### Guided Prompting in SAM for Weakly Supervised Cell Segmentation in Histopathological Images
---
- Prompting method
---
### Method (SAM workflow)
- D-SAM: Object detction
- SAM-S: SAM as psuedo mask generator
- SAM-W: SAM with weakly supervised loss
- SAM-M: Mask prompting (SAM-S + D-SAM)
- SAM-ILP (SAM-S+D-SAM and some constraints)
- do Integer Linear Programming over each image
---
- model architecture
---
### SAM-ILP
- two GMM models to model the distribution of intensities of the foreground and background (determined by D-SAM and SAM-S)
- objective function for ILP: (only search over the set of ambiguous pixels) <font color="#f00">Note: ambiguous pixel: 前景背景的edge $e$ 為1的情況 （猜測）</font>
---
- 疑問點： 若 D-SAM 與 SAM-S 預測的結果不一樣，則直接以mask為主?
---
- $x_{ij}$ denotes the pixel belongs to forground/background
- $e_ij$ denotes the edge
---
- different prompt type:
---
- weakly supervised method:
---
- SAM trained with supervised loss peforms worse than D-SAM
---
- D-SAM (local) better distinct instance; while SAM-S (full-view)
---
- SAM-ILP reduce FN; improve quality for wrongly sized box
---
{"contributors":"[{\"id\":\"4724c2c9-74e7-418c-87c3-79a619b97461\",\"add\":1713,\"del\":136}]","description":"type:slide","title":"Guided Prompting in SAM for Weakly Supervised Cell Segmentation in Histopathological Images"}