评审老师的意见：第1.2.1节关于图像分割的讨论比较含糊。图像分割分为三类：语义分割、实例分割和全景...

TL;DR

The dissertation “基于小样本学习的医学图像分割方法研究 (Few‑shot Medical Image Segmentation)” develops two complementary networks that make it much easier to delineate organs and lesions when you only have a handful of annotated images:

1. MS‑ECANet – a general few‑shot segmenter that couples multi‑scale “memory” prototypes with an edge‑context module so the model still recognises objects of very different sizes and recovers crisp boundaries.
2. DPCP‑MNet – a medical‑specific few‑shot segmenter that fuses two learnable prompts (geometry‑aware edges + centre‑focused semantics) into a “collaborative prototype,” then injects it into a lightweight Mamba state‑space block to capture long‑range context in high‑resolution CT/MRI slices.

Both networks outperform state‑of‑the‑art baselines on PASCAL‑5i / COCO‑20i (natural images) and on abdominal & cardiac CT/MRI (medical images) while using only one annotated support slice per class.

---

1  Research background & motivation

Medical images have unique headaches: privacy limits data sharing, expert annotation is expensive, and organs look different from patient to patient. Conventional deep networks therefore over‑fit when training data are scarce. Few‑shot learning is attractive, but existing general‑purpose methods often lose edge detail and cannot adapt to blurred or low‑contrast boundaries that are common in CT/MRI.

---

2  Main contributions

#	Contribution	What’s new	Where it helps
1	MS‑ECANet (Multi‑Scale Edge Context Aggregation Network)	• Multi‑Scale Memory Coordination (MSMC) blends the current support prototype with historical prototypes stored in an external memory.<br>• Edge Context Aggregation Module (ECAM) mixes Sobel‑based edges with channel attention to keep borders sharp.	Natural scenes (PASCAL‑5i, COCO‑20i) and a preliminary abdominal MRI test.
2	DPCP‑MNet (Dual‑Prompt Collaborative‑Prototype Mamba Net)	• Edge‑aware geometric prompt + centre semantic prompt → merged into a collaborative prototype.<br>• A selective‑state Mamba block models full‑slice context in linear time, so high‑resolution volumes are tractable.<br>• Hierarchical decoder restores fine details.	Medical CT/MRI datasets (abdomen CT, abdomen MRI, cardiac MRI).
3	Systematic comparison with foundation models (SAM2, MedSAM2).	Shows that purpose‑built few‑shot nets can beat giant models when annotations are extremely limited.	Guides practitioners on when to choose which tool.

---

3  Experimental highlights

Dataset	Setting	Metric	Best baseline	Proposed
PASCAL‑5i	1‑shot	mIoU	HSNet 66.2 %	MS‑ECANet 72.9 %
COCO‑20i	1‑shot	mIoU	PI‑CLIP 56.8 %	MS‑ECANet 57.4 %
Abdomen MRI	1‑shot	Dice (mean of 4 organs)	Q‑Net 81.0 %	DPCP‑MNet 83 %
Cardiac MRI	1‑shot	Dice (LV, RV, Myo)	MedSAM2 (tuned) –	DPCP‑MNet improves by ≈2–4 pp

pp = percentage‑point gain

Ablation studies confirm:

• ECAM alone adds ~3 pp Dice on blurred organs; MSMC alone adds ~5 pp on multi‑scale targets; together they give the full 8 pp jump.
• Dual‑prompt + Mamba is superior to single‑prompt or Transformer‑style attention while cutting GPU memory.

---

4  Take‑away for practitioners

• If you only have a few labelled scans (≈1–5 per class) and care about both speed and sharp borders, the proposed pipelines are safer bets than fine‑tuning huge foundation models.
• Edge cues matter – explicit edge prompts or ECAM‑style modules consistently lift Dice/Hausdorff on fuzzy organs.
• State‑space (Mamba) layers are a promising middle ground: near‑Transformer quality without quadratic cost, great for full‑resolution 512×512 slices.

---

5  Limitations & future work

• Current memory module stores prototypes in Euclidean space; extending to probabilistic or hierarchical memories could reduce noise.
• Most tests are 2‑D; the authors plan a 3‑D variant for volumetric segmentation.
• Integrating cross‑modality priors (e.g., fusing CT & MRI) and uncertainty quantification are named as next steps.

---

Citation

基于小样本学习的医学图像分割方法研究. 硕士学位论文, 2025.