Amplifying Action-context Greater: Image segmentation-guided Active Bleeding Localization in intraoperative gastrectomy

Proposed Active Bleeding Framework

The schematic diagram of active bleeding framework using our proposed model AMAGI.

Proposed fusion-based active bleeding recognition model AMAGI.

GradCAM Visualization

Example gifs

Data Type	frames	Fast Pathway (SlowFast)	Fast Pathway (AMAGI)	Fusion Layer (AMAGI)
Ex1
Ex2

Surgical Analysis Index

Comparisons of SlowFast and AMAGI on blood count and duration for each 30 test cases.

Split 1

Split2

Split 3

Code with pretrained model

Here is our modified version of mmaction2 for active bleeding framework.

1. Modified implementation of fusion architecture is in

   ./mmaction/models/{segmentors|heads|backbones}
   

2. train and inference scripts are in

   ./my_tools/{amagi_train.sh|amagi_inf.sh} 
   

3. Pretrained semantic segmentation model and fusion model AMAGI are publicly available in our google drive here.