3D Medical Image Segmentation on ACDC (Structure-specific Dice and HD95)
89.8Dice (RV)UNet++
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| UNet++2026.01 | 89.8 | 87.1 | 95.2 | 90.7 | 5.01 | 1.14 | 2.5 | 2.88 | |
| AutoSAM (CNN)2026.01 | 88.8 | 86 | 94.2 | 89.7 | 8.68 | 5.48 | 5.63 | 6.6 | |
| Attention UNet2026.01 | 87.8 | 85.8 | 94.7 | 89.4 | 9.05 | 1.44 | 1.23 | 3.91 | |
| MambaSAM (IFA Dec)Inference Speed (FPS/Volume)=2.63, Max VRAM (GB)=10.99, Total Params (M)=113.64, Trainable Params (M)=23.97 (≈ 21%)2026.01 | 87.1 | 87.4 | 93.4 | 89.3 | 12.48 | 8.36 | 7.22 | 9.35 | |
| TP_MFGCInference Speed (FPS/Volume)=4.77, Max VRAM (GB)=12.62, Total Params (M)=112.61, Trainable Params (M)=22.93 (≈ 20%)2026.01 | 86.8 | 68 | 89.7 | 88 | 53.08 | 19.87 | 24.21 | 32.39 | |
| MambaSAM-BaseInference Speed (FPS/Volume)=2.78, Max VRAM (GB)=11.57, Total Params (M)=113.55, Trainable Params (M)=23.88 (≈ 21%)2026.01 | 83.6 | 91 | 97.1 | 90.6 | 11.81 | 5.76 | 5.01 | 7.53 | |
| MambaUNet2026.01 | 83.5 | 78.9 | 91.8 | 84.7 | 7.73 | 2.29 | 3.71 | 4.58 | |
| TP-Mamba (LoRA)Inference Speed (FPS/Volume)=0.72, Max VRAM (GB)=1.9, Total Params (M)=≈ 90M + Adapters + LoRA, Trainable Params (M)=Very Few (Adapters + LoRA only)2026.01 | 75.8 | 76.9 | 86 | 79.6 | 4.83 | 13.77 | 7.02 | 8.54 | |
| TP-Mamba (simple)Inference Speed (FPS/Volume)=0.76, Max VRAM (GB)=2.41, Total Params (M)=199.21, Trainable Params (M)=109.53 (≈ 55%)2026.01 | 62.8 | 66.4 | 74.6 | 67.9 | 14.79 | 5.94 | — | — | |
| SwinUNet2026.01 | 57.2 | 60.7 | 78.2 | 65.4 | 34.46 | 7.38 | 9.71 | 17.18 |