---
type: slide
---

#### Towards Segment Anything Model (SAM) for Medical Image Segmentation: A Survey

---

#### Problems:
- exisiting model are often tailored for specific modalities and targets
- limits their capacity for further generalization to tasks

---

#### Introduction:
- evaluate the performancee of SAM with different modes in different task
- investigate methods to better adapt SAM to medical image segementation task

---

#### Background: SAM architecture

---

#### Performance of SAM in Medical Image Segmentation
- **Pathology image segmentation**
- **Liver tumor segementation from CECT**
- **Polyps segemntation from colonoscopy images**
- **Brain MRI segmentation**
- SAM can obtain comparable or even better performance compared with BET
- **Abdominal CT organ segmentation**
- **Endoscopic Surgical instrument segmentation**
- **Other multi-dataset evaluations**

---

#### evaluate SAM on 12 datasets


---

#### summary of SAM dirctly used in medical image
- SAM requires substantial human information to obtain overall moderate performance using only few point or bounding box prompts
- SAM even totally fails in situations when the segmentation targets have weak boundaries with low-contrast and smaller and irregular shape

---

#### Adapting SAM to Medical Image Segmentation
- fine-tuning SAM on medical datasets (small fraction/ parameter-efficient fine-tuning)
- CNN as complementary encoder
- extending the usability of SAM to medical images
- enhancing the robustness against different prompts
- result is highly dependent on the prompt, and the model tends to be more sensitive to wrong prompts
- Input augmentation with SAM

---

#### fine-tuning

---

#### fine-tuning

---

#### enhancing the robustness against different prompts

---

#### enhancing the robustness against different prompts

---

#### Conclusion
- segmentation performance varies significantly across different datasets and tasks

---

#### Future Directions
- buliding large datasets
- integrating clinical domain knowledge
- adapt SAM 2D to 3D
- Reduce the annotation cost
- beyond points and box prompts
- integration into clinical workflow

---

- wilson experiment


---