referring Segmentation Models2026.04 | 56.5 | 70.3 | |
| U-NetModel Category=Standard Medical Segmentation Models2026.04 | 54.2 | 68.5 |