Research map / Label-efficient and robust learning / Self-supervised and contrastive
SAMora: Enhancing SAM through Hierarchical Self-Supervised Pre-Training for Medical Images
Shuhang Chen, Hangjie Yuan, Pengwei Liu, Hanxue Gu, Tao Feng, Dong Ni
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hierarchical image/patch/pixel self-supervised LoRA experts to SAMed | SAMed · Synapse (10% train) | Mean Dice (%): 75.57 -> 79.41 | Improved |
| add SAMora to H-SAM | H-SAM · Synapse (fully supervised) | Mean Dice (%): 86.49 -> 88.59 | Improved |
| add SAMora pre-trained LoRA experts, cutting fine-tuning epochs | SAMed · Synapse (10% train) | Fine-tuning epochs: 200 -> 20 | Improved |
| use HL-Attn instead of LAC fusion of LoRA experts (ablation) | H-SAMora · Synapse (10% train) | Mean Dice (%): 82.41 -> 84.34 | Improved |
| use HL-Attn instead of MOLE fusion (ablation) | H-SAMora · Synapse (10% train) | Mean Dice (%): 83.91 -> 84.34 | Improved |
| fuse image-level LoRA last (order 2-1-1) rather than patch-level last (1-2-1) (ablation) | HL-Attn · Synapse (10% train) | Mean Dice (%): 83.86 -> 84.34 | Improved |
| add continual pre-training to teacher-student distillation (patch-level LoRA) (ablation) | SAMora · Synapse (10% train) | Mean Dice (%): 77.19 -> 78.81 | Improved |
| add teacher-student distillation without CPT (image-level LoRA) (ablation) | SAMora · Synapse (10% train) | Mean Dice (%): 77.20 -> 77.31 | No clear effect |
| add teacher-student distillation with CPT instead of training LoRA from scratch (patch-level) (ablation) | SAMora · Synapse (10% train) | Mean Dice (%): 76.54 -> 78.81 | Improved |
| increase LoRA rank from 4 to 16 (ablation) | SAMora and baselines · Synapse (10% train) | Mean Dice (%): 79.41 -> 76.88 | Hurt |
| replace HL-Attn with hierarchical cross-attention transformer (HCAT) (ablation) | H-SAMora · Synapse (10% train) | Mean Dice (%): 84.34 -> 84.40 | No clear effect |
| shorten continual pre-training (no CPT, minimal pre-training) (ablation) | H-SAMora · Synapse (10% train) | Mean Dice (%): 84.34 -> 80.72 | Hurt |
Closest papers by meaning
- Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.89)
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- BLO-SAM: Bi-level Optimization Based Finetuning of the Segment Anything Model for Overfitting-Preventing Semantic Segmentation (ICML 2024 · similarity 0.86)
- Revisiting MAE Pre-training for 3D Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.86)
- Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
